schema: 1 proofs: P001: script: proofs/aggregation/attention_sum_in_range.py page: docs/proofs/aggregation/attention-sum-in-range.py historical: status: Provisional pass source_revision: 810b3fe3b0b25c2c7a718571d83f6006f23110c3 sources: - https://github.com/relflow/relflow/blob/810b3fe/proofs/aggregation/cardinality_generalization/test_attention_sum_in_range.py - https://github.com/relflow/relflow/blob/810b3fe/proofs/aggregation/cardinality_generalization/support.py summary: |- The checked-in report says every behavioral gate passes at this fixed seed. The result remains provisional until the controls are repeated across seeds. insights: |- **Attention can learn totals across varying collection lengths without a supplied count field.** Independent signed amounts prevent count alone from predicting the answer. The reduction retains additive evidence and present-token count alongside its learned summary, providing a route for both content and multiplicity to survive. The held-out accuracy and item-permutation controls support this route within the trained length range. They do not establish a general sum rule: a model could still learn relationships specialized to familiar lengths. The unseen-length proof separately checks longer bags and duplication. Treat this case as evidence for ordinary interpolation, rather than a guarantee that every accepted input shape will be handled accurately. evidence: |- The proof uses deterministic CPU training at seed 3609 for 1,100 updates. The split contains 1,536 training, 384 validation, and 768 test records. The source reports gate clearance but does not record measured scores. nRMSE divides test RMSE by the training-mean baseline error; lower is better. | Check | Required gate | Recorded evidence | | --- | --- | --- | | Held-out sum nRMSE | < 0.35 | Gate reported passing | | Complete-item permutation drift | < 0.12 target SD | Gate reported passing | runs: - started_at: '2026-09-15T02:13:54.778513+00:00' mode: smoke seed: 3609 steps_override: 2 accelerator: gpu threads: 1 provenance: revision: null source_sha256: 12844f7c58729a2d44756463446b10202bdbfc01cf6bce8d81875196243b6eb2 script_sha256: 7407794e6eb66bed5e779a7d57e95d523b17bd81310ef51deaa8f8e72d19abd2 python: 3.12.6 platform: Linux-6.8.0-138-generic-x86_64-with-glibc2.35 host: lab packages: relflow: 0.2.1 torch: 2.12.0 lightning: 2.6.4 numpy: 2.4.6 cuda: '13.0' gpu: NVIDIA GeForce RTX 3090 code_sha256: 0da77456a92d1fef568c8fe84bb26390dcc87322020ede204d81dfc5f912b974 metrics: measured: rmse: 0.9715104338018339 baseline_rmse: 1.0961391615079292 nrmse: 0.8863020936733553 permutation_drift: 0.012183398449126632 steps: 2 checks: Seen-length nRMSE below 0.35: false Permutation drift below 0.12 target SD: true outcome: not_met duration_seconds: 0.862 - started_at: '2026-09-15T02:16:40.875298+00:00' mode: full seed: 3609 steps_override: null accelerator: gpu threads: 1 provenance: revision: null source_sha256: 12844f7c58729a2d44756463446b10202bdbfc01cf6bce8d81875196243b6eb2 script_sha256: 7407794e6eb66bed5e779a7d57e95d523b17bd81310ef51deaa8f8e72d19abd2 python: 3.12.6 platform: Linux-6.8.0-138-generic-x86_64-with-glibc2.35 host: lab packages: relflow: 0.2.1 torch: 2.12.0 lightning: 2.6.4 numpy: 2.4.6 cuda: '13.0' gpu: NVIDIA GeForce RTX 3090 code_sha256: 0da77456a92d1fef568c8fe84bb26390dcc87322020ede204d81dfc5f912b974 metrics: measured: rmse: 0.011440918430587489 baseline_rmse: 1.0961391615079292 nrmse: 0.010437468920322603 permutation_drift: 0.0029657489343600614 steps: 1100 checks: Seen-length nRMSE below 0.35: true Permutation drift below 0.12 target SD: true outcome: met duration_seconds: 35.782 P002: script: proofs/aggregation/attention_sum_unseen_lengths.py page: docs/proofs/aggregation/attention-sum-unseen-lengths.py historical: status: Provisional pass source_revision: 810b3fe3b0b25c2c7a718571d83f6006f23110c3 sources: - https://github.com/relflow/relflow/blob/810b3fe/proofs/aggregation/cardinality_generalization/test_attention_sum_unseen_lengths.py - https://github.com/relflow/relflow/blob/810b3fe/proofs/aggregation/cardinality_generalization/support.py summary: |- The checked-in report says every behavioral gate passes at this fixed seed. The result remains provisional until the controls are repeated across seeds. insights: |- **The learned sum extends beyond trained collection lengths and responds approximately to duplication.** Longer held-out bags test length extrapolation. Duplicating complete bags checks that predictions roughly double, while equal-value probes isolate the need to distinguish different repetition counts. Spare branch capacity only permits longer inputs; it does not establish their meaning. These behavioral checks provide the evidence that the model uses accumulated mass. The equal-value probes themselves use familiar lengths, so they complement rather than replace the unseen-length test. The result supports a bounded extension beyond training, not arbitrary-length or exact summation; missing values and nested totals require separate checks. evidence: |- The proof uses deterministic CPU training at seed 3615 for 1,100 updates. Training uses 1,536 records; validation uses 384. Separate seen and unseen tests contain 512 records each, with 256 short records for duplication. The source reports gate clearance but does not record measured scores. nRMSE divides test RMSE by the training-mean baseline error; lower is better. | Check | Required gate | Recorded evidence | | --- | --- | --- | | Seen-length sum nRMSE | < 0.35 | Gate reported passing | | Unseen-length sum nRMSE | < 0.30 | Gate reported passing | | Complete-bag doubling error | < 0.20 unseen-target SD | Gate reported passing | | Equal-value cardinality error | < 0.20 unseen-target SD | Gate reported passing | runs: - started_at: '2026-09-15T02:13:58.975883+00:00' mode: smoke seed: 3615 steps_override: 2 accelerator: gpu threads: 1 provenance: revision: null source_sha256: 70dfbf74a5344939544a0869e108d29b141f31a431040ac421f69ac2484d3787 script_sha256: 940a234cfd0627f31ce4c070050203072b2954f8ab70587f28e1894f7db2965a python: 3.12.6 platform: Linux-6.8.0-138-generic-x86_64-with-glibc2.35 host: lab packages: relflow: 0.2.1 torch: 2.12.0 lightning: 2.6.4 numpy: 2.4.6 cuda: '13.0' gpu: NVIDIA GeForce RTX 3090 code_sha256: f06fdf3ee348cdab290eee5280ca3978bb3e06ae8fcb501830852bdddb7e87bc metrics: in_range_score: rmse: 0.6269405964118039 baseline_rmse: 1.0401230906449561 nrmse: 0.6027561565074501 unseen_score: rmse: 1.412499210910415 baseline_rmse: 1.7874220810107455 nrmse: 0.7902437963123291 duplication_error: 0.34753360299755176 probe_error: 1.0865928441284074 probe_target: - 0.65 - 1.9500000000000002 - 3.9000000000000004 probe_prediction: - 0.7397217154502869 - 0.7481788396835327 - 0.7602300643920898 steps: 2 checks: All measurements finite: true Seen nRMSE < 0.35, unseen nRMSE < 0.30, duplication and cardinality errors < 0.20: false outcome: not_met duration_seconds: 0.724 - started_at: '2026-09-15T02:17:20.021979+00:00' mode: full seed: 3615 steps_override: null accelerator: gpu threads: 1 provenance: revision: null source_sha256: 70dfbf74a5344939544a0869e108d29b141f31a431040ac421f69ac2484d3787 script_sha256: 940a234cfd0627f31ce4c070050203072b2954f8ab70587f28e1894f7db2965a python: 3.12.6 platform: Linux-6.8.0-138-generic-x86_64-with-glibc2.35 host: lab packages: relflow: 0.2.1 torch: 2.12.0 lightning: 2.6.4 numpy: 2.4.6 cuda: '13.0' gpu: NVIDIA GeForce RTX 3090 code_sha256: f06fdf3ee348cdab290eee5280ca3978bb3e06ae8fcb501830852bdddb7e87bc metrics: in_range_score: rmse: 0.01776947570306214 baseline_rmse: 1.0401230906449561 nrmse: 0.01708401232785218 unseen_score: rmse: 0.1318423008366465 baseline_rmse: 1.7874220810107455 nrmse: 0.07376114586326066 duplication_error: 0.06995505807797699 probe_error: 0.048478743460122514 probe_target: - 0.65 - 1.9500000000000002 - 3.9000000000000004 probe_prediction: - 0.6695117950439453 - 1.9607654809951782 - 3.751617193222046 steps: 1100 checks: All measurements finite: true Seen nRMSE < 0.35, unseen nRMSE < 0.30, duplication and cardinality errors < 0.20: true outcome: met duration_seconds: 34.076 P003: script: proofs/aggregation/mean_duplicate_invariance.py page: docs/proofs/aggregation/mean-duplicate-invariance.py historical: status: Provisional pass source_revision: 810b3fe3b0b25c2c7a718571d83f6006f23110c3 sources: - https://github.com/relflow/relflow/blob/810b3fe/proofs/aggregation/cardinality_generalization/test_mean_duplicate_invariance.py - https://github.com/relflow/relflow/blob/810b3fe/proofs/aggregation/cardinality_generalization/support.py summary: |- The checked-in report says every behavioral gate passes at this fixed seed. The result remains provisional until the controls are repeated across seeds. insights: |- **With item attention disabled, Mean preserves the answer when a whole bag is repeated.** It averages encoded Number tokens, so duplicating every item leaves the representation unchanged. The decoder must still learn how that representation corresponds to the arithmetic mean of the raw amounts. That invariance is useful for an average and destructive for a total. The same duplicated bag has twice the sum, which cannot be recovered from an unchanged summary without additional information. This proof checks learned mean accuracy together with permutation and duplication stability. It does not imply that every architecture containing Mean discards count; earlier item interaction can change what reaches the reducer. evidence: |- The proof uses deterministic CPU training at seed 3600 for 350 updates. Training uses 768 records of lengths one through six. Validation uses 192; the 384 test records have at most three items before duplication. The source reports gate clearance but does not record measured scores. nRMSE divides test RMSE by the training-mean baseline error; lower is better. | Check | Required gate | Recorded evidence | | --- | --- | --- | | Held-out mean nRMSE | < 0.25 | Gate reported passing | | Complete-item permutation drift | < 0.000001 target SD | Gate reported passing | | Complete-bag duplication drift | < 0.000001 target SD | Gate reported passing | runs: - started_at: '2026-09-15T02:14:03.087597+00:00' mode: smoke seed: 3600 steps_override: 2 accelerator: gpu threads: 1 provenance: revision: null source_sha256: 02efc670933c194c70390e7b9d4a1d3f8003d548a006b38fbae516e4f40ce6d6 script_sha256: 0bd2109beffe831baa765f127e9c1786b7c6a557907fc25c0d8d19bdae806cc0 python: 3.12.6 platform: Linux-6.8.0-138-generic-x86_64-with-glibc2.35 host: lab packages: relflow: 0.2.1 torch: 2.12.0 lightning: 2.6.4 numpy: 2.4.6 cuda: '13.0' gpu: NVIDIA GeForce RTX 3090 code_sha256: 2c62be872f128a42d93cbba77f5f789045c57a03fc7b9c1b05a0d4bc965a4b5c metrics: measured: rmse: 0.38763687422212156 baseline_rmse: 0.4464820375017577 nrmse: 0.8682026188356917 permutation_drift: 9.660340108552e-08 duplication_drift: 1.521048304462154e-07 steps: 2 checks: Mean nRMSE below 0.25: false Permutation drift below 1e-6 target SD: true Duplication drift below 1e-6 target SD: true outcome: not_met duration_seconds: 0.692 - started_at: '2026-09-15T02:17:57.471956+00:00' mode: full seed: 3600 steps_override: null accelerator: gpu threads: 1 provenance: revision: null source_sha256: 02efc670933c194c70390e7b9d4a1d3f8003d548a006b38fbae516e4f40ce6d6 script_sha256: 0bd2109beffe831baa765f127e9c1786b7c6a557907fc25c0d8d19bdae806cc0 python: 3.12.6 platform: Linux-6.8.0-138-generic-x86_64-with-glibc2.35 host: lab packages: relflow: 0.2.1 torch: 2.12.0 lightning: 2.6.4 numpy: 2.4.6 cuda: '13.0' gpu: NVIDIA GeForce RTX 3090 code_sha256: 2c62be872f128a42d93cbba77f5f789045c57a03fc7b9c1b05a0d4bc965a4b5c metrics: measured: rmse: 0.00870920478372268 baseline_rmse: 0.4464820375017577 nrmse: 0.019506282565036885 permutation_drift: 9.016303112837792e-08 duplication_drift: 1.4843119244744396e-07 steps: 350 checks: Mean nRMSE below 0.25: true Permutation drift below 1e-6 target SD: true Duplication drift below 1e-6 target SD: true outcome: met duration_seconds: 7.091 P004: script: proofs/aggregation/visible_count_sum.py page: docs/proofs/aggregation/visible-count-sum.py historical: status: Provisional pass source_revision: 810b3fe3b0b25c2c7a718571d83f6006f23110c3 sources: - https://github.com/relflow/relflow/blob/810b3fe/proofs/aggregation/cardinality_generalization/test_visible_count_sum.py - https://github.com/relflow/relflow/blob/810b3fe/proofs/aggregation/cardinality_generalization/support.py summary: |- The checked-in report says every behavioral gate passes at this fixed seed. The result remains provisional until the controls are repeated across seeds. insights: |- **Providing an item count lets the model distinguish totals that an average alone cannot.** Every item in a record repeats the same amount, so the branch summary stays unchanged when only the number of copies changes. The ordinary root count field supplies the missing factor needed for a total. The matched probes require different predictions for the same repeated value at different counts. This isolates learning the amount-times-count relationship from recovering multiplicity through the collection itself. It is a useful diagnostic when a reduction appears to lose mass. The proof covers repeated-value bags and familiar counts; it does not establish sums over arbitrary mixed-value collections. evidence: |- The proof uses deterministic CPU training at seed 3605 for 700 updates. The split contains 1,024 training, 256 validation, and 512 test records. The source reports gate clearance but does not record measured scores. nRMSE divides test RMSE by the training-mean baseline error; lower is better. | Check | Required gate | Recorded evidence | | --- | --- | --- | | Held-out sum nRMSE | < 0.25 | Gate reported passing | | Six-minus-one probe prediction | > 2.5 | Gate reported passing | | Probe RMSE | < 0.35 | Gate reported passing | runs: - started_at: '2026-09-15T02:14:07.120269+00:00' mode: smoke seed: 3605 steps_override: 2 accelerator: gpu threads: 1 provenance: revision: null source_sha256: 6ce281d0517c8197cd92a45e4c1c5a22985eee2ecfa2b19afde05484eb5547db script_sha256: c1885747ad039202f9c713a2dc8c9c2905da05ddf11b891f85c0086dde9d69fe python: 3.12.6 platform: Linux-6.8.0-138-generic-x86_64-with-glibc2.35 host: lab packages: relflow: 0.2.1 torch: 2.12.0 lightning: 2.6.4 numpy: 2.4.6 cuda: '13.0' gpu: NVIDIA GeForce RTX 3090 code_sha256: f7fb9b4e437be5f31e8378a4d66f929898e67dc0298d9a20ed9ff900d9015c61 metrics: measured: rmse: 2.141662911831524 baseline_rmse: 2.2334354369152405 nrmse: 0.9589097031564654 probe_error: 3.231573420102224 probe_target: - 0.7 - 4.199999999999999 probe_prediction: - -0.22359752655029297 - -0.275835245847702 steps: 2 checks: Visible-count nRMSE below 0.25: false Matched-count prediction gap above 2.5: false Matched-count RMSE below 0.35: false outcome: not_met duration_seconds: 0.686 - started_at: '2026-09-15T02:18:07.972553+00:00' mode: full seed: 3605 steps_override: null accelerator: gpu threads: 1 provenance: revision: null source_sha256: 6ce281d0517c8197cd92a45e4c1c5a22985eee2ecfa2b19afde05484eb5547db script_sha256: c1885747ad039202f9c713a2dc8c9c2905da05ddf11b891f85c0086dde9d69fe python: 3.12.6 platform: Linux-6.8.0-138-generic-x86_64-with-glibc2.35 host: lab packages: relflow: 0.2.1 torch: 2.12.0 lightning: 2.6.4 numpy: 2.4.6 cuda: '13.0' gpu: NVIDIA GeForce RTX 3090 code_sha256: f7fb9b4e437be5f31e8378a4d66f929898e67dc0298d9a20ed9ff900d9015c61 metrics: measured: rmse: 0.04991937532832079 baseline_rmse: 2.2334354369152405 nrmse: 0.022350937261597342 probe_error: 0.043074799555061556 probe_target: - 0.7 - 4.199999999999999 probe_prediction: - 0.7487784624099731 - 4.236490249633789 steps: 700 checks: Visible-count nRMSE below 0.25: true Matched-count prediction gap above 2.5: true Matched-count RMSE below 0.35: true outcome: met duration_seconds: 16.503 P005: script: proofs/aggregation/visible_count_unseen_lengths.py page: docs/proofs/aggregation/visible-count-unseen-lengths.py historical: status: Provisional pass source_revision: 810b3fe3b0b25c2c7a718571d83f6006f23110c3 sources: - https://github.com/relflow/relflow/blob/810b3fe/proofs/aggregation/cardinality_generalization/test_visible_count_unseen_lengths.py - https://github.com/relflow/relflow/blob/810b3fe/proofs/aggregation/cardinality_generalization/support.py summary: |- The checked-in report says every behavioral gate passes at this fixed seed. The result remains provisional until the controls are repeated across seeds. insights: |- **Supplying count lets this model extend its learned total to counts absent from training.** Mean carries the repeated amount, while the visible count carries multiplicity. Doubling both the bag and its count checks whether predictions respond approximately like a product. This separates numerical extrapolation from structural counting: the model is told how many items exist. It therefore answers a different question from the Attention proof without a count field. The result supports a limited extension of the amount-times-count relationship, using bags whose values are identical within each record. It does not establish wider mixed-value sums or reliable behavior at arbitrarily large counts. evidence: |- The proof uses deterministic CPU training at seed 3621 for 700 updates. Training uses 1,024 records and validation 256. The unseen test contains 512 records, with 256 separate short records for duplication. The source reports gate clearance but does not record measured scores. nRMSE divides test RMSE by the training-mean baseline error; lower is better. | Check | Required gate | Recorded evidence | | --- | --- | --- | | Unseen-count sum nRMSE | < 0.40 | Gate reported passing | | Complete-bag doubling error | < 0.20 unseen-target SD | Gate reported passing | runs: - started_at: '2026-09-15T02:14:11.138217+00:00' mode: smoke seed: 3621 steps_override: 2 accelerator: gpu threads: 1 provenance: revision: null source_sha256: c1409aed71c9e427c0038327918a3bc6e4f9be0ddc8d781ca7906e76d651f3bb script_sha256: 7925a4c3f6179c4455db597bb005017a232a7267d1e83e7fa3ca117146cc005b python: 3.12.6 platform: Linux-6.8.0-138-generic-x86_64-with-glibc2.35 host: lab packages: relflow: 0.2.1 torch: 2.12.0 lightning: 2.6.4 numpy: 2.4.6 cuda: '13.0' gpu: NVIDIA GeForce RTX 3090 code_sha256: 416dfcbc48fa5dc8eafd8b6a8a79404bc2e3cded6afe662535504301badd5119 metrics: unseen_score: rmse: 4.629937546017952 baseline_rmse: 4.8922481485064635 nrmse: 0.9463824003758698 duplication_error: 0.12502611168610908 steps: 2 checks: All measurements finite: true Unseen nRMSE below 0.40 and duplication error below 0.20: false outcome: not_met duration_seconds: 0.653 - started_at: '2026-09-15T02:18:27.827905+00:00' mode: full seed: 3621 steps_override: null accelerator: gpu threads: 1 provenance: revision: null source_sha256: c1409aed71c9e427c0038327918a3bc6e4f9be0ddc8d781ca7906e76d651f3bb script_sha256: 7925a4c3f6179c4455db597bb005017a232a7267d1e83e7fa3ca117146cc005b python: 3.12.6 platform: Linux-6.8.0-138-generic-x86_64-with-glibc2.35 host: lab packages: relflow: 0.2.1 torch: 2.12.0 lightning: 2.6.4 numpy: 2.4.6 cuda: '13.0' gpu: NVIDIA GeForce RTX 3090 code_sha256: 416dfcbc48fa5dc8eafd8b6a8a79404bc2e3cded6afe662535504301badd5119 metrics: unseen_score: rmse: 1.0328117902916467 baseline_rmse: 4.8922481485064635 nrmse: 0.21111189762664637 duplication_error: 0.17863387934508984 steps: 700 checks: All measurements finite: true Unseen nRMSE below 0.40 and duplication error below 0.20: true outcome: met duration_seconds: 16.509 P006: script: proofs/aggregation/aligned_pair_covariance.py page: docs/proofs/aggregation/aligned-pair-covariance.py historical: status: Provisional pass source_revision: 810b3fe3b0b25c2c7a718571d83f6006f23110c3 sources: - https://github.com/relflow/relflow/blob/810b3fe/proofs/aggregation/distribution_statistics/test_aligned_pair_covariance.py - https://github.com/relflow/relflow/blob/810b3fe/proofs/aggregation/distribution_statistics/support.py summary: |- The checked-in report says every behavioral gate passes at this fixed seed. The result remains provisional until the controls are repeated across seeds. insights: |- **Covariance depends on which X belongs with which Y, not just the two value distributions.** Circularly shifting only Y preserves both marginal sets of numbers but breaks accuracy against the original covariance labels. That is evidence that the model uses item-level pairing. Keep aligned fields on the same repeated item so their relationship remains available before aggregation. The supplied-cross-product control helps distinguish that interaction from the final averaging step. This proof supports learned population covariance in its tested setting; it does not establish Pearson correlation, which also requires normalization by both spreads. Complete-pair permutation and varying collection lengths remain separate checks. evidence: |- The proof uses deterministic CPU training at seed 3508 for 1,100 updates. The split contains 1,536 training, 384 validation, and 768 test records. The source reports gate clearance but does not record measured scores. nRMSE divides test RMSE by the training-mean baseline error; lower is better. | Check | Required gate | Recorded evidence | | --- | --- | --- | | Held-out covariance nRMSE | < 0.35 | Gate reported passing | | Shifted-y nRMSE increase | ≥ 0.35 | Gate reported passing | | Baseline RMSE | > 0.35 | Gate reported passing | runs: - started_at: '2026-09-15T02:14:15.111777+00:00' mode: smoke seed: 3508 steps_override: 2 accelerator: gpu threads: 1 provenance: revision: null source_sha256: 9f4bb0bff69a8408d3a32a4d322c5d54ab1f93ee02890b0750817f6fde1dfa88 script_sha256: 2ffe78169520d9859ef392e497a2147d965579e3a333e729101bfe252b1829af python: 3.12.6 platform: Linux-6.8.0-138-generic-x86_64-with-glibc2.35 host: lab packages: relflow: 0.2.1 torch: 2.12.0 lightning: 2.6.4 numpy: 2.4.6 cuda: '13.0' gpu: NVIDIA GeForce RTX 3090 code_sha256: c22da3412e3bc4bab52bc043f4656508dc3352ad4eacbfa34f570170e974776c metrics: intact: rmse: 0.5274288524922595 baseline_rmse: 0.47722980830253015 nrmse: 1.1051884088470574 corrupted: rmse: 0.5368457160690728 baseline_rmse: 0.47722980830253015 nrmse: 1.1249207545911515 steps: 2 checks: All measurements finite: true Baseline RMSE above 0.35: true Corruption preserves baseline: true Covariance nRMSE below 0.35: false Shuffled pairing raises nRMSE by at least 0.35: false outcome: not_met duration_seconds: 0.942 - started_at: '2026-09-15T02:18:47.729841+00:00' mode: full seed: 3508 steps_override: null accelerator: gpu threads: 1 provenance: revision: null source_sha256: 9f4bb0bff69a8408d3a32a4d322c5d54ab1f93ee02890b0750817f6fde1dfa88 script_sha256: 2ffe78169520d9859ef392e497a2147d965579e3a333e729101bfe252b1829af python: 3.12.6 platform: Linux-6.8.0-138-generic-x86_64-with-glibc2.35 host: lab packages: relflow: 0.2.1 torch: 2.12.0 lightning: 2.6.4 numpy: 2.4.6 cuda: '13.0' gpu: NVIDIA GeForce RTX 3090 code_sha256: c22da3412e3bc4bab52bc043f4656508dc3352ad4eacbfa34f570170e974776c metrics: intact: rmse: 0.03611944439780519 baseline_rmse: 0.47722980830253015 nrmse: 0.0756856419473865 corrupted: rmse: 0.6538480142116774 baseline_rmse: 0.47722980830253015 nrmse: 1.3700904738900628 steps: 1100 checks: All measurements finite: true Baseline RMSE above 0.35: true Corruption preserves baseline: true Covariance nRMSE below 0.35: true Shuffled pairing raises nRMSE by at least 0.35: true outcome: met duration_seconds: 29.055 P007: script: proofs/aggregation/raw_value_variance.py page: docs/proofs/aggregation/raw-value-variance.py historical: status: Provisional pass source_revision: 810b3fe3b0b25c2c7a718571d83f6006f23110c3 sources: - https://github.com/relflow/relflow/blob/810b3fe/proofs/aggregation/distribution_statistics/test_raw_value_variance.py - https://github.com/relflow/relflow/blob/810b3fe/proofs/aggregation/distribution_statistics/support.py summary: |- The checked-in report says every behavioral gate passes at this fixed seed. The result remains provisional until the controls are repeated across seeds. insights: |- **The model can learn how spread out values are, even when their averages are the same.** Independently varied location and scale prevent the mean alone from supplying the variance label. The matched-mean control requires a larger prediction for the more dispersed collection, without providing squared deviations as inputs. This supports learning beyond a single raw-average statistic. It does not mean `rf.Mean()` always erases spread: averaging nonlinear encoded features is different from supplying only the arithmetic mean. This particular proof uses Attention. The squared-deviation control isolates the easier reduction task. This proof uses population variance; a sample-variance target, varying lengths, and missing values need separate checks. evidence: |- The proof uses deterministic CPU training at seed 3504 for 900 updates. The split contains 1,280 training, 320 validation, and 640 test records. The source reports gate clearance but does not record measured scores. nRMSE divides test RMSE by the training-mean baseline error; lower is better. | Check | Required gate | Recorded evidence | | --- | --- | --- | | Held-out variance nRMSE | < 0.30 | Gate reported passing | | Matched-mean prediction separation | > 0.65 | Gate reported passing | runs: - started_at: '2026-09-15T02:14:19.399578+00:00' mode: smoke seed: 3504 steps_override: 2 accelerator: gpu threads: 1 provenance: revision: null source_sha256: 51f0fe1ca3c0f68991bab59bd92cf0d1e8a755b8a60f3b66fcdcf41e342f0d8c script_sha256: c15e9c78bd4282de4c2275d8308d02778eb0d466f0c81934e04132e43fade4a6 python: 3.12.6 platform: Linux-6.8.0-138-generic-x86_64-with-glibc2.35 host: lab packages: relflow: 0.2.1 torch: 2.12.0 lightning: 2.6.4 numpy: 2.4.6 cuda: '13.0' gpu: NVIDIA GeForce RTX 3090 code_sha256: e5acc260bda808202519fbac2453d782447bd949cba214ae3b802040f9a3bdba metrics: measured: rmse: 0.6614792398888426 baseline_rmse: 0.4630170659412655 nrmse: 1.4286282051917982 paired_prediction: - 1.135880708694458 - 1.1724250316619873 steps: 2 checks: Variance nRMSE below 0.30: false Matched-mean variance prediction gap above 0.65: false outcome: not_met duration_seconds: 0.735 - started_at: '2026-09-15T02:19:20.158395+00:00' mode: full seed: 3504 steps_override: null accelerator: gpu threads: 1 provenance: revision: null source_sha256: 51f0fe1ca3c0f68991bab59bd92cf0d1e8a755b8a60f3b66fcdcf41e342f0d8c script_sha256: c15e9c78bd4282de4c2275d8308d02778eb0d466f0c81934e04132e43fade4a6 python: 3.12.6 platform: Linux-6.8.0-138-generic-x86_64-with-glibc2.35 host: lab packages: relflow: 0.2.1 torch: 2.12.0 lightning: 2.6.4 numpy: 2.4.6 cuda: '13.0' gpu: NVIDIA GeForce RTX 3090 code_sha256: e5acc260bda808202519fbac2453d782447bd949cba214ae3b802040f9a3bdba metrics: measured: rmse: 0.01230695448216799 baseline_rmse: 0.4630170659412655 nrmse: 0.02657991548788646 paired_prediction: - 0.037198811769485474 - 1.2225090265274048 steps: 900 checks: Variance nRMSE below 0.30: true Matched-mean variance prediction gap above 0.65: true outcome: met duration_seconds: 23.07 P008: script: proofs/aggregation/supplied_cross_deviations.py page: docs/proofs/aggregation/supplied-cross-deviations.py historical: status: Provisional pass source_revision: 810b3fe3b0b25c2c7a718571d83f6006f23110c3 sources: - https://github.com/relflow/relflow/blob/810b3fe/proofs/aggregation/distribution_statistics/test_supplied_cross_deviations.py - https://github.com/relflow/relflow/blob/810b3fe/proofs/aggregation/distribution_statistics/support.py summary: |- The checked-in report says every behavioral gate passes at this fixed seed. The result remains provisional until the controls are repeated across seeds. insights: |- **Once centered products are supplied, the model can learn their average as covariance.** The difficult information about pairing and centering has already been calculated outside the model. Positive, negative, and canceling contributions then share the same reduction task. This control helps localize a failure in raw-pair covariance. If supplied products are easy but raw pairs are not, the missing behavior is more likely in the interaction than in final averaging. Passing here does not demonstrate learned centering, multiplication, or sensitivity to raw X–Y pairing. The proof checks numerical accuracy for its fixed collection length, without independently testing broader covariance identities. evidence: |- The proof uses deterministic CPU training at seed 3513 for 350 updates. The split contains 768 training, 192 validation, and 384 test records. The source reports gate clearance but does not record measured scores. nRMSE divides test RMSE by the training-mean baseline error; lower is better. | Check | Required gate | Recorded evidence | | --- | --- | --- | | Held-out covariance nRMSE | < 0.20 | Gate reported passing | runs: - started_at: '2026-09-15T02:14:23.484928+00:00' mode: smoke seed: 3513 steps_override: 2 accelerator: gpu threads: 1 provenance: revision: null source_sha256: 772bd81f23b43f14312f70a3f585f5914ad9dc8d3ac095ae146bdd51990259f0 script_sha256: 0769700abc6770b6cf9548da350dc9676ab84c1e7f14a449ad0ecc32cbc27d94 python: 3.12.6 platform: Linux-6.8.0-138-generic-x86_64-with-glibc2.35 host: lab packages: relflow: 0.2.1 torch: 2.12.0 lightning: 2.6.4 numpy: 2.4.6 cuda: '13.0' gpu: NVIDIA GeForce RTX 3090 code_sha256: afb47f7c85ae5d5999437ba681b54168ccf00b3567b810f5b95650749bc823c7 metrics: measured: rmse: 0.6519931857112699 baseline_rmse: 0.4641245625126908 nrmse: 1.4047806092862023 steps: 2 checks: Covariance nRMSE below 0.20: false outcome: not_met duration_seconds: 0.698 - started_at: '2026-09-15T02:19:46.599284+00:00' mode: full seed: 3513 steps_override: null accelerator: gpu threads: 1 provenance: revision: null source_sha256: 772bd81f23b43f14312f70a3f585f5914ad9dc8d3ac095ae146bdd51990259f0 script_sha256: 0769700abc6770b6cf9548da350dc9676ab84c1e7f14a449ad0ecc32cbc27d94 python: 3.12.6 platform: Linux-6.8.0-138-generic-x86_64-with-glibc2.35 host: lab packages: relflow: 0.2.1 torch: 2.12.0 lightning: 2.6.4 numpy: 2.4.6 cuda: '13.0' gpu: NVIDIA GeForce RTX 3090 code_sha256: afb47f7c85ae5d5999437ba681b54168ccf00b3567b810f5b95650749bc823c7 metrics: measured: rmse: 0.008793690333400911 baseline_rmse: 0.4641245625126908 nrmse: 0.018946832474871358 steps: 350 checks: Covariance nRMSE below 0.20: true outcome: met duration_seconds: 7.622 P009: script: proofs/aggregation/supplied_squared_deviations.py page: docs/proofs/aggregation/supplied-squared-deviations.py historical: status: Provisional pass source_revision: 810b3fe3b0b25c2c7a718571d83f6006f23110c3 sources: - https://github.com/relflow/relflow/blob/810b3fe/proofs/aggregation/distribution_statistics/test_supplied_squared_deviations.py - https://github.com/relflow/relflow/blob/810b3fe/proofs/aggregation/distribution_statistics/support.py summary: |- The checked-in report says every behavioral gate passes at this fixed seed. The result remains provisional until the controls are repeated across seeds. insights: |- **Providing squared deviations tests averaging; it does not show that the model can calculate variance from raw values.** The generator has already centered and squared each value. The model receives those sufficient statistics and learns to decode their average as population variance. That makes the case a useful diagnostic beside raw-value variance: if this succeeds while the raw route fails, investigate centering or the learned nonlinear interaction before blaming final reduction. Mean averages encoded tokens, so numerical decoding still has to work. The accuracy gate does not independently establish permutation behavior, varying collection lengths, or handling of missing values, and it supplies no evidence that the model learned the preprocessing operation. evidence: |- The proof uses deterministic CPU training at seed 3500 for 350 updates. The split contains 768 training, 192 validation, and 384 test records. The source reports gate clearance but does not record measured scores. nRMSE divides test RMSE by the training-mean baseline error; lower is better. | Check | Required gate | Recorded evidence | | --- | --- | --- | | Held-out variance nRMSE | < 0.20 | Gate reported passing | runs: - started_at: '2026-09-15T02:14:27.504678+00:00' mode: smoke seed: 3500 steps_override: 2 accelerator: gpu threads: 1 provenance: revision: null source_sha256: 9794c74e27d86088dfb4ea8d03abccb0a8bc826edc859bf652613b0ffc95017b script_sha256: 342287554d0bcab4367eedfd87f4f8f5361511ebd304cf1d100a9d4e1480e5d8 python: 3.12.6 platform: Linux-6.8.0-138-generic-x86_64-with-glibc2.35 host: lab packages: relflow: 0.2.1 torch: 2.12.0 lightning: 2.6.4 numpy: 2.4.6 cuda: '13.0' gpu: NVIDIA GeForce RTX 3090 code_sha256: 19fb61d95346c08bc1691b7b2efd5acb4eddba8782100b18e3777490f80e9f6d metrics: measured: rmse: 0.6144500835113652 baseline_rmse: 0.4800716415268894 nrmse: 1.2799133095158028 steps: 2 checks: Variance nRMSE below 0.20: false outcome: not_met duration_seconds: 0.677 - started_at: '2026-09-15T02:19:57.574262+00:00' mode: full seed: 3500 steps_override: null accelerator: gpu threads: 1 provenance: revision: null source_sha256: 9794c74e27d86088dfb4ea8d03abccb0a8bc826edc859bf652613b0ffc95017b script_sha256: 342287554d0bcab4367eedfd87f4f8f5361511ebd304cf1d100a9d4e1480e5d8 python: 3.12.6 platform: Linux-6.8.0-138-generic-x86_64-with-glibc2.35 host: lab packages: relflow: 0.2.1 torch: 2.12.0 lightning: 2.6.4 numpy: 2.4.6 cuda: '13.0' gpu: NVIDIA GeForce RTX 3090 code_sha256: 19fb61d95346c08bc1691b7b2efd5acb4eddba8782100b18e3777490f80e9f6d metrics: measured: rmse: 0.010172192102680725 baseline_rmse: 0.4800716415268894 nrmse: 0.02118890436920542 steps: 350 checks: Variance nRMSE below 0.20: true outcome: met duration_seconds: 7.135 P010: script: proofs/aggregation/selected_group_raw_weighted_sum.py page: docs/proofs/aggregation/selected-group-raw-weighted-sum.py historical: status: Provisional pass source_revision: 810b3fe3b0b25c2c7a718571d83f6006f23110c3 sources: - https://github.com/relflow/relflow/blob/810b3fe/proofs/aggregation/grouped_weighted_aggregation/test_selected_group_raw_weighted_sum.py - https://github.com/relflow/relflow/blob/810b3fe/proofs/aggregation/grouped_weighted_aggregation/support.py summary: |- The checked-in report says every behavioral gate passes at this fixed seed. The result remains provisional until the controls are repeated across seeds. insights: |- **The model can select a requested group and learn its weighted total from raw values and weights.** The same bag appears with different requested groups, so bag content alone cannot determine the answer. Rotating group labels tests selection; swapping weights within each group separately tests the value–weight relationship. Both controls remove accuracy against retained original labels. Keeping group, value, and weight together preserves the relationships the answer needs. The supplied-contribution control helps separate multiplication from selection and reduction. The result is still narrow: groups are familiar and equally populated. Unseen labels, absent groups, and varying group sizes are additional tasks, not consequences of this pass. evidence: |- The proof uses deterministic CPU training at seed 3605 for 1,100 updates. The split uses 512 training, 128 validation, and 256 test bags, each expanded into three requests. The source reports gate clearance but does not record measured scores. nRMSE divides test RMSE by the training-mean baseline error; lower is better. | Check | Required gate | Recorded evidence | | --- | --- | --- | | Held-out answer nRMSE | < 0.50 | Gate reported passing | | Group-label corruption nRMSE increase | ≥ 0.20 | Gate reported passing | | Within-group weight-swap nRMSE increase | ≥ 0.20 | Gate reported passing | | Complete-item permutation drift | < 0.10 target SD | Gate reported passing | runs: - started_at: '2026-09-15T02:14:31.538988+00:00' mode: smoke seed: 3605 steps_override: 2 accelerator: gpu threads: 1 provenance: revision: null source_sha256: 27937931b9c7d89f9cced00722a005874718f90849eb68704f32bb43c221fa0d script_sha256: 7c8a462450b4d74edd3322f2ecad3ad037b22434fcf56896f842e33ab6403ccf python: 3.12.6 platform: Linux-6.8.0-138-generic-x86_64-with-glibc2.35 host: lab packages: relflow: 0.2.1 torch: 2.12.0 lightning: 2.6.4 numpy: 2.4.6 cuda: '13.0' gpu: NVIDIA GeForce RTX 3090 code_sha256: 10677fafe774f9ac7a11e615d3dd85df77b22b4b918ccbf8475a99f5033b455b metrics: intact: rmse: 0.8720222974706681 baseline_rmse: 0.8856217062634284 nrmse: 0.9846442237169882 labels: rmse: 0.8719627359023779 baseline_rmse: 0.8856217062634284 nrmse: 0.984576969755315 pairing: rmse: 0.8717713469755564 baseline_rmse: 0.8856217062634284 nrmse: 0.9843608628944871 permutation_drift: 0.005300220432817928 steps: 2 checks: All measurements finite: true Raw grouped sum nRMSE below 0.50: false Rotated labels raise nRMSE by at least 0.20: false Swapped weights raise nRMSE by at least 0.20: false Permutation drift below 0.10 target SD: true outcome: not_met duration_seconds: 1.434 - started_at: '2026-09-15T02:20:08.067014+00:00' mode: full seed: 3605 steps_override: null accelerator: gpu threads: 1 provenance: revision: null source_sha256: 27937931b9c7d89f9cced00722a005874718f90849eb68704f32bb43c221fa0d script_sha256: 7c8a462450b4d74edd3322f2ecad3ad037b22434fcf56896f842e33ab6403ccf python: 3.12.6 platform: Linux-6.8.0-138-generic-x86_64-with-glibc2.35 host: lab packages: relflow: 0.2.1 torch: 2.12.0 lightning: 2.6.4 numpy: 2.4.6 cuda: '13.0' gpu: NVIDIA GeForce RTX 3090 code_sha256: 10677fafe774f9ac7a11e615d3dd85df77b22b4b918ccbf8475a99f5033b455b metrics: intact: rmse: 0.043258763522469844 baseline_rmse: 0.8856217062634284 nrmse: 0.04884564506101041 labels: rmse: 1.2619153938747705 baseline_rmse: 0.8856217062634284 nrmse: 1.424892123747714 pairing: rmse: 0.5351155153189048 baseline_rmse: 0.8856217062634284 nrmse: 0.6042258354039648 permutation_drift: 0.02378622793395685 steps: 1100 checks: All measurements finite: true Raw grouped sum nRMSE below 0.50: true Rotated labels raise nRMSE by at least 0.20: true Swapped weights raise nRMSE by at least 0.20: true Permutation drift below 0.10 target SD: true outcome: met duration_seconds: 34.949 P011: script: proofs/aggregation/selected_group_supplied_contribution_sum.py page: docs/proofs/aggregation/selected-group-supplied-contribution-sum.py historical: status: Provisional pass source_revision: 810b3fe3b0b25c2c7a718571d83f6006f23110c3 sources: - https://github.com/relflow/relflow/blob/810b3fe/proofs/aggregation/grouped_weighted_aggregation/test_selected_group_supplied_contribution_sum.py - https://github.com/relflow/relflow/blob/810b3fe/proofs/aggregation/grouped_weighted_aggregation/support.py summary: |- The checked-in report says every behavioral gate passes at this fixed seed. The result remains provisional until the controls are repeated across seeds. insights: |- **The model can select and sum a requested group’s supplied contributions.** Each bag is reused with different requests, and rotating item group labels breaks accuracy against the original answer. That makes the association between groups and contributions relevant, rather than allowing one answer for the whole bag. The products are already present, so this case tests selection and reduction without establishing learned multiplication. Compare it with the raw-pair proof when diagnosing a difficult schema. The two routes expose different numbers of input fields, so their scores are not a matched comparison of model capacity. Missing groups and new category labels also remain outside this case. evidence: |- The proof uses deterministic CPU training at seed 3600 for 800 updates. The split uses 384 training, 96 validation, and 192 test bags, each expanded into three requests. The source reports gate clearance but does not record measured scores. nRMSE divides test RMSE by the training-mean baseline error; lower is better. | Check | Required gate | Recorded evidence | | --- | --- | --- | | Held-out answer nRMSE | < 0.50 | Gate reported passing | | Group-label corruption nRMSE increase | ≥ 0.20 | Gate reported passing | | Complete-item permutation drift | < 0.10 target SD | Gate reported passing | runs: - started_at: '2026-09-15T02:14:36.319678+00:00' mode: smoke seed: 3600 steps_override: 2 accelerator: gpu threads: 1 provenance: revision: null source_sha256: b8adda32c5f5c52e2ac75517b67ba47e34f325d46ad60ae14baef2822def6f30 script_sha256: ae820fdf66aa1d6daf4738a363a65cd6b8430edab73b834eac71cd943fe76643 python: 3.12.6 platform: Linux-6.8.0-138-generic-x86_64-with-glibc2.35 host: lab packages: relflow: 0.2.1 torch: 2.12.0 lightning: 2.6.4 numpy: 2.4.6 cuda: '13.0' gpu: NVIDIA GeForce RTX 3090 code_sha256: b761289e7e48ae19e8a183638ff5da2d292099945133c2c25bd57d1b4c9e9e4c metrics: intact: rmse: 0.9259963059149292 baseline_rmse: 0.9164778582723881 nrmse: 1.0103859002774862 corrupted: rmse: 0.9261492348204072 baseline_rmse: 0.9164778582723881 nrmse: 1.010552766180571 permutation_drift: 0.001964887514259805 steps: 2 checks: All measurements finite: true Intact calibration nRMSE below 1.60: true Corrupt calibration nRMSE below 1.80: true Grouped contribution nRMSE below 0.50: false Rotated labels raise nRMSE by at least 0.20: false Permutation drift below 0.10 target SD: true outcome: not_met duration_seconds: 0.986 - started_at: '2026-09-15T02:20:46.436531+00:00' mode: full seed: 3600 steps_override: null accelerator: gpu threads: 1 provenance: revision: null source_sha256: b8adda32c5f5c52e2ac75517b67ba47e34f325d46ad60ae14baef2822def6f30 script_sha256: ae820fdf66aa1d6daf4738a363a65cd6b8430edab73b834eac71cd943fe76643 python: 3.12.6 platform: Linux-6.8.0-138-generic-x86_64-with-glibc2.35 host: lab packages: relflow: 0.2.1 torch: 2.12.0 lightning: 2.6.4 numpy: 2.4.6 cuda: '13.0' gpu: NVIDIA GeForce RTX 3090 code_sha256: b761289e7e48ae19e8a183638ff5da2d292099945133c2c25bd57d1b4c9e9e4c metrics: intact: rmse: 0.06792581690778365 baseline_rmse: 0.9164778582723881 nrmse: 0.07411615708406487 corrupted: rmse: 1.277367547523599 baseline_rmse: 0.9164778582723881 nrmse: 1.393778950570075 permutation_drift: 0.030823682800309726 steps: 800 checks: All measurements finite: true Intact calibration nRMSE below 1.60: true Corrupt calibration nRMSE below 1.80: true Grouped contribution nRMSE below 0.50: true Rotated labels raise nRMSE by at least 0.20: true Permutation drift below 0.10 target SD: true outcome: met duration_seconds: 23.792 P012: script: proofs/aggregation/requested_rank_values.py page: docs/proofs/aggregation/requested-rank-values.py historical: status: Provisional pass source_revision: 810b3fe3b0b25c2c7a718571d83f6006f23110c3 sources: - https://github.com/relflow/relflow/blob/810b3fe/proofs/aggregation/order_statistics/test_requested_rank_values.py - https://github.com/relflow/relflow/blob/810b3fe/proofs/aggregation/order_statistics/support.py summary: |- The checked-in report says every behavioral gate passes at this fixed seed. The result remains provisional until the controls are repeated across seeds. insights: |- **The model can use a request to pick the minimum, maximum, median, or a quartile from an unsorted collection.** Each bag appears with every request, so a bag-only answer cannot solve the task. Hiding rank collapses those predictions, while cycling rank labels destroys accuracy against the original targets. Preserving encoded item coordinates leaves evidence available for request-conditioned decoding; it does not itself perform sorting. Separate gates for every rank and data shape prevent good minimum or maximum predictions from hiding failed quartiles. The tested quantiles select observed values from an odd-length bag. Even lengths, interpolation conventions, empty collections, and much larger bags are additional questions. evidence: |- The proof uses deterministic CPU training at seed 3700 for 800 updates. The split contains 1,024 training, 192 validation, and 384 test bags, each expanded into five requests. nRMSE baselines are matched to each reported rank or data-shape cell. The values below are recorded in the source report, not a new run. Lower nRMSE is better; 1.0 matches the corresponding constant-mean baseline. | Check | Required gate | Recorded evidence | | --- | --- | --- | | Minimum / maximum nRMSE | < 0.30 each | 0.0423 / 0.0550 | | Q25 / median / Q75 nRMSE | < 0.35 each | 0.0975 / 0.0882 / 0.1243 | | Symmetric / left-skewed / right-skewed nRMSE | < 0.45 each | 0.0580 / 0.0549 / 0.0522 | | Duplicate-bearing nRMSE | < 0.45 | 0.0820 | | Complete-item permutation drift | < 0.10 target SD | 0.0367 target SD | | Cycled-rank nRMSE | ≥ 0.80 and increase ≥ 0.35 | 1.4380 (intact overall: 0.0823) | | Hidden-rank within-bag spread | ≤ 0.000001 | 0.0000002384 | | Hidden-rank degradation | Both endpoints and ≥ 3 ranks at nRMSE ≥ 0.65 | Gate reported passing | runs: - started_at: '2026-09-15T02:14:40.625295+00:00' mode: smoke seed: 3700 steps_override: 2 accelerator: gpu threads: 1 provenance: revision: null source_sha256: 39f67bb6c9feee68195d0831b099e7d1c7baa9126415cbc747ca26177e703b2b script_sha256: edf7585291d13f57fef4a81aa9a1a541b958e220c979df889d23d9e9ee45bdf4 python: 3.12.6 platform: Linux-6.8.0-138-generic-x86_64-with-glibc2.35 host: lab packages: relflow: 0.2.1 torch: 2.12.0 lightning: 2.6.4 numpy: 2.4.6 cuda: '13.0' gpu: NVIDIA GeForce RTX 3090 code_sha256: 8731b8d90639964fe900d976f62a7a9898f57242f861056429469d709e0cd718 metrics: intact: rmse: 1.72987399014268 baseline_rmse: 0.9290372023301314 nrmse: 1.8620072326532873 by_rank: maximum: rmse: 1.2088060619832266 baseline_rmse: 0.9466038978237984 nrmse: 1.2769924830884594 median: rmse: 1.4201135867295638 baseline_rmse: 0.8438626855475819 nrmse: 1.682872831150311 minimum: rmse: 2.517855671484593 baseline_rmse: 1.0886907856680468 nrmse: 2.312737192810516 q25: rmse: 1.9098741998891142 baseline_rmse: 0.8870723799603668 nrmse: 2.1530083035326215 q75: rmse: 1.2235884816106544 baseline_rmse: 0.8574627753116424 nrmse: 1.4269872895251279 by_shape: duplicates: rmse: 1.7452287371682402 baseline_rmse: 1.2044208081983223 nrmse: 1.4490190847656523 left_skewed: rmse: 1.8125247855763011 baseline_rmse: 1.2724328485237015 nrmse: 1.4244561413824106 right_skewed: rmse: 1.6108291608029603 baseline_rmse: 1.1207206271078058 nrmse: 1.4373155288128816 symmetric: rmse: 1.7447108810428138 baseline_rmse: 1.2584867784356901 nrmse: 1.3863561468730758 cycled: rmse: 1.8391260087321306 baseline_rmse: 0.9290372023301314 nrmse: 1.9796042657058215 hidden: maximum: rmse: 0.8487453493215398 baseline_rmse: 0.9466038978237984 nrmse: 0.8966214393082141 median: rmse: 1.6411632223135708 baseline_rmse: 0.8438626855475819 nrmse: 1.9448226001942734 minimum: rmse: 2.868993195177375 baseline_rmse: 1.0886907856680468 nrmse: 2.6352691075794232 q25: rmse: 2.0961026769485365 baseline_rmse: 0.8870723799603668 nrmse: 2.362944359785148 q75: rmse: 1.2235404929666727 baseline_rmse: 0.8574627753116424 nrmse: 1.426931323662395 permutation_drift: 0.0012358029816055868 hidden_family_spread: 0.0 target_scale: 1.2148096616638178 steps: 2 checks: Expected number of rank requests: true All ranks represented: true All bag shapes represented: true Training bags separate from validation and test: true Validation bags separate from test: true All measurements finite: true Target SD above 0.80: true Per-rank baseline RMSE above 0.60: true Per-shape baseline RMSE above 0.80: true Hidden requests yield identical predictions: true Endpoint, interior-rank, shape, and causal-control gates: false outcome: not_met duration_seconds: 1.905 - started_at: '2026-09-15T02:21:13.612512+00:00' mode: full seed: 3700 steps_override: null accelerator: gpu threads: 1 provenance: revision: null source_sha256: 39f67bb6c9feee68195d0831b099e7d1c7baa9126415cbc747ca26177e703b2b script_sha256: edf7585291d13f57fef4a81aa9a1a541b958e220c979df889d23d9e9ee45bdf4 python: 3.12.6 platform: Linux-6.8.0-138-generic-x86_64-with-glibc2.35 host: lab packages: relflow: 0.2.1 torch: 2.12.0 lightning: 2.6.4 numpy: 2.4.6 cuda: '13.0' gpu: NVIDIA GeForce RTX 3090 code_sha256: 8731b8d90639964fe900d976f62a7a9898f57242f861056429469d709e0cd718 metrics: intact: rmse: 0.06546551629972189 baseline_rmse: 0.9290372023301314 nrmse: 0.07046597933379513 by_rank: maximum: rmse: 0.03148641920531384 baseline_rmse: 0.9466038978237984 nrmse: 0.03326250745184946 median: rmse: 0.0777021138527348 baseline_rmse: 0.8438626855475819 nrmse: 0.09207909673398339 minimum: rmse: 0.048558041332576324 baseline_rmse: 1.0886907856680468 nrmse: 0.0446022341438115 q25: rmse: 0.0775761143980616 baseline_rmse: 0.8870723799603668 nrmse: 0.08745184288290837 q75: rmse: 0.07761262221592768 baseline_rmse: 0.8574627753116424 nrmse: 0.09051427589695611 by_shape: duplicates: rmse: 0.07047749576287497 baseline_rmse: 1.2044208081983223 nrmse: 0.05851567432507361 left_skewed: rmse: 0.06460019800743773 baseline_rmse: 1.2724328485237015 nrmse: 0.05076904300481396 right_skewed: rmse: 0.05263029709184923 baseline_rmse: 1.1207206271078058 nrmse: 0.04696112110265152 symmetric: rmse: 0.07233757069444668 baseline_rmse: 1.2584867784356901 nrmse: 0.05747980188108365 cycled: rmse: 1.3455171411280074 baseline_rmse: 0.9290372023301314 nrmse: 1.4482919927784341 hidden: maximum: rmse: 1.8848899419521308 baseline_rmse: 0.9466038978237984 nrmse: 1.9912129522025122 median: rmse: 0.7047759204544627 baseline_rmse: 0.8438626855475819 nrmse: 0.8351784390100554 minimum: rmse: 0.8839106462852007 baseline_rmse: 1.0886907856680468 nrmse: 0.8119023858026059 q25: rmse: 0.21711282571787752 baseline_rmse: 0.8870723799603668 nrmse: 0.2447520975994967 q75: rmse: 1.2040925365318929 baseline_rmse: 0.8574627753116424 nrmse: 1.4042505064948958 permutation_drift: 0.029800371800515996 hidden_family_spread: 0.0 target_scale: 1.2148096616638178 steps: 800 checks: Expected number of rank requests: true All ranks represented: true All bag shapes represented: true Training bags separate from validation and test: true Validation bags separate from test: true All measurements finite: true Target SD above 0.80: true Per-rank baseline RMSE above 0.60: true Per-shape baseline RMSE above 0.80: true Hidden requests yield identical predictions: true Endpoint, interior-rank, shape, and causal-control gates: true outcome: met duration_seconds: 21.578 P013: script: proofs/aggregation/mean_erases_contribution_count.py page: docs/proofs/aggregation/mean-erases-contribution-count.py historical: status: Expected limitation source_revision: 810b3fe3b0b25c2c7a718571d83f6006f23110c3 sources: - https://github.com/relflow/relflow/blob/810b3fe/proofs/aggregation/weighted_aggregation/test_mean_erases_contribution_count.py - https://github.com/relflow/relflow/blob/810b3fe/proofs/aggregation/weighted_aggregation/support.py summary: |- The checked-in report confirms the intended count-erasure boundary. Passing this proof does not mean its sum predictions are correct. insights: |- **This Mean configuration discards the count required to recover a sum.** With item attention disabled, averaging identical encoded contributions produces the same summary for one copy or many. A larger decoder cannot recover a distinction that no longer reaches it. The model still learns the contribution average, showing that useful content survives. Its matching sum predictions are the intended limitation, even though the correct totals differ. To predict variable-length totals, preserve multiplicity through a suitable reduction or supply an explicit count. Mean remains useful when repeating the whole collection should leave the desired answer unchanged. evidence: |- The proof uses deterministic CPU training at seed 3309 for 300 updates. The split contains 1,024 training, 256 validation, and 512 test records. The source reports gate clearance but does not record measured scores. nRMSE divides test RMSE by the training-mean baseline error; lower is better. | Check | Required gate | Recorded evidence | | --- | --- | --- | | Contribution-average nRMSE | < 0.25 | Gate reported passing | | One-versus-six sum prediction difference | ≤ 0.000001 | Gate reported passing | runs: - started_at: '2026-09-15T02:14:45.888837+00:00' mode: smoke seed: 3309 steps_override: 2 accelerator: gpu threads: 1 provenance: revision: null source_sha256: 7e5aab558b63eaef5c34cffc99b3a41b4be03057be065a22a5f68d770c0e209f script_sha256: eeea4bcdd55825b7b630e130d623ab78c41e7e4142776d8c8d1de2765c795461 python: 3.12.6 platform: Linux-6.8.0-138-generic-x86_64-with-glibc2.35 host: lab packages: relflow: 0.2.1 torch: 2.12.0 lightning: 2.6.4 numpy: 2.4.6 cuda: '13.0' gpu: NVIDIA GeForce RTX 3090 code_sha256: be07c95b2046830c894f00b369878efe9bdd1a96d793eb8990d3fec5d52033b9 metrics: mean_score: rmse: 0.43001463612786694 baseline_rmse: 0.711858212210264 nrmse: 0.6040734359061549 paired_sum: - 0.47120678424835205 - 0.47120681405067444 steps: 2 checks: Mean contribution nRMSE below 0.25: false Mean reduction erases contribution count: true outcome: not_met duration_seconds: 0.635 - started_at: '2026-09-15T02:21:38.565842+00:00' mode: full seed: 3309 steps_override: null accelerator: gpu threads: 1 provenance: revision: null source_sha256: 7e5aab558b63eaef5c34cffc99b3a41b4be03057be065a22a5f68d770c0e209f script_sha256: eeea4bcdd55825b7b630e130d623ab78c41e7e4142776d8c8d1de2765c795461 python: 3.12.6 platform: Linux-6.8.0-138-generic-x86_64-with-glibc2.35 host: lab packages: relflow: 0.2.1 torch: 2.12.0 lightning: 2.6.4 numpy: 2.4.6 cuda: '13.0' gpu: NVIDIA GeForce RTX 3090 code_sha256: be07c95b2046830c894f00b369878efe9bdd1a96d793eb8990d3fec5d52033b9 metrics: mean_score: rmse: 0.018810305821033203 baseline_rmse: 0.711858212210264 nrmse: 0.02642423097519473 paired_sum: - 2.6523079872131348 - 2.652308464050293 steps: 300 checks: Mean contribution nRMSE below 0.25: true Mean reduction erases contribution count: true outcome: met duration_seconds: 7.326 P014: script: proofs/aggregation/raw_value_weight_sum.py page: docs/proofs/aggregation/raw-value-weight-sum.py historical: status: Provisional pass source_revision: 810b3fe3b0b25c2c7a718571d83f6006f23110c3 sources: - https://github.com/relflow/relflow/blob/810b3fe/proofs/aggregation/weighted_aggregation/test_raw_value_weight_sum.py - https://github.com/relflow/relflow/blob/810b3fe/proofs/aggregation/weighted_aggregation/support.py summary: |- The checked-in report says every behavioral gate passes at this fixed seed. The result remains provisional until the controls are repeated across seeds. insights: |- **The model uses the weight attached to each value, rather than only the collection’s separate values and weights.** Swapping weights between items preserves both sets of numbers but worsens predictions against the original labels. Reordering complete items leaves predictions approximately stable, while scaling all weights changes the answer proportionally. Keeping related fields on the same item gives the model a route to learn their interaction before aggregation. The supplied-product control helps isolate that interaction when diagnosing a failure. This proof supports the raw paired schema for its fixed collection length; it does not establish variable-length totals, negative-weight behavior, or exact multiplication. evidence: |- The proof uses deterministic CPU training at seed 3304 for 900 updates. The split contains 1,024 training, 256 validation, and 512 test records. The source reports gate clearance but does not record measured scores. nRMSE divides test RMSE by the training-mean baseline error; lower is better. | Check | Required gate | Recorded evidence | | --- | --- | --- | | Held-out nRMSE | < 0.30 | Gate reported passing | | Weight-permutation nRMSE increase | ≥ 0.35 | Gate reported passing | | Scaled-weight nRMSE | < 0.35 | Gate reported passing | | Complete-item permutation drift | < 0.08 target SD | Gate reported passing | | Weight-scaling error | < 0.15 scaled-target SD | Gate reported passing | runs: - started_at: '2026-09-15T02:14:49.869994+00:00' mode: smoke seed: 3304 steps_override: 2 accelerator: gpu threads: 1 provenance: revision: null source_sha256: 663ebdfbebc94e9be8bc9a1ae859d351aa37aa7fab3e21279afbeca2e745122c script_sha256: 22ad4b6ad0f2c2e18e5d2f4ebfb49e7849f22e5bbf04828fb94e3a05edb5a2aa python: 3.12.6 platform: Linux-6.8.0-138-generic-x86_64-with-glibc2.35 host: lab packages: relflow: 0.2.1 torch: 2.12.0 lightning: 2.6.4 numpy: 2.4.6 cuda: '13.0' gpu: NVIDIA GeForce RTX 3090 code_sha256: 4c7d7d064a251e97905d4c30e5d117c0110b815f702a9233e77d701f7280c815 metrics: intact: rmse: 1.102865133530183 baseline_rmse: 1.2971704626253104 nrmse: 0.8502083305983719 corrupted: rmse: 1.1030686681928599 baseline_rmse: 1.2971704626253104 nrmse: 0.8503652372413624 scaled: rmse: 0.7985595666588989 baseline_rmse: 0.9728778469689828 nrmse: 0.8208220272944076 permutation_drift: 0.028831436786201575 scaling_error: 0.12094008745694862 steps: 2 checks: Weighted sum nRMSE below 0.30: false Shuffled weights raise nRMSE by at least 0.35: false Scaled-weight nRMSE below 0.35: false Permutation drift below 0.08 target SD: true Weight scaling error below 0.15 target SD: true outcome: not_met duration_seconds: 0.849 - started_at: '2026-09-15T02:21:49.304365+00:00' mode: full seed: 3304 steps_override: null accelerator: gpu threads: 1 provenance: revision: null source_sha256: 663ebdfbebc94e9be8bc9a1ae859d351aa37aa7fab3e21279afbeca2e745122c script_sha256: 22ad4b6ad0f2c2e18e5d2f4ebfb49e7849f22e5bbf04828fb94e3a05edb5a2aa python: 3.12.6 platform: Linux-6.8.0-138-generic-x86_64-with-glibc2.35 host: lab packages: relflow: 0.2.1 torch: 2.12.0 lightning: 2.6.4 numpy: 2.4.6 cuda: '13.0' gpu: NVIDIA GeForce RTX 3090 code_sha256: 4c7d7d064a251e97905d4c30e5d117c0110b815f702a9233e77d701f7280c815 metrics: intact: rmse: 0.02914664462386432 baseline_rmse: 1.2971704626253104 nrmse: 0.022469402028223156 corrupted: rmse: 0.7510852935035259 baseline_rmse: 1.2971704626253104 nrmse: 0.5790181900869246 scaled: rmse: 0.03725923202770983 baseline_rmse: 0.9728778469689828 nrmse: 0.03829795502466377 permutation_drift: 0.01038750135440631 scaling_error: 0.032901590615004826 steps: 900 checks: Weighted sum nRMSE below 0.30: true Shuffled weights raise nRMSE by at least 0.35: true Scaled-weight nRMSE below 0.35: true Permutation drift below 0.08 target SD: true Weight scaling error below 0.15 target SD: true outcome: met duration_seconds: 26.607 P015: script: proofs/aggregation/supplied_contribution_sum.py page: docs/proofs/aggregation/supplied-contribution-sum.py historical: status: Provisional pass source_revision: 810b3fe3b0b25c2c7a718571d83f6006f23110c3 sources: - https://github.com/relflow/relflow/blob/810b3fe/proofs/aggregation/weighted_aggregation/test_supplied_contribution_sum.py - https://github.com/relflow/relflow/blob/810b3fe/proofs/aggregation/weighted_aggregation/support.py summary: |- The checked-in report says every behavioral gate passes at this fixed seed. The result remains provisional until the controls are repeated across seeds. insights: |- **The model can sum contributions when their products have already been calculated.** This removes value–weight binding and multiplication from the task, leaving numerical encoding, reduction, and decoding. It is useful evidence that the simpler aggregation route works. A failure on raw pairs alongside success here would point toward the missing interaction, rather than summation alone. However, every collection has the same length, so learning an average and a fixed scaling factor could also solve this task. The proof has an accuracy gate but no separate duplication or permutation gate. Its success should not be described as general arithmetic over arbitrary collection lengths. evidence: |- The proof uses deterministic CPU training at seed 3300 for 500 updates. The split contains 768 training, 192 validation, and 384 test records. The source reports gate clearance but does not record measured scores. nRMSE divides test RMSE by the training-mean baseline error; lower is better. | Check | Required gate | Recorded evidence | | --- | --- | --- | | Held-out sum nRMSE | < 0.25 | Gate reported passing | runs: - started_at: '2026-09-15T02:14:54.062805+00:00' mode: smoke seed: 3300 steps_override: 2 accelerator: gpu threads: 1 provenance: revision: null source_sha256: aa9e4c1923a4dccf37b5737ea295f9c6ec0b22c4bc6c3c11921c7e1ad6e4cd5b script_sha256: 053b5681918b4d8dc4270d297ddd7803f8806a916f8c4340556cc7610ee1b727 python: 3.12.6 platform: Linux-6.8.0-138-generic-x86_64-with-glibc2.35 host: lab packages: relflow: 0.2.1 torch: 2.12.0 lightning: 2.6.4 numpy: 2.4.6 cuda: '13.0' gpu: NVIDIA GeForce RTX 3090 code_sha256: 7bf47e7f1f7173602f6d5d132f4aee9d0955fbf9e42640b5065a15de127c636e metrics: measured: rmse: 1.0598801036619754 baseline_rmse: 1.2867287797077822 nrmse: 0.8237012495381315 steps: 2 checks: Contribution sum nRMSE below 0.25: false outcome: not_met duration_seconds: 0.678 - started_at: '2026-09-15T02:22:19.301684+00:00' mode: full seed: 3300 steps_override: null accelerator: gpu threads: 1 provenance: revision: null source_sha256: aa9e4c1923a4dccf37b5737ea295f9c6ec0b22c4bc6c3c11921c7e1ad6e4cd5b script_sha256: 053b5681918b4d8dc4270d297ddd7803f8806a916f8c4340556cc7610ee1b727 python: 3.12.6 platform: Linux-6.8.0-138-generic-x86_64-with-glibc2.35 host: lab packages: relflow: 0.2.1 torch: 2.12.0 lightning: 2.6.4 numpy: 2.4.6 cuda: '13.0' gpu: NVIDIA GeForce RTX 3090 code_sha256: 7bf47e7f1f7173602f6d5d132f4aee9d0955fbf9e42640b5065a15de127c636e metrics: measured: rmse: 0.018869576500494013 baseline_rmse: 1.2867287797077822 nrmse: 0.014664766031563636 steps: 500 checks: Contribution sum nRMSE below 0.25: true outcome: met duration_seconds: 13.0 P016: script: proofs/aggregation/variable_cardinality_weighted_mean.py page: docs/proofs/aggregation/variable-cardinality-weighted-mean.py historical: status: Provisional pass source_revision: 810b3fe3b0b25c2c7a718571d83f6006f23110c3 sources: - https://github.com/relflow/relflow/blob/810b3fe/proofs/aggregation/weighted_aggregation/test_variable_cardinality_weighted_mean.py - https://github.com/relflow/relflow/blob/810b3fe/proofs/aggregation/weighted_aggregation/support.py summary: |- The checked-in report says every behavioral gate passes at this fixed seed. The result remains provisional until the controls are repeated across seeds. insights: |- **The model responds to relative weights while keeping the same answer when all weights are scaled together.** The model is reported to preserve its answer when complete items are reordered or duplicated, and when all weights are scaled equally. Swapping weights between values instead removes accuracy against the original labels. Those controls distinguish useful invariance from simply ignoring the weights. Keep each weight beside its value, and choose the target’s algebra deliberately: duplicating items preserves a weighted mean but doubles a weighted sum. This proof covers positive weights and lengths within the trained range; a zero denominator, signed weights, or substantially longer bags needs separate treatment. evidence: |- The proof uses deterministic CPU training at seed 3313 for 1,100 updates. The split contains 1,536 training, 384 validation, and 768 test records. The source reports gate clearance but does not record measured scores. nRMSE divides test RMSE by the training-mean baseline error; lower is better. | Check | Required gate | Recorded evidence | | --- | --- | --- | | Held-out weighted-mean nRMSE | < 0.40 | Gate reported passing | | Weight-permutation nRMSE increase | ≥ 0.25 | Gate reported passing | | Complete-item permutation drift | < 0.10 target SD | Gate reported passing | | Weight-scaling drift | < 0.10 target SD | Gate reported passing | | Complete-item duplication drift | < 0.10 target SD | Gate reported passing | runs: - started_at: '2026-09-15T02:14:58.090165+00:00' mode: smoke seed: 3313 steps_override: 2 accelerator: gpu threads: 1 provenance: revision: null source_sha256: 34adfb1bbd259cfa3cffaef8d4b502ab0fa44ab136c8895e33999b05c4508b5c script_sha256: fa07c983ac756a34f954f103000c70969015f2119d2eb3b5f3fd7dda5231a3a5 python: 3.12.6 platform: Linux-6.8.0-138-generic-x86_64-with-glibc2.35 host: lab packages: relflow: 0.2.1 torch: 2.12.0 lightning: 2.6.4 numpy: 2.4.6 cuda: '13.0' gpu: NVIDIA GeForce RTX 3090 code_sha256: 7ac2b0bda39432a83e0715cbc65ab5026aa8ac0724f315dd9c338a98902ebebd metrics: intact: rmse: 0.5037878240325339 baseline_rmse: 0.39181947854115323 nrmse: 1.2857651332401037 corrupted: rmse: 0.5022027721421557 baseline_rmse: 0.39181947854115323 nrmse: 1.2817197705739094 permutation_drift: 0.06346261181616443 scaling_drift: 0.08875849487822701 duplication_drift: 0.09085059851980674 steps: 2 checks: Weighted mean nRMSE below 0.40: false Shuffled weights raise nRMSE by at least 0.25: false Permutation drift below 0.10 target SD: true Weight scaling drift below 0.10 target SD: true Duplication drift below 0.10 target SD: true outcome: not_met duration_seconds: 0.941 - started_at: '2026-09-15T02:22:35.682947+00:00' mode: full seed: 3313 steps_override: null accelerator: gpu threads: 1 provenance: revision: null source_sha256: 34adfb1bbd259cfa3cffaef8d4b502ab0fa44ab136c8895e33999b05c4508b5c script_sha256: fa07c983ac756a34f954f103000c70969015f2119d2eb3b5f3fd7dda5231a3a5 python: 3.12.6 platform: Linux-6.8.0-138-generic-x86_64-with-glibc2.35 host: lab packages: relflow: 0.2.1 torch: 2.12.0 lightning: 2.6.4 numpy: 2.4.6 cuda: '13.0' gpu: NVIDIA GeForce RTX 3090 code_sha256: 7ac2b0bda39432a83e0715cbc65ab5026aa8ac0724f315dd9c338a98902ebebd metrics: intact: rmse: 0.01875336608718387 baseline_rmse: 0.39181947854115323 nrmse: 0.04786226084779546 corrupted: rmse: 0.38477710023252537 baseline_rmse: 0.39181947854115323 nrmse: 0.9820264721528177 permutation_drift: 0.01338920024463998 scaling_drift: 0.029665908104889224 duplication_drift: 0.02135327380034293 steps: 1100 checks: Weighted mean nRMSE below 0.40: true Shuffled weights raise nRMSE by at least 0.25: true Permutation drift below 0.10 target SD: true Weight scaling drift below 0.10 target SD: true Duplication drift below 0.10 target SD: true outcome: met duration_seconds: 38.561 P017: script: proofs/calibration/signal_vs_noise.py page: docs/proofs/calibration/signal-vs-noise.py historical: status: Provisional pass source_revision: 810b3fe3b0b25c2c7a718571d83f6006f23110c3 sources: - https://github.com/relflow/relflow/blob/810b3fe/proofs/calibration/signal_detection/test_signal_vs_noise.py - https://github.com/relflow/relflow/blob/810b3fe/proofs/calibration/signal_detection/support.py summary: |- The source records one paired seed clearing every gate. Exact AUC values are not recorded. The thresholds below are acceptance criteria, not measurements. insights: |- **A meaningful learning check should recognize a real signal and remain at chance when that signal is absent.** The two models have the same schema and training budget; only the relationship between the visible Boolean and the target changes. The recorded run clears both the positive AUC gate and the independent-noise chance band. Separate train, validation, and test streams help distinguish learned relationships from memorized observations. Exact scores are not recorded. Use this as a basic check before interpreting harder proof failures. Copying a Boolean does not establish aggregation or relational reasoning, and the one-seed result does not measure training efficiency. evidence: |- | Evaluation | Required gate | Recorded result | | --- | --- | --- | | Signal AUC | At least 0.98 | Gate reported passing | | Independent-noise AUC | 0.42–0.58 | Gate reported passing | | Signal minus noise AUC | At least 0.40 | Gate reported passing | This calibrates the learning and evaluation harness. It does not establish that a more complicated architecture can recover every relationship. runs: - started_at: '2026-09-15T02:15:02.357151+00:00' mode: smoke seed: 7 steps_override: 2 accelerator: gpu threads: 1 provenance: revision: null source_sha256: 0afefb0737862e2bf07d3de59454b2b86b762ca70c0dfad3a5b6913493288603 script_sha256: 1574fd9a843a615e2237292905ea63bf92efcb5812dbc5382d24c21bb55eb6f5 python: 3.12.6 platform: Linux-6.8.0-138-generic-x86_64-with-glibc2.35 host: lab packages: relflow: 0.2.1 torch: 2.12.0 lightning: 2.6.4 numpy: 2.4.6 cuda: '13.0' gpu: NVIDIA GeForce RTX 3090 code_sha256: 5ae87a38313d48d31f82d6a7244d53a5931db208ccd010b6ac0979f03e2f3bdd metrics: signal_auc: 0.9069156646728516 noise_auc: 0.49146175384521484 auc_gap: 0.4154539108276367 checks: Signal AUC is at least 0.98: false Independent noise AUC remains between 0.42 and 0.58: true Signal exceeds noise by at least 0.40 AUC: true outcome: not_met duration_seconds: 1.264 - started_at: '2026-09-15T02:23:17.604572+00:00' mode: full seed: 7 steps_override: null accelerator: gpu threads: 1 provenance: revision: null source_sha256: 388d96f0024c691fa3c4ea225efa769f77981b7577578bf8077f8e1ca00decb5 script_sha256: 1574fd9a843a615e2237292905ea63bf92efcb5812dbc5382d24c21bb55eb6f5 python: 3.12.6 platform: Linux-6.8.0-138-generic-x86_64-with-glibc2.35 host: lab relflow_path: /home/grantham/relflow-proof-scripts.RB4VJH/src/relflow packages: relflow: 0.2.1 torch: 2.12.0 lightning: 2.6.4 numpy: 2.4.6 pyarrow: 24.0.0 polars: 1.44.1 torchmetrics: 1.9.0 tensordict: 0.12.4 pydantic: 2.13.4 cuda: '13.0' gpu: NVIDIA GeForce RTX 3090 code_sha256: 5ae87a38313d48d31f82d6a7244d53a5931db208ccd010b6ac0979f03e2f3bdd metrics: signal_auc: 1.0 noise_auc: 0.4753994941711426 auc_gap: 0.5246005058288574 checks: Signal AUC is at least 0.98: true Independent noise AUC remains between 0.42 and 0.58: true Signal exceeds noise by at least 0.40 AUC: true outcome: met duration_seconds: 6.596 P018: script: proofs/cluster/hidden_regression_regimes.py page: docs/proofs/cluster/hidden-regression-regimes.py historical: status: Partial evidence source_revision: 810b3fe3b0b25c2c7a718571d83f6006f23110c3 sources: - https://github.com/relflow/relflow/blob/810b3fe/proofs/cluster/cluster_convergence/test_hidden_regression_regimes.py - https://github.com/relflow/relflow/blob/810b3fe/proofs/cluster/cluster_convergence/support.py summary: |- The source records one run meeting the count and usage gates. The experiment does not yet measure held-out regression quality or whether identities were assigned to their correct generating groups. insights: |- **Settling near five clusters does not yet show that the model discovered the five generating functions.** Repeated identities follow stable regimes while their numeric inputs vary; reconstruction engages adaptive Cluster state alongside the supervised numeric task. The recorded gates inspect terminal count and usage. Count is calculated from usage perplexity, so those diagnostics are coupled. Neither tells us whether identities share the correct regime or whether predictions work on independent observations. The current train and validation tables are identical. Evaluate held-out regression and assignment agreement before treating this as regime discovery. The existing result supports adaptive-state behavior, with the harder behavioral claim still open. evidence: |- | Diagnostic | Required gate | Recorded result | | --- | --- | --- | | Committed count in each final epoch | 4–7 for true count 5 | Gate reported passing | | Usage perplexity in each final epoch | 3.5–6.5 | Gate reported passing | | Terminal committed count | Neither lower bound 3 nor upper bound 15 | Gate reported passing | The evidence concerns adaptive state. Settling on approximately five groups does not establish that the model discovered the five functions correctly. runs: - started_at: '2026-09-15T02:15:06.979855+00:00' mode: smoke seed: 0 steps_override: 2 accelerator: gpu threads: 1 provenance: revision: null source_sha256: a855289777003bf0ec18ebf59ad0bc9a6180205fca745c1a42baa8cbd5ff521b script_sha256: 748439ffc5099809de9b0ce4b32dc2d46115de2a8ef143ffed3488546694fd93 python: 3.12.6 platform: Linux-6.8.0-138-generic-x86_64-with-glibc2.35 host: lab packages: relflow: 0.2.1 torch: 2.12.0 lightning: 2.6.4 numpy: 2.4.6 cuda: '13.0' gpu: NVIDIA GeForce RTX 3090 code_sha256: 51ce395948cba002dc428923ce333f93b6b4cb198ce5be247c92ccf5041773c9 metrics: trajectory: - epoch: 0 n_committed: 3 perplexity: 3.258863687515259 adherence: 0.012305669486522675 final_committed: - 3 final_perplexity: - 3.258863687515259 checks: Final committed counts remain between 4 and 7: false Final perplexity remains within 1.5 of five: false Terminal commitment is above the lower bound: false Terminal commitment is below the upper bound: true outcome: not_met duration_seconds: 0.645 - started_at: '2026-09-15T02:23:27.629749+00:00' mode: full seed: 0 steps_override: null accelerator: gpu threads: 1 provenance: revision: null source_sha256: 4035bdca26e7f43662ce98fbb02dc392c9435b4288004b86dd61d42aa111ed8a script_sha256: 748439ffc5099809de9b0ce4b32dc2d46115de2a8ef143ffed3488546694fd93 python: 3.12.6 platform: Linux-6.8.0-138-generic-x86_64-with-glibc2.35 host: lab relflow_path: /home/grantham/relflow-proof-scripts.RB4VJH/src/relflow packages: relflow: 0.2.1 torch: 2.12.0 lightning: 2.6.4 numpy: 2.4.6 pyarrow: 24.0.0 polars: 1.44.1 torchmetrics: 1.9.0 tensordict: 0.12.4 pydantic: 2.13.4 cuda: '13.0' gpu: NVIDIA GeForce RTX 3090 code_sha256: 51ce395948cba002dc428923ce333f93b6b4cb198ce5be247c92ccf5041773c9 metrics: trajectory: - epoch: 0 n_committed: 4 perplexity: 4.372092247009277 adherence: 0.05254526063799858 - epoch: 1 n_committed: 8 perplexity: 8.402069091796875 adherence: 0.0076850783079862595 - epoch: 2 n_committed: 7 perplexity: 7.199680805206299 adherence: 0.012563434429466724 - epoch: 3 n_committed: 6 perplexity: 6.230902671813965 adherence: 0.015585856512188911 - epoch: 4 n_committed: 6 perplexity: 6.341165065765381 adherence: 0.014966081827878952 - epoch: 5 n_committed: 6 perplexity: 6.3358659744262695 adherence: 0.01352382730692625 - epoch: 6 n_committed: 6 perplexity: 6.37974214553833 adherence: 0.01200682483613491 - epoch: 7 n_committed: 6 perplexity: 6.326610565185547 adherence: 0.01046148780733347 - epoch: 8 n_committed: 6 perplexity: 6.308366775512695 adherence: 0.00900269579142332 - epoch: 9 n_committed: 6 perplexity: 6.272604465484619 adherence: 0.007946990430355072 - epoch: 10 n_committed: 6 perplexity: 6.247581481933594 adherence: 0.0070823985151946545 - epoch: 11 n_committed: 6 perplexity: 6.185416221618652 adherence: 0.006402456201612949 - epoch: 12 n_committed: 6 perplexity: 6.179567813873291 adherence: 0.005967267323285341 - epoch: 13 n_committed: 6 perplexity: 6.129400730133057 adherence: 0.005367706995457411 - epoch: 14 n_committed: 6 perplexity: 6.055844306945801 adherence: 0.004808566067367792 - epoch: 15 n_committed: 6 perplexity: 5.836822986602783 adherence: 0.004326104186475277 - epoch: 16 n_committed: 6 perplexity: 5.672842502593994 adherence: 0.0036972553934901953 - epoch: 17 n_committed: 6 perplexity: 5.597299575805664 adherence: 0.0031334117520600557 - epoch: 18 n_committed: 6 perplexity: 5.673702716827393 adherence: 0.0026480776723474264 - epoch: 19 n_committed: 6 perplexity: 5.611253261566162 adherence: 0.002550099976360798 - epoch: 20 n_committed: 6 perplexity: 5.72269344329834 adherence: 0.0024596890434622765 - epoch: 21 n_committed: 6 perplexity: 5.699801445007324 adherence: 0.0021622052881866693 - epoch: 22 n_committed: 6 perplexity: 5.627553939819336 adherence: 0.0018641826463863254 - epoch: 23 n_committed: 6 perplexity: 5.53046178817749 adherence: 0.0016020906623452902 - epoch: 24 n_committed: 5 perplexity: 5.434558868408203 adherence: 0.0013695001835003495 - epoch: 25 n_committed: 5 perplexity: 5.346553325653076 adherence: 0.001190957147628069 - epoch: 26 n_committed: 5 perplexity: 5.26736307144165 adherence: 0.0010388604132458568 - epoch: 27 n_committed: 5 perplexity: 5.203525543212891 adherence: 0.0009039812721312046 - epoch: 28 n_committed: 5 perplexity: 5.130213260650635 adherence: 0.0008230844978243113 - epoch: 29 n_committed: 5 perplexity: 5.073955535888672 adherence: 0.0010630140313878655 final_committed: - 5 - 5 - 5 - 5 - 5 final_perplexity: - 5.346553325653076 - 5.26736307144165 - 5.203525543212891 - 5.130213260650635 - 5.073955535888672 checks: Final committed counts remain between 4 and 7: true Final perplexity remains within 1.5 of five: true Terminal commitment is above the lower bound: true Terminal commitment is below the upper bound: true outcome: met duration_seconds: 23.844 P019: script: proofs/cluster/plain_input_is_dormant.py page: docs/proofs/cluster/plain-input-is-dormant.py historical: status: Expected limitation source_revision: 810b3fe3b0b25c2c7a718571d83f6006f23110c3 sources: - https://github.com/relflow/relflow/blob/810b3fe/proofs/cluster/cluster_convergence/test_plain_input_is_dormant.py - https://github.com/relflow/relflow/blob/810b3fe/proofs/cluster/cluster_convergence/support.py summary: |- The recorded test passes by observing unchanged commitment and zero adherence. Those are the expected signs of a dormant mechanism, not evidence that the model recovered a five-group partition. insights: |- **A prediction target beside a Cluster field does not activate that field’s adaptive grouping.** The identities repeat and carry stable labels, but the Cluster field is only an input. The expected result is unchanged committed count and zero adherence, which the source reports observing. This does not mean every parameter or the label decoder is frozen: the assertions concern these specific adaptive-state diagnostics. No label-prediction score is established here. When adaptive grouping is intended, add a reconstructing mask to the Cluster field before changing capacity or training longer. Then check predictive usefulness and assignment quality separately; merely activating commitment does not establish either. evidence: |- | Diagnostic | Required gate | Recorded result | | --- | --- | --- | | Committed count over all epochs | Constant | Reported unchanged | | Adherence over all epochs | Exactly 0 | Reported zero | The check makes no claim about the label decoder's predictive quality or about every model parameter remaining fixed. Its scope is adaptive Cluster state. runs: - started_at: '2026-09-15T02:15:10.965587+00:00' mode: smoke seed: 0 steps_override: 2 accelerator: gpu threads: 1 provenance: revision: null source_sha256: 96f47589ff173377f03085e26c207c147dbdbb906f6a8d285d4b41d41a030da9 script_sha256: 725a199da0187cc8dbc2ce31fcafbe5303510eb081793c27a2c57aa79b7ab00c python: 3.12.6 platform: Linux-6.8.0-138-generic-x86_64-with-glibc2.35 host: lab packages: relflow: 0.2.1 torch: 2.12.0 lightning: 2.6.4 numpy: 2.4.6 cuda: '13.0' gpu: NVIDIA GeForce RTX 3090 code_sha256: 854f5153350c97a00943ff12741d68980750a65a947d818771b3e86d080bf2a8 metrics: trajectory: - epoch: 0 n_committed: 3 perplexity: 3.0 adherence: 0.0 committed: - 3 adherence: - 0.0 checks: Committed count remains constant: true Adherence remains exactly zero: true outcome: met duration_seconds: 0.585 - started_at: '2026-09-15T02:23:54.876896+00:00' mode: full seed: 0 steps_override: null accelerator: gpu threads: 1 provenance: revision: null source_sha256: 49e7b928e110a6f0c7b138740d2cb92f4973b4a27a8bb4ed8a2e38831e9ab2f7 script_sha256: 725a199da0187cc8dbc2ce31fcafbe5303510eb081793c27a2c57aa79b7ab00c python: 3.12.6 platform: Linux-6.8.0-138-generic-x86_64-with-glibc2.35 host: lab relflow_path: /home/grantham/relflow-proof-scripts.RB4VJH/src/relflow packages: relflow: 0.2.1 torch: 2.12.0 lightning: 2.6.4 numpy: 2.4.6 pyarrow: 24.0.0 polars: 1.44.1 torchmetrics: 1.9.0 tensordict: 0.12.4 pydantic: 2.13.4 cuda: '13.0' gpu: NVIDIA GeForce RTX 3090 code_sha256: 854f5153350c97a00943ff12741d68980750a65a947d818771b3e86d080bf2a8 metrics: trajectory: - epoch: 0 n_committed: 3 perplexity: 3.0 adherence: 0.0 - epoch: 1 n_committed: 3 perplexity: 3.0 adherence: 0.0 - epoch: 2 n_committed: 3 perplexity: 3.0 adherence: 0.0 - epoch: 3 n_committed: 3 perplexity: 3.0 adherence: 0.0 - epoch: 4 n_committed: 3 perplexity: 3.0 adherence: 0.0 committed: - 3 - 3 - 3 - 3 - 3 adherence: - 0.0 - 0.0 - 0.0 - 0.0 - 0.0 checks: Committed count remains constant: true Adherence remains exactly zero: true outcome: met duration_seconds: 2.331 P020: script: proofs/cluster/reconstructing_category_labels.py page: docs/proofs/cluster/reconstructing-category-labels.py historical: status: Partial evidence source_revision: 810b3fe3b0b25c2c7a718571d83f6006f23110c3 sources: - https://github.com/relflow/relflow/blob/810b3fe/proofs/cluster/cluster_convergence/test_reconstructing_category_labels.py - https://github.com/relflow/relflow/blob/810b3fe/proofs/cluster/cluster_convergence/support.py summary: |- The recorded single-seed run settles near five clusters. Training and validation reuse the same observations, and the test checks neither held-out prediction nor recovery of the actual identity partition. insights: |- **The right number of clusters does not mean the right identities were grouped together.** Repeated identities provide consistent label evidence, and the Cluster reconstruction loss updates usage and commitment. The recorded run ends near five groups. However, committed count is derived by rounding and clamping usage perplexity: these are coupled diagnostics, not independent confirmations of recovery. Incorrect groups can satisfy both gates. Training and validation reuse observations, with no held-out skill or assignment-agreement requirement. Treat this as evidence about mechanism activation. Demonstrating useful clustering still requires independent predictions, partition recovery, and controls without genuine generating groups. evidence: |- | Diagnostic | Required gate | Recorded result | | --- | --- | --- | | Committed count in each final epoch | 4–7 for true count 5 | Gate reported passing | | Usage perplexity in each final epoch | 3.5–6.5 | Gate reported passing | | Terminal committed count | Neither lower bound 3 nor upper bound 15 | Gate reported passing | Usage perplexity summarizes how many clusters carry meaningful assignment mass. A plausible count and perplexity do not prove that the right identities share a cluster. runs: - started_at: '2026-09-15T02:15:14.899120+00:00' mode: smoke seed: 0 steps_override: 2 accelerator: gpu threads: 1 provenance: revision: null source_sha256: 95f6a46ba99894f901600a04312fc1fd1f3354b5c362210ca1a213db64cdbfbc script_sha256: b327e09d1524c7eddc51d0c48472f183727b30cf364bd72ce1fee9822f7161be python: 3.12.6 platform: Linux-6.8.0-138-generic-x86_64-with-glibc2.35 host: lab packages: relflow: 0.2.1 torch: 2.12.0 lightning: 2.6.4 numpy: 2.4.6 cuda: '13.0' gpu: NVIDIA GeForce RTX 3090 code_sha256: af5644a349eddf217b9c669653bbef5be2fd4150e115838d2e4027b40a8c668c metrics: trajectory: - epoch: 0 n_committed: 3 perplexity: 3.3275811672210693 adherence: 0.01613517478108406 final_committed: - 3 final_perplexity: - 3.3275811672210693 checks: Final committed counts remain between 4 and 7: false Final perplexity remains within 1.5 of five: false Terminal commitment is above the lower bound: false Terminal commitment is below the upper bound: true outcome: not_met duration_seconds: 0.657 - started_at: '2026-09-15T02:24:00.576894+00:00' mode: full seed: 0 steps_override: null accelerator: gpu threads: 1 provenance: revision: null source_sha256: ad145610d04bc8af8d40fb104042fd9aa3abb73ef86d4fade1cf1001bc933df5 script_sha256: b327e09d1524c7eddc51d0c48472f183727b30cf364bd72ce1fee9822f7161be python: 3.12.6 platform: Linux-6.8.0-138-generic-x86_64-with-glibc2.35 host: lab relflow_path: /home/grantham/relflow-proof-scripts.RB4VJH/src/relflow packages: relflow: 0.2.1 torch: 2.12.0 lightning: 2.6.4 numpy: 2.4.6 pyarrow: 24.0.0 polars: 1.44.1 torchmetrics: 1.9.0 tensordict: 0.12.4 pydantic: 2.13.4 cuda: '13.0' gpu: NVIDIA GeForce RTX 3090 code_sha256: af5644a349eddf217b9c669653bbef5be2fd4150e115838d2e4027b40a8c668c metrics: trajectory: - epoch: 0 n_committed: 4 perplexity: 4.334924221038818 adherence: 0.05807165056467056 - epoch: 1 n_committed: 7 perplexity: 6.767301559448242 adherence: 0.008113355375826359 - epoch: 2 n_committed: 6 perplexity: 6.064903259277344 adherence: 0.01602412946522236 - epoch: 3 n_committed: 5 perplexity: 5.252449035644531 adherence: 0.020100772380828857 - epoch: 4 n_committed: 5 perplexity: 5.424206256866455 adherence: 0.019789520651102066 - epoch: 5 n_committed: 5 perplexity: 5.493758678436279 adherence: 0.018425457179546356 - epoch: 6 n_committed: 6 perplexity: 5.526225566864014 adherence: 0.016715876758098602 - epoch: 7 n_committed: 6 perplexity: 5.5231709480285645 adherence: 0.015048263594508171 - epoch: 8 n_committed: 6 perplexity: 5.506030082702637 adherence: 0.01349477469921112 - epoch: 9 n_committed: 6 perplexity: 6.471004009246826 adherence: 0.0005394776235334575 - epoch: 10 n_committed: 5 perplexity: 5.432131290435791 adherence: 0.0026502199470996857 - epoch: 11 n_committed: 5 perplexity: 5.432953357696533 adherence: 0.003791071241721511 - epoch: 12 n_committed: 5 perplexity: 5.4138641357421875 adherence: 0.004280708730220795 - epoch: 13 n_committed: 5 perplexity: 5.387415409088135 adherence: 0.0044521489180624485 - epoch: 14 n_committed: 5 perplexity: 5.35729455947876 adherence: 0.004438720177859068 - epoch: 15 n_committed: 5 perplexity: 5.324090003967285 adherence: 0.00431922497227788 - epoch: 16 n_committed: 5 perplexity: 5.293123245239258 adherence: 0.004147147759795189 - epoch: 17 n_committed: 5 perplexity: 5.265658855438232 adherence: 0.003949499223381281 - epoch: 18 n_committed: 5 perplexity: 5.241479873657227 adherence: 0.0037347483448684216 - epoch: 19 n_committed: 5 perplexity: 5.218961715698242 adherence: 0.0035151427146047354 - epoch: 20 n_committed: 5 perplexity: 5.1984944343566895 adherence: 0.003297733375802636 - epoch: 21 n_committed: 5 perplexity: 5.180423259735107 adherence: 0.003088444471359253 - epoch: 22 n_committed: 5 perplexity: 5.1614885330200195 adherence: 0.002890538889914751 - epoch: 23 n_committed: 5 perplexity: 5.146783828735352 adherence: 0.002702163066715002 - epoch: 24 n_committed: 5 perplexity: 5.133707046508789 adherence: 0.0025267833843827248 - epoch: 25 n_committed: 5 perplexity: 5.120759963989258 adherence: 0.002360387472435832 - epoch: 26 n_committed: 5 perplexity: 5.110231399536133 adherence: 0.002208678750321269 - epoch: 27 n_committed: 5 perplexity: 5.101234436035156 adherence: 0.002066380809992552 - epoch: 28 n_committed: 5 perplexity: 5.0931925773620605 adherence: 0.0019342011073604226 - epoch: 29 n_committed: 5 perplexity: 5.0839056968688965 adherence: 0.001811657682992518 final_committed: - 5 - 5 - 5 - 5 - 5 final_perplexity: - 5.120759963989258 - 5.110231399536133 - 5.101234436035156 - 5.0931925773620605 - 5.0839056968688965 checks: Final committed counts remain between 4 and 7: true Final perplexity remains within 1.5 of five: true Terminal commitment is above the lower bound: true Terminal commitment is below the upper bound: true outcome: met duration_seconds: 14.894 P021: script: proofs/identity/unseen_identity_equality.py page: docs/proofs/identity/unseen-identity-equality.py historical: status: Provisional pass source_revision: 810b3fe3b0b25c2c7a718571d83f6006f23110c3 sources: - https://github.com/relflow/relflow/blob/810b3fe/proofs/identity/hash_equality/test_unseen_identity_equality.py - https://github.com/relflow/relflow/blob/810b3fe/proofs/identity/hash_equality/support.py summary: |- The source records one seed clearing the Hash equality gate and both negative controls. It reports passing thresholds, without exact AUC measurements. insights: |- **Hash can preserve equality across fields without learning each identifier in advance.** Compatible batch-local representations let matching unseen strings supply a comparison signal to the shared model. The recorded Hash model clears the held-out equality gate on a disjoint namespace. Shuffling labels removes that advantage, while matched Category fields cannot reliably distinguish unseen strings after their content becomes unavailable. These are passing thresholds, not recorded exact AUC values. Use Hash when observation-local equality matters more than persistent category semantics. This result does not establish invertible identifiers, stable embeddings across separately encoded batches, or general relational joins. Prediction stability under a shared identity renaming remains untested. evidence: |- | Evaluation | Required gate | Recorded result | | --- | --- | --- | | Hash on unseen identities | AUC ≥ 0.95 | Gate reported passing | | Shuffled test labels | AUC 0.42–0.58 | Gate reported passing | | Category on unseen identities | AUC ≤ 0.65 | Gate reported passing | | Hash minus shuffled-label AUC | At least 0.35 | Gate reported passing | This establishes a provisional route for equality, not an invertible or persistent identity representation. Hash collisions remain possible. runs: - started_at: '2026-09-15T02:15:18.910151+00:00' mode: smoke seed: 29 steps_override: 2 accelerator: gpu threads: 1 provenance: revision: null source_sha256: 9e3113cf14a05d6faf0c703bba652d39193e5711836e6f170227bee35b841090 script_sha256: 3ab7430e2f102ab4f801c636b7c4cde373225f5577851b5f6fd468bf26594b12 python: 3.12.6 platform: Linux-6.8.0-138-generic-x86_64-with-glibc2.35 host: lab packages: relflow: 0.2.1 torch: 2.12.0 lightning: 2.6.4 numpy: 2.4.6 cuda: '13.0' gpu: NVIDIA GeForce RTX 3090 code_sha256: 1feddae272636fdf4e05b9cad1b0b4f610ad7b055f866a29fd4f8dc04736f3ba metrics: hash_auc: 0.510551929473877 shuffled_auc: 0.4934225082397461 category_oov_auc: 0.5 auc_gap: 0.01712942123413086 checks: Unseen Hash equality AUC is at least 0.95: false Shuffled labels remain between 0.42 and 0.58 AUC: true Category OOV AUC is at most 0.65: true Hash exceeds shuffled labels by at least 0.35 AUC: false outcome: not_met duration_seconds: 1.711 - started_at: '2026-09-15T02:24:18.855867+00:00' mode: full seed: 29 steps_override: null accelerator: gpu threads: 1 provenance: revision: null source_sha256: f8f9e5eafa2d1516fa31c67b12cdedad7e96f80290c0c5ed32f0b6300cb6c0b0 script_sha256: 3ab7430e2f102ab4f801c636b7c4cde373225f5577851b5f6fd468bf26594b12 python: 3.12.6 platform: Linux-6.8.0-138-generic-x86_64-with-glibc2.35 host: lab relflow_path: /home/grantham/relflow-proof-scripts.RB4VJH/src/relflow packages: relflow: 0.2.1 torch: 2.12.0 lightning: 2.6.4 numpy: 2.4.6 pyarrow: 24.0.0 polars: 1.44.1 torchmetrics: 1.9.0 tensordict: 0.12.4 pydantic: 2.13.4 cuda: '13.0' gpu: NVIDIA GeForce RTX 3090 code_sha256: 1feddae272636fdf4e05b9cad1b0b4f610ad7b055f866a29fd4f8dc04736f3ba metrics: hash_auc: 0.9998767375946045 shuffled_auc: 0.49859464168548584 category_oov_auc: 0.5 auc_gap: 0.5012820959091187 checks: Unseen Hash equality AUC is at least 0.95: true Shuffled labels remain between 0.42 and 0.58 AUC: true Category OOV AUC is at most 0.65: true Hash exceeds shuffled labels by at least 0.35 AUC: true outcome: met duration_seconds: 29.041 P022: script: proofs/relational/attention_number_payload.py page: docs/proofs/relational/attention-number-payload.py historical: status: Provisional pass source_revision: 810b3fe3b0b25c2c7a718571d83f6006f23110c3 sources: - https://github.com/relflow/relflow/blob/810b3fe/proofs/relational/argmax_retrieval/test_attention_number_payload.py - https://github.com/relflow/relflow/blob/810b3fe/proofs/relational/argmax_retrieval/__init__.py summary: |- The checked-in report records approximately **0.30 nRMSE** on one deterministic CPU seed. The matched pass-through route records approximately 0.16. Neither result establishes capacity for longer collections. insights: |- **A compressed summary can preserve enough information to retrieve the value attached to the highest score.** The summary slots do not represent named candidates. The item branch reduces first, so requesting several root outputs cannot recreate information that the earlier summary failed to retain. The payload-rotation control removes accuracy while preserving score and payload marginals. That supports relational information surviving the tested route, rather than success from collection statistics alone. The retained-token route records better accuracy on the matched task, but this comparison supplies no speed or memory measurements. Treat summary width as task-specific capacity; longer collections, other payload types, and exact retrieval remain separate questions. evidence: |- Both routes use identical splits: 4,096 training, 1,024 validation, and 2,048 test records, trained for 25 CPU epochs. nRMSE uses the training-mean baseline. | Check | Required gate | Recorded result | | --- | --- | --- | | Intact retrieval | nRMSE ≤ 0.35 | Approximately 0.30 | | Rotated payloads | nRMSE ≥ 0.90 | Gate reported passing | | Corruption gap | At least 0.35 nRMSE | Gate reported passing | runs: - started_at: '2026-09-15T02:15:24.019056+00:00' mode: smoke seed: 23 steps_override: 2 accelerator: gpu threads: 1 provenance: revision: null source_sha256: 5575a43226d5d7d4b349e101e3f98d57f0a7362da711f5a777c3495a3ed6e93b script_sha256: 503311defddbad62bd63fbe125a23ef712b2bd1cee711fb96c5af2b2f32e3f09 python: 3.12.6 platform: Linux-6.8.0-138-generic-x86_64-with-glibc2.35 host: lab packages: relflow: 0.2.1 torch: 2.12.0 lightning: 2.6.4 numpy: 2.4.6 cuda: '13.0' gpu: NVIDIA GeForce RTX 3090 metrics: rmse: 1.00318426502101 baseline_rmse: 0.9911584043638657 intact_nrmse: 1.012133136947835 broken_nrmse: 1.012690002090478 checks: Intact retrieval nRMSE <= 0.35: false Broken pairing nRMSE >= 0.90: true Breaking pairing increases nRMSE by >= 0.35: false outcome: not_met duration_seconds: 0.903 - started_at: '2026-09-15T02:24:51.311825+00:00' mode: full seed: 23 steps_override: null accelerator: gpu threads: 1 provenance: revision: null source_sha256: d57439096e83b67b4fab821b5fed80a4aef76a6bf15860436d866f012b3b7b54 script_sha256: 9daa6891a91d7daa86e24ef8862f4ac97576ae382b6c2ff2a418e90dffc2cf00 python: 3.12.6 platform: Linux-6.8.0-138-generic-x86_64-with-glibc2.35 host: lab relflow_path: /home/grantham/relflow-proof-scripts.RB4VJH/src/relflow packages: relflow: 0.2.1 torch: 2.12.0 lightning: 2.6.4 numpy: 2.4.6 pyarrow: 24.0.0 polars: 1.44.1 torchmetrics: 1.9.0 tensordict: 0.12.4 pydantic: 2.13.4 cuda: '13.0' gpu: NVIDIA GeForce RTX 3090 code_sha256: a064d7690ae53ac7f06d95845d3d3d56897664bae9cb61fd1f68591d032878a9 metrics: rmse: 0.2813291196788821 baseline_rmse: 0.9911584043638657 intact_nrmse: 0.2838387067498475 broken_nrmse: 1.3881021097428687 checks: Intact retrieval nRMSE <= 0.35: true Broken pairing nRMSE >= 0.90: true Breaking pairing increases nRMSE by >= 0.35: true outcome: met duration_seconds: 19.478 P023: script: proofs/relational/pass_through_number_payload.py page: docs/proofs/relational/pass-through-number-payload.py historical: status: Provisional pass source_revision: 810b3fe3b0b25c2c7a718571d83f6006f23110c3 sources: - https://github.com/relflow/relflow/blob/810b3fe/proofs/relational/argmax_retrieval/test_pass_through_number_payload.py - https://github.com/relflow/relflow/blob/810b3fe/proofs/relational/argmax_retrieval/__init__.py summary: |- The checked-in report records approximately **0.16 nRMSE** for three candidates on one deterministic CPU seed. Rotating payloads removes the retrieval skill. These are recorded results, not a fresh run. insights: |- **The model learns which payload belongs to the highest score, rather than merely predicting a statistic of the whole collection.** All candidate scores and payloads are visible, and their shared item coordinates preserve the pairing. The root answer is hidden; its scalar decoder uses the retained encoded evidence to produce one Number. Rotating payloads preserves both marginal distributions but breaks the original answer's association with its score. The resulting loss of accuracy supports use of that pairing. It does not establish exact argmax computation or behavior beyond the tested three candidates. Preserve candidate evidence when diagnosing similar learned retrieval tasks; compute a known selection rule directly when its result must be exact. evidence: |- The protocol uses 4,096 training, 1,024 validation, and 2,048 test records over 25 CPU epochs. nRMSE divides error by the training-mean predictor's test error. | Check | Required gate | Recorded result | | --- | --- | --- | | Intact retrieval | nRMSE ≤ 0.35 | Approximately 0.16 | | Rotated payloads | nRMSE ≥ 0.90 | Gate reported passing | | Corruption gap | At least 0.35 nRMSE | Gate reported passing | runs: - started_at: '2026-09-15T02:15:28.285395+00:00' mode: smoke seed: 23 steps_override: 2 accelerator: gpu threads: 1 provenance: revision: null source_sha256: 900cce04e87494f814ea387300d091bc0dbd5d6e5061949b90fe0eee152a6865 script_sha256: db70cebd0a97d453687fc2419f2aaead18fc65d0c8ecb49a2bf7b681c8f15edc python: 3.12.6 platform: Linux-6.8.0-138-generic-x86_64-with-glibc2.35 host: lab packages: relflow: 0.2.1 torch: 2.12.0 lightning: 2.6.4 numpy: 2.4.6 cuda: '13.0' gpu: NVIDIA GeForce RTX 3090 metrics: rmse: 0.9003777396122906 baseline_rmse: 0.9911584043638657 intact_nrmse: 0.9084095293427503 broken_nrmse: 0.9110087154088085 train_nrmse: 0.9087091985224002 checks: Intact retrieval nRMSE <= 0.35: false Broken pairing nRMSE >= 0.90: true Breaking pairing increases nRMSE by >= 0.35: false outcome: not_met duration_seconds: 1.165 - started_at: '2026-09-15T02:25:14.219811+00:00' mode: full seed: 23 steps_override: null accelerator: gpu threads: 1 provenance: revision: null source_sha256: bf49f8710d19a73ca107a284a243d846ed29b9f4b97d9db0738ff1a0e43b2e57 script_sha256: 04959a39d9d1a7129efdf2eb244fd0f24d7636490f894aa9bf1fb3f5441a34f3 python: 3.12.6 platform: Linux-6.8.0-138-generic-x86_64-with-glibc2.35 host: lab relflow_path: /home/grantham/relflow-proof-scripts.RB4VJH/src/relflow packages: relflow: 0.2.1 torch: 2.12.0 lightning: 2.6.4 numpy: 2.4.6 pyarrow: 24.0.0 polars: 1.44.1 torchmetrics: 1.9.0 tensordict: 0.12.4 pydantic: 2.13.4 cuda: '13.0' gpu: NVIDIA GeForce RTX 3090 code_sha256: b6b705dfe51f8869b4e15f0009a20ba2c1bd05167b35c59d1a96ffbc4500de1c metrics: rmse: 0.2512988007621003 baseline_rmse: 0.9911584043638657 intact_nrmse: 0.2535405033702823 broken_nrmse: 1.372341575993056 train_nrmse: 0.25405967626833875 checks: Intact retrieval nRMSE <= 0.35: true Broken pairing nRMSE >= 0.90: true Breaking pairing increases nRMSE by >= 0.35: true outcome: met duration_seconds: 16.776 P024: script: proofs/relational/aligned_position_control.py page: docs/proofs/relational/aligned-position-control.py historical: status: Provisional pass source_revision: 810b3fe3b0b25c2c7a718571d83f6006f23110c3 sources: - https://github.com/relflow/relflow/blob/810b3fe/proofs/relational/associative_recall/test_aligned_position_control.py - https://github.com/relflow/relflow/blob/810b3fe/proofs/relational/associative_recall/__init__.py summary: |- The control is reported passing on one deterministic CPU seed with both valid and broken keys. That is evidence for positional copying, not associative recall. The [shuffled recall proof](hash-recall.qmd) tests identity dependence. insights: |- **A model can copy the right value using position while ignoring matching identities.** Aligned source and query positions give the model another way to recover every target. Rotating query keys while retaining target positions leaves accuracy high, so this control does not establish identity-based recall. Repeated decoder queries retain positional information. Although branch and root reductions are compressed, decoder memory also includes the target value field's own parcel, which still contains visible source values. Success therefore does not show that one summary preserves all source information. Keep this control beside shuffled-key recall: their different responses to broken keys distinguish copying by position from using identity. evidence: |- The protocol fits 2,048 records, validates on 512, and tests 1,024 after 800 CPU steps. nRMSE is relative to predicting zero for every query value. | Check | Required gate | Recorded result | | --- | --- | --- | | Aligned copying | nRMSE ≤ 0.25 | Gate reported passing | | Broken query keys | nRMSE ≤ 0.25 | Gate reported passing | | Change after key rotation | Absolute nRMSE difference ≤ 0.10 | Gate reported passing | runs: - started_at: '2026-09-15T02:15:32.810216+00:00' mode: smoke seed: 17 steps_override: 2 accelerator: gpu threads: 1 provenance: revision: null source_sha256: 8066d228439b45ab2bd515e8eb74119d1629cc4bf26c039c89fc62018f46b990 script_sha256: 7f440546bb98a590a86b30285c8c3dea47d431040545088512c6f747f78cedeb python: 3.12.6 platform: Linux-6.8.0-138-generic-x86_64-with-glibc2.35 host: lab packages: relflow: 0.2.1 torch: 2.12.0 lightning: 2.6.4 numpy: 2.4.6 cuda: '13.0' gpu: NVIDIA GeForce RTX 3090 metrics: intact_nrmse: 1.3284462697596782 broken_nrmse: 1.319193460072285 corruption_gap: -0.009252809687393171 checks: Aligned copying nRMSE <= 0.25: false Broken identities retain copying nRMSE <= 0.25: false Identity change alters nRMSE by <= 0.10: true outcome: not_met duration_seconds: 1.028 - started_at: '2026-09-15T02:25:34.435010+00:00' mode: full seed: 17 steps_override: null accelerator: gpu threads: 1 provenance: revision: null source_sha256: c2018887f84ae8aded1142132631ffa7a366bb5f242a5bdb153b12e9b6909eef script_sha256: 678769f758499d0f3f88ad57c6934bd7e0beda583ab7d3c153558c30e787041b python: 3.12.6 platform: Linux-6.8.0-138-generic-x86_64-with-glibc2.35 host: lab relflow_path: /home/grantham/relflow-proof-scripts.RB4VJH/src/relflow packages: relflow: 0.2.1 torch: 2.12.0 lightning: 2.6.4 numpy: 2.4.6 pyarrow: 24.0.0 polars: 1.44.1 torchmetrics: 1.9.0 tensordict: 0.12.4 pydantic: 2.13.4 cuda: '13.0' gpu: NVIDIA GeForce RTX 3090 code_sha256: fca216695f5ed371bc7347114e5c00e12a80a72cffac9a4b455786435ddc326c metrics: intact_nrmse: 0.023022250453277816 broken_nrmse: 0.022929462165871755 corruption_gap: -9.278828740606065e-05 checks: Aligned copying nRMSE <= 0.25: true Broken identities retain copying nRMSE <= 0.25: true Identity change alters nRMSE by <= 0.10: true outcome: met duration_seconds: 21.143 P025: script: proofs/relational/hash_recall.py page: docs/proofs/relational/hash-recall.py historical: status: Provisional pass source_revision: 810b3fe3b0b25c2c7a718571d83f6006f23110c3 sources: - https://github.com/relflow/relflow/blob/810b3fe/proofs/relational/associative_recall/test_hash_recall.py - https://github.com/relflow/relflow/blob/810b3fe/proofs/relational/associative_recall/__init__.py summary: |- The checked-in report records passing accuracy and broken-key gates for one deterministic CPU seed with two source/query pairs. It does not report exact nRMSE values or a larger supported memory capacity. insights: |- **The model recalls values for unseen keys, and breaking the key relationship removes that skill.** Fresh values and independently shuffled records prevent fixed positions or remembered answers from reliably solving the task. Hash preserves equality without a learned vocabulary; the Boolean query mask hides only query values, leaving source values visible. The decoder can use both retained ancestor context and the value field's own visible source parcel. Aligned identity and role fields condition each repeated query. Rotating query keys while retaining original targets tests whether that relationship matters. This supports learned recall for two pairs, not a deterministic lookup guarantee. Larger memories, absent or duplicate keys, and collisions remain untested. evidence: |- The protocol uses 2,048 training, 512 validation, and 1,024 test records with 800 CPU training steps. nRMSE uses a zero-value predictor as its baseline. | Check | Required gate | Recorded result | | --- | --- | --- | | Unseen-key recall | nRMSE ≤ 0.25 | Gate reported passing | | Broken query keys | nRMSE ≥ 0.80 | Gate reported passing | | Broken-key gap | At least 0.50 nRMSE | Gate reported passing | runs: - started_at: '2026-09-15T02:15:37.206852+00:00' mode: smoke seed: 17 steps_override: 2 accelerator: gpu threads: 1 provenance: revision: null source_sha256: b85144ea5b72ef48269097a25fdc5a83d0aae6b270e71bd3428c4c4a6ccee51c script_sha256: bd11e0de7a0290b37c65f046dc14ee373653042e641ad38ecca1c55a4b9167c3 python: 3.12.6 platform: Linux-6.8.0-138-generic-x86_64-with-glibc2.35 host: lab packages: relflow: 0.2.1 torch: 2.12.0 lightning: 2.6.4 numpy: 2.4.6 cuda: '13.0' gpu: NVIDIA GeForce RTX 3090 metrics: intact_nrmse: 1.3100486307342105 broken_nrmse: 1.3160914357450295 corruption_gap: 0.006042805010818997 checks: Finite errors: true Recall calibration nRMSE <= 0.90: false Identity recall nRMSE <= 0.25: false Broken identities nRMSE >= 0.80: true Corruption increases nRMSE by >= 0.50: false outcome: not_met duration_seconds: 1.321 - started_at: '2026-09-15T02:25:58.975476+00:00' mode: full seed: 17 steps_override: null accelerator: gpu threads: 1 provenance: revision: null source_sha256: 1416b1e5eae1bca63fe91c913072ace23e081c14c5fcd75fcd5a9b7b1eb9c9d5 script_sha256: 878047d7b42d34092cc055942a7fa6d2c07acc26d54ec1a024d5216d70f3c934 python: 3.12.6 platform: Linux-6.8.0-138-generic-x86_64-with-glibc2.35 host: lab relflow_path: /home/grantham/relflow-proof-scripts.RB4VJH/src/relflow packages: relflow: 0.2.1 torch: 2.12.0 lightning: 2.6.4 numpy: 2.4.6 pyarrow: 24.0.0 polars: 1.44.1 torchmetrics: 1.9.0 tensordict: 0.12.4 pydantic: 2.13.4 cuda: '13.0' gpu: NVIDIA GeForce RTX 3090 code_sha256: 42e2656c4c6174d3130fcb1e5c58a75385b7a77203226f60d0a76aa0d249a76f metrics: intact_nrmse: 0.22284473798592025 broken_nrmse: 1.3745850849698824 corruption_gap: 1.1517403469839622 checks: Finite errors: true Recall calibration nRMSE <= 0.90: true Identity recall nRMSE <= 0.25: true Broken identities nRMSE >= 0.80: true Corruption increases nRMSE by >= 0.50: true outcome: met duration_seconds: 18.046 P026: script: proofs/relational/filtered_mean.py page: docs/proofs/relational/filtered-mean.py historical: status: Provisional pass source_revision: 810b3fe3b0b25c2c7a718571d83f6006f23110c3 sources: - https://github.com/relflow/relflow/blob/810b3fe/proofs/relational/category_conditioned_reduction/test_filtered_mean.py - https://github.com/relflow/relflow/blob/810b3fe/proofs/relational/category_conditioned_reduction/__init__.py summary: |- The checked-in report records passing accuracy and label-corruption gates for all three groups on one deterministic CPU seed. Exact per-group scores are not recorded. insights: |- **The model uses group membership to select which values contribute to its answer.** The same bag supplies separate requests for each group, while the operation stays fixed at mean. Item coordinates keep group and value together; the visible root request can condition how the decoder uses the learned collection summary. Shuffling only group labels preserves all values and category counts but makes the retained original answers inconsistent with the new membership. Increased error supports use of the group/value relationship. It does not establish an exact filtering algorithm or behavior for absent groups and unknown categories. This small task also shows that pass-through reduction is not required for every learned group selection. evidence: |- Training uses 512 bags, validation 64, and evaluation 128, each with two items per group, over 800 CPU steps. Each group's nRMSE uses its own training-mean baseline. | Check | Required gate | Recorded result | | --- | --- | --- | | Each selected group | nRMSE ≤ 0.55 | Gate reported passing | | Corrupted labels, each group | At least 0.25 nRMSE worse | Gate reported passing | runs: - started_at: '2026-09-15T02:15:41.934570+00:00' mode: smoke seed: 16 steps_override: 2 accelerator: gpu threads: 1 provenance: revision: null source_sha256: d0969247caa4a66c61d244e6f5b19e8435e14f29d98b1bc8bea094de43162c77 script_sha256: 24dce1cd9121d4ce217c351000eb2c7d738d134d30bb7fb03a3a29b68bb9672b python: 3.12.6 platform: Linux-6.8.0-138-generic-x86_64-with-glibc2.35 host: lab packages: relflow: 0.2.1 torch: 2.12.0 lightning: 2.6.4 numpy: 2.4.6 cuda: '13.0' gpu: NVIDIA GeForce RTX 3090 metrics: intact/A/rmse: 1.4523209310098621 intact/A/baseline_rmse: 0.8356639150214514 intact/A/nrmse: 1.73792466672751 intact/B/rmse: 1.4579753429050628 intact/B/baseline_rmse: 0.8356760790930771 intact/B/nrmse: 1.7446656418445532 intact/C/rmse: 1.3736421037993132 intact/C/baseline_rmse: 0.8854705200121815 intact/C/nrmse: 1.5513131976210974 corrupted/A/rmse: 1.4521147074815042 corrupted/A/baseline_rmse: 0.8356639150214514 corrupted/A/nrmse: 1.7376778886572224 corrupted/B/rmse: 1.460544778113098 corrupted/B/baseline_rmse: 0.8356760790930771 corrupted/B/nrmse: 1.747740320266393 corrupted/C/rmse: 1.3736453993344222 corrupted/C/baseline_rmse: 0.8854705200121815 corrupted/C/nrmse: 1.5513169194108516 checks: Group A nRMSE <= 0.55: false Group B nRMSE <= 0.55: false Group C nRMSE <= 0.55: false Group A corruption increases nRMSE by >= 0.25: false Group B corruption increases nRMSE by >= 0.25: false Group C corruption increases nRMSE by >= 0.25: false outcome: not_met duration_seconds: 0.838 - started_at: '2026-09-15T02:26:20.438891+00:00' mode: full seed: 16 steps_override: null accelerator: gpu threads: 1 provenance: revision: null source_sha256: fbf0ec5b8f7b7d088d28ccd39477847b27d364a3b7ea3eb1a036e6e04da6d153 script_sha256: 0f63f0d8908f3f35cc2a58a6584ba703f971215e9eced5e9b1ba73695298f6fa python: 3.12.6 platform: Linux-6.8.0-138-generic-x86_64-with-glibc2.35 host: lab relflow_path: /home/grantham/relflow-proof-scripts.RB4VJH/src/relflow packages: relflow: 0.2.1 torch: 2.12.0 lightning: 2.6.4 numpy: 2.4.6 pyarrow: 24.0.0 polars: 1.44.1 torchmetrics: 1.9.0 tensordict: 0.12.4 pydantic: 2.13.4 cuda: '13.0' gpu: NVIDIA GeForce RTX 3090 code_sha256: 35a76e5b8dd67fb06ad22b0303f47f0543e320c432f5ad13f3277e8b444075d8 metrics: intact/A/rmse: 0.022685833035542592 intact/A/baseline_rmse: 0.8356639150214514 intact/A/nrmse: 0.027147077464700926 intact/B/rmse: 0.016297808242230382 intact/B/baseline_rmse: 0.8356760790930771 intact/B/nrmse: 0.01950254249220306 intact/C/rmse: 0.01636764611243637 intact/C/baseline_rmse: 0.8854705200121815 intact/C/nrmse: 0.018484687793119525 corrupted/A/rmse: 0.7936227859752805 corrupted/A/baseline_rmse: 0.8356639150214514 corrupted/A/nrmse: 0.9496913432655618 corrupted/B/rmse: 0.7921123931195244 corrupted/B/baseline_rmse: 0.8356760790930771 corrupted/B/nrmse: 0.9478701292720615 corrupted/C/rmse: 0.8906145939256858 corrupted/C/baseline_rmse: 0.8854705200121815 corrupted/C/nrmse: 1.0058094242521292 checks: Group A nRMSE <= 0.55: true Group B nRMSE <= 0.55: true Group C nRMSE <= 0.55: true Group A corruption increases nRMSE by >= 0.25: true Group B corruption increases nRMSE by >= 0.25: true Group C corruption increases nRMSE by >= 0.25: true outcome: met duration_seconds: 27.037 P027: script: proofs/relational/group_operation_composition.py page: docs/proofs/relational/group-operation-composition.py historical: status: Provisional pass source_revision: 810b3fe3b0b25c2c7a718571d83f6006f23110c3 sources: - https://github.com/relflow/relflow/blob/810b3fe/proofs/relational/category_conditioned_reduction/test_group_operation_composition.py - https://github.com/relflow/relflow/blob/810b3fe/proofs/relational/category_conditioned_reduction/__init__.py summary: |- The checked-in report records all twelve group-operation cells passing their nRMSE gate on one deterministic CPU seed. Corruption controls for this composed case remain unfinished; the simpler filtered-mean proof supplies the current label-corruption evidence. insights: |- **The model can answer requests for different groups and operations, but this combined task still needs checks that it uses both requests.** The same bag supplies different answers according to visible group and operation requests. Measuring every cell prevents an easy subset from hiding a failed operation or group. Item coordinates bind categories to values, retained item slots preserve their evidence, and request fields can condition the scalar decoder. The twelve root outputs are learned summaries, not twelve declared statistics. Label and request corruptions have not been applied to this exact route; neighboring proofs do not substitute for them. The evidence supports bounded conditional prediction, with broader relational composition still requiring validation. evidence: |- The protocol trains on 512 bags, validates on 64, and tests 128 after 1,200 CPU steps. Each cell's error is normalized by its matching training-mean baseline. | Check | Required gate | Recorded result | | --- | --- | --- | | Every group-operation cell | nRMSE ≤ 0.30 | All twelve gates reported passing | | Numerical calibration | Finite nRMSE in every cell | Gate reported passing | runs: - started_at: '2026-09-15T02:15:46.160072+00:00' mode: smoke seed: 16 steps_override: 2 accelerator: gpu threads: 1 provenance: revision: null source_sha256: 88a51fd340db0ef3750bd8f39a8d4b51ab649e93c5d25cf2ed1dde0b7897c44c script_sha256: 9c92d932819a07221efb44a05b1f054a78a964ae168be02ebf45505a08819555 python: 3.12.6 platform: Linux-6.8.0-138-generic-x86_64-with-glibc2.35 host: lab packages: relflow: 0.2.1 torch: 2.12.0 lightning: 2.6.4 numpy: 2.4.6 cuda: '13.0' gpu: NVIDIA GeForce RTX 3090 metrics: intact/max/A/rmse: 1.343814516534358 intact/max/A/baseline_rmse: 0.8798919617581095 intact/max/A/nrmse: 1.5272494521364717 intact/max/B/rmse: 1.1504907006995917 intact/max/B/baseline_rmse: 0.8645386409025352 intact/max/B/nrmse: 1.3307568294443575 intact/max/C/rmse: 1.0778907341768926 intact/max/C/baseline_rmse: 0.8226223328561437 intact/max/C/nrmse: 1.3103105655233762 intact/mean/A/rmse: 0.8382183208600148 intact/mean/A/baseline_rmse: 0.8703165895406241 intact/mean/A/nrmse: 0.9631188591986376 intact/mean/B/rmse: 0.8416726901990014 intact/mean/B/baseline_rmse: 0.8474965036769826 intact/mean/B/nrmse: 0.9931282153345603 intact/mean/C/rmse: 0.8131480088868479 intact/mean/C/baseline_rmse: 0.814604977900874 intact/mean/C/nrmse: 0.9982114410621691 intact/min/A/rmse: 1.433727790190114 intact/min/A/baseline_rmse: 0.8766077148642057 intact/min/A/nrmse: 1.635540921987221 intact/min/B/rmse: 1.579062064564362 intact/min/B/baseline_rmse: 0.8365111183932998 intact/min/B/nrmse: 1.8876761227003074 intact/min/C/rmse: 1.671122153392365 intact/min/C/baseline_rmse: 0.8391276941102654 intact/min/C/nrmse: 1.9914992260674589 intact/sum/A/rmse: 3.516778776872889 intact/sum/A/baseline_rmse: 3.4812663581624963 intact/sum/A/nrmse: 1.0102010059147377 intact/sum/B/rmse: 3.357539921339317 intact/sum/B/baseline_rmse: 3.3899860147079304 intact/sum/B/nrmse: 0.9904288415268259 intact/sum/C/rmse: 3.2731957658294757 intact/sum/C/baseline_rmse: 3.258419911603496 intact/sum/C/nrmse: 1.0045346685285594 checks: max/A finite calibration nRMSE <= 1.60: true max/B finite calibration nRMSE <= 1.60: true max/C finite calibration nRMSE <= 1.60: true mean/A finite calibration nRMSE <= 1.60: true mean/B finite calibration nRMSE <= 1.60: true mean/C finite calibration nRMSE <= 1.60: true min/A finite calibration nRMSE <= 1.60: false min/B finite calibration nRMSE <= 1.60: false min/C finite calibration nRMSE <= 1.60: false sum/A finite calibration nRMSE <= 1.60: true sum/B finite calibration nRMSE <= 1.60: true sum/C finite calibration nRMSE <= 1.60: true max/A nRMSE <= 0.30: false max/B nRMSE <= 0.30: false max/C nRMSE <= 0.30: false mean/A nRMSE <= 0.30: false mean/B nRMSE <= 0.30: false mean/C nRMSE <= 0.30: false min/A nRMSE <= 0.30: false min/B nRMSE <= 0.30: false min/C nRMSE <= 0.30: false sum/A nRMSE <= 0.30: false sum/B nRMSE <= 0.30: false sum/C nRMSE <= 0.30: false outcome: not_met duration_seconds: 1.525 - started_at: '2026-09-15T02:26:50.904617+00:00' mode: full seed: 16 steps_override: null accelerator: gpu threads: 1 provenance: revision: null source_sha256: cb22c1a08f196d542a455a3c2afb632dfc40ce2daaaf964ec83921cd95d3b2b1 script_sha256: b8f1279fd689729389850b05e122db2b30185f813165cf70741f0f9200a12ef7 python: 3.12.6 platform: Linux-6.8.0-138-generic-x86_64-with-glibc2.35 host: lab relflow_path: /home/grantham/relflow-proof-scripts.RB4VJH/src/relflow packages: relflow: 0.2.1 torch: 2.12.0 lightning: 2.6.4 numpy: 2.4.6 pyarrow: 24.0.0 polars: 1.44.1 torchmetrics: 1.9.0 tensordict: 0.12.4 pydantic: 2.13.4 cuda: '13.0' gpu: NVIDIA GeForce RTX 3090 code_sha256: 733ea4403e53b9e7862481f18886205de05f779a0a316e12a379716cbd0b9789 metrics: intact/max/A/rmse: 0.18308504484072816 intact/max/A/baseline_rmse: 0.8798919617581095 intact/max/A/nrmse: 0.20807673305130153 intact/max/B/rmse: 0.16941651406402344 intact/max/B/baseline_rmse: 0.8645386409025352 intact/max/B/nrmse: 0.19596176046817418 intact/max/C/rmse: 0.1579989945141643 intact/max/C/baseline_rmse: 0.8226223328561437 intact/max/C/nrmse: 0.19206747519921077 intact/mean/A/rmse: 0.06480399142037688 intact/mean/A/baseline_rmse: 0.8703165895406241 intact/mean/A/nrmse: 0.07446025067105996 intact/mean/B/rmse: 0.07211585837754324 intact/mean/B/baseline_rmse: 0.8474965036769826 intact/mean/B/nrmse: 0.08509280930913397 intact/mean/C/rmse: 0.049025701102873395 intact/mean/C/baseline_rmse: 0.814604977900874 intact/mean/C/nrmse: 0.06018340475797968 intact/min/A/rmse: 0.14529593248443903 intact/min/A/baseline_rmse: 0.8766077148642057 intact/min/A/nrmse: 0.16574795090292657 intact/min/B/rmse: 0.17183883376188822 intact/min/B/baseline_rmse: 0.8365111183932998 intact/min/B/nrmse: 0.2054232513871923 intact/min/C/rmse: 0.17559495849844933 intact/min/C/baseline_rmse: 0.8391276941102654 intact/min/C/nrmse: 0.20925892415532088 intact/sum/A/rmse: 0.1923188640205741 intact/sum/A/baseline_rmse: 3.4812663581624963 intact/sum/A/nrmse: 0.05524393833572822 intact/sum/B/rmse: 0.25666843512215765 intact/sum/B/baseline_rmse: 3.3899860147079304 intact/sum/B/nrmse: 0.07571371504441776 intact/sum/C/rmse: 0.2897579721848155 intact/sum/C/baseline_rmse: 3.258419911603496 intact/sum/C/nrmse: 0.08892591502800606 checks: max/A finite calibration nRMSE <= 1.60: true max/B finite calibration nRMSE <= 1.60: true max/C finite calibration nRMSE <= 1.60: true mean/A finite calibration nRMSE <= 1.60: true mean/B finite calibration nRMSE <= 1.60: true mean/C finite calibration nRMSE <= 1.60: true min/A finite calibration nRMSE <= 1.60: true min/B finite calibration nRMSE <= 1.60: true min/C finite calibration nRMSE <= 1.60: true sum/A finite calibration nRMSE <= 1.60: true sum/B finite calibration nRMSE <= 1.60: true sum/C finite calibration nRMSE <= 1.60: true max/A nRMSE <= 0.30: true max/B nRMSE <= 0.30: true max/C nRMSE <= 0.30: true mean/A nRMSE <= 0.30: true mean/B nRMSE <= 0.30: true mean/C nRMSE <= 0.30: true min/A nRMSE <= 0.30: true min/B nRMSE <= 0.30: true min/C nRMSE <= 0.30: true sum/A nRMSE <= 0.30: true sum/B nRMSE <= 0.30: true sum/C nRMSE <= 0.30: true outcome: met duration_seconds: 36.59 P028: script: proofs/relational/direct_sibling_hash_overlap.py page: docs/proofs/relational/direct-sibling-hash-overlap.py historical: status: Expected limitation source_revision: 810b3fe3b0b25c2c7a718571d83f6006f23110c3 sources: - https://github.com/relflow/relflow/blob/810b3fe/proofs/relational/collection_overlap/test_direct_sibling_hash_overlap.py - https://github.com/relflow/relflow/blob/810b3fe/proofs/relational/collection_overlap/support.py summary: |- The recorded intact **AUC is 0.4972**. This proof deliberately passes when the model remains near chance and barely responds to removing an overlap. A green test records the known limitation; it does not mean overlap detection works. insights: |- **Retaining every identity token does not make the current model reliably detect overlap between two collections.** The scalar target must discover a match anywhere across both sides, which is a different demand from a repeated target carrying its own visible query identity. The passing flat equality control shows that unseen Hash comparison itself can work. This proof deliberately passes when predictions remain near chance and barely respond to removing the shared member. Low drift under renaming and permutation is not useful invariance by itself: an uninformative predictor can also stay stable. The result describes the tested architecture and budget, not an impossibility theorem. Compute overlap directly when an application requires it today. evidence: |- One deterministic CPU run uses 3,072 training, 768 validation, and 1,024 test records over 700 steps. AUC of 0.5 corresponds to chance ranking. | Check | Limitation gate | Recorded result | | --- | --- | --- | | Intact AUC | 0.35–0.65 | 0.4972 | | Consistent renaming and permutation AUC | 0.35–0.65 | 0.5153 | | Removed-overlap AUC against retained labels | 0.35–0.65 | 0.4940 | | Mean invariant probability drift | ≤ 0.15 | 0.0006 | | Mean positive probability drop after removal | Absolute value ≤ 0.10 | Approximately 0.0000 | runs: - started_at: '2026-09-15T02:15:51.082706+00:00' mode: smoke seed: 3711 steps_override: 2 accelerator: gpu threads: 1 provenance: revision: null source_sha256: 16d423174935a6a34268f877e935d2d75756405850dad1daef33919698a2c475 script_sha256: dbadefe57cc473749cefe590eda0e37dceb6737669eb2f205869393771b37e70 python: 3.12.6 platform: Linux-6.8.0-138-generic-x86_64-with-glibc2.35 host: lab packages: relflow: 0.2.1 torch: 2.12.0 lightning: 2.6.4 numpy: 2.4.6 cuda: '13.0' gpu: NVIDIA GeForce RTX 3090 code_sha256: 105c5ef8d9e7cb724caead45e22a50bd200a140fdea28a581c58a12c7b532548 metrics: intact_auc: 0.5161590576171875 invariant_auc: 0.5326423645019531 broken_auc: 0.499969482421875 invariant_drift: 0.012097313185222447 positive_break_drop: 0.0007783769397065043 checks: Labels are exactly balanced: true Labels match actual overlap: true Renaming and permutation preserve overlap: true Intervention removes all overlaps: true Finite metrics: true Intact AUC remains in the calibrated chance band: true Renamed and permuted AUC remains in the chance band: true Broken AUC remains in the chance band: true Invariant probability drift <= 0.15: true Absolute positive break response <= 0.10: true train generator has balanced truthful labels: true train sides have two to five unique members: true train positive rows have exactly one overlap: true validate generator has balanced truthful labels: true validate sides have two to five unique members: true validate positive rows have exactly one overlap: true test generator has balanced truthful labels: true test sides have two to five unique members: true test positive rows have exactly one overlap: true invariant generator has balanced truthful labels: true invariant sides have two to five unique members: true invariant positive rows have exactly one overlap: true outcome: met duration_seconds: 1.02 - started_at: '2026-09-15T02:27:30.953191+00:00' mode: full seed: 3711 steps_override: null accelerator: gpu threads: 1 provenance: revision: null source_sha256: da4241972f82b268dd5602e23ee7406ae38d32d5efe15b016539e9fd38034103 script_sha256: dbadefe57cc473749cefe590eda0e37dceb6737669eb2f205869393771b37e70 python: 3.12.6 platform: Linux-6.8.0-138-generic-x86_64-with-glibc2.35 host: lab relflow_path: /home/grantham/relflow-proof-scripts.RB4VJH/src/relflow packages: relflow: 0.2.1 torch: 2.12.0 lightning: 2.6.4 numpy: 2.4.6 pyarrow: 24.0.0 polars: 1.44.1 torchmetrics: 1.9.0 tensordict: 0.12.4 pydantic: 2.13.4 cuda: '13.0' gpu: NVIDIA GeForce RTX 3090 code_sha256: 105c5ef8d9e7cb724caead45e22a50bd200a140fdea28a581c58a12c7b532548 metrics: intact_auc: 0.47298240661621094 invariant_auc: 0.5077915191650391 broken_auc: 0.47818946838378906 invariant_drift: 0.001136141800088808 positive_break_drop: -2.6322610210627317e-05 checks: Labels are exactly balanced: true Labels match actual overlap: true Renaming and permutation preserve overlap: true Intervention removes all overlaps: true Finite metrics: true Intact AUC remains in the calibrated chance band: true Renamed and permuted AUC remains in the chance band: true Broken AUC remains in the chance band: true Invariant probability drift <= 0.15: true Absolute positive break response <= 0.10: true train generator has balanced truthful labels: true train sides have two to five unique members: true train positive rows have exactly one overlap: true validate generator has balanced truthful labels: true validate sides have two to five unique members: true validate positive rows have exactly one overlap: true test generator has balanced truthful labels: true test sides have two to five unique members: true test positive rows have exactly one overlap: true invariant generator has balanced truthful labels: true invariant sides have two to five unique members: true invariant positive rows have exactly one overlap: true outcome: met duration_seconds: 24.44 P029: script: proofs/relational/flat_unseen_hash_equality_control.py page: docs/proofs/relational/flat-unseen-hash-equality-control.py historical: status: Provisional pass source_revision: 810b3fe3b0b25c2c7a718571d83f6006f23110c3 sources: - https://github.com/relflow/relflow/blob/810b3fe/proofs/relational/collection_overlap/test_flat_unseen_hash_equality_control.py - https://github.com/relflow/relflow/blob/810b3fe/proofs/relational/collection_overlap/support.py summary: |- The checked-in report records **0.9890 AUC** for unseen equality on one deterministic CPU seed. This supports the identity primitive; it does not establish [sibling collection overlap](direct-sibling-hash-overlap.qmd). insights: |- **The model can compare two unseen scalar identities, so failed collection overlap cannot be explained solely by an unusable Hash representation.** The fields share the same root context and preserve equality through Hash encoding. Fresh identities and balanced labels rule out a persistent lookup table or a majority-label solution. Making formerly equal pairs unequal while retaining their positive labels lowers their predicted probabilities and removes ranking skill against those labels. That supports sensitivity to equality, rather than merely a favorable intact score. It does not establish calibrated probabilities or an all-pairs comparison across repeated branches. Keep this primitive control when testing more complex identity-routing schemas, so different failure locations remain distinguishable. evidence: |- The protocol uses 3,072 training, 768 validation, and 1,024 test records with 600 CPU training steps. AUC evaluates positive/negative probability ranking. | Check | Required gate | Recorded result | | --- | --- | --- | | Intact equality | AUC ≥ 0.95 | 0.9890 | | All-unequal intervention | AUC between 0.35 and 0.65 | Gate reported passing | | Formerly positive mean probability drop | ≥ 0.25 | Gate reported passing | runs: - started_at: '2026-09-15T02:15:55.481514+00:00' mode: smoke seed: 3701 steps_override: 2 accelerator: gpu threads: 1 provenance: revision: null source_sha256: 382a44adb3b7cbbdd0fad996496997f360efc46c4f347fa34439d41c063d4bcc script_sha256: b74c5b04fbbd8a062b08955346d0302fc697c48fae2ebdd82fd87b81b232a980 python: 3.12.6 platform: Linux-6.8.0-138-generic-x86_64-with-glibc2.35 host: lab packages: relflow: 0.2.1 torch: 2.12.0 lightning: 2.6.4 numpy: 2.4.6 cuda: '13.0' gpu: NVIDIA GeForce RTX 3090 code_sha256: d643cee2bef8d7a7b905b658008a5437b069c80d0a935394ec8788a6524e7aca metrics: intact_auc: 0.48752593994140625 broken_auc: 0.5044612884521484 positive_probability_drop: -0.0024551497772336006 checks: Finite metrics: true Unseen equality AUC >= 0.95: false All-unequal control AUC is in [0.35, 0.65]: true Breaking equality lowers positive probability by >= 0.25: false outcome: not_met duration_seconds: 0.746 - started_at: '2026-09-15T02:27:58.822559+00:00' mode: full seed: 3701 steps_override: null accelerator: gpu threads: 1 provenance: revision: null source_sha256: 91bd2a492989c34950f499cf1412eca0775dd53798c15806bfb6596c10c9cc4c script_sha256: b74c5b04fbbd8a062b08955346d0302fc697c48fae2ebdd82fd87b81b232a980 python: 3.12.6 platform: Linux-6.8.0-138-generic-x86_64-with-glibc2.35 host: lab relflow_path: /home/grantham/relflow-proof-scripts.RB4VJH/src/relflow packages: relflow: 0.2.1 torch: 2.12.0 lightning: 2.6.4 numpy: 2.4.6 pyarrow: 24.0.0 polars: 1.44.1 torchmetrics: 1.9.0 tensordict: 0.12.4 pydantic: 2.13.4 cuda: '13.0' gpu: NVIDIA GeForce RTX 3090 code_sha256: d643cee2bef8d7a7b905b658008a5437b069c80d0a935394ec8788a6524e7aca metrics: intact_auc: 0.9998245239257812 broken_auc: 0.5046043395996094 positive_probability_drop: 0.9927793605953639 checks: Finite metrics: true Unseen equality AUC >= 0.95: true All-unequal control AUC is in [0.35, 0.65]: true Breaking equality lowers positive probability by >= 0.25: true outcome: met duration_seconds: 12.087 P030: script: proofs/relational/repeated_subtotal_reconstruction.py page: docs/proofs/relational/repeated-subtotal-reconstruction.py historical: status: Provisional pass source_revision: 810b3fe3b0b25c2c7a718571d83f6006f23110c3 sources: - https://github.com/relflow/relflow/blob/810b3fe/proofs/relational/item_alignment/test_repeated_subtotal_reconstruction.py - https://github.com/relflow/relflow/blob/810b3fe/proofs/relational/item_alignment/__init__.py summary: |- The checked-in report records passing accuracy, target-permutation, and output geometry gates for one seed. It reports threshold success rather than an exact measured nRMSE. insights: |- **The model learns each item's subtotal and keeps its prediction attached to the correct coordinate.** Repeated decoder queries can use aligned quantity and price inputs as well as ancestor context; they are not limited to one identical root representation for every item. Permuting only target subtotals leaves requests and predictions unchanged but raises error against the altered labels. That isolates coordinate alignment. Separate assertions verify that real items are inferred and padded coordinates are not, including empty orders. The arithmetic score is still aggregated over lengths, so a weak individual length could be hidden. This establishes local learned arithmetic and output geometry, not general cross-item reasoning or exact billing calculations. evidence: |- The protocol fits 4,096 orders, validates on 1,024, and tests 2,048 over 25 CPU epochs. nRMSE divides error by the training-mean subtotal baseline's test error. | Check | Required gate | Recorded result | | --- | --- | --- | | Aligned subtotals | nRMSE ≤ 0.25 | Gate reported passing | | Permuted targets | At least 0.50 nRMSE worse | Gate reported passing | | Output shape | Exactly five coordinates per order | Gate reported passing | | Inferred flags | Match every real and padded position | Gate reported passing | runs: - started_at: '2026-09-15T02:15:59.644893+00:00' mode: smoke seed: 31 steps_override: 2 accelerator: gpu threads: 1 provenance: revision: null source_sha256: 09034fe14c4e7d2bcb9140d8874c5d9c560f5be662cb507b93f68379a6ad677f script_sha256: a36c5276358abc7ad4fd394d00ba84875037a2e0a70c181bde322eb31912c870 python: 3.12.6 platform: Linux-6.8.0-138-generic-x86_64-with-glibc2.35 host: lab packages: relflow: 0.2.1 torch: 2.12.0 lightning: 2.6.4 numpy: 2.4.6 cuda: '13.0' gpu: NVIDIA GeForce RTX 3090 code_sha256: b5ee5a1268175a2973329686c3cffc200b555c22b1c4e6f8c5fc3c2cc6ebc214 metrics: aligned_nrmse: 0.7709520467502883 permuted_target_nrmse: 0.9386946831751944 baseline_rmse: 1.0618151651706766 checks: Target permutation preserves visible inputs: true Output retains the fixed branch length: true Inferred flags match real and padded coordinates: true Aligned subtotal nRMSE <= 0.25: false Permuting targets increases nRMSE by >= 0.50: false outcome: not_met duration_seconds: 0.96 - started_at: '2026-09-15T02:28:14.376443+00:00' mode: full seed: 31 steps_override: null accelerator: gpu threads: 1 provenance: revision: null source_sha256: fda8373e15faa5243cd2403d7f817b0918a80af65494718cbace2bcbe2769df9 script_sha256: a36c5276358abc7ad4fd394d00ba84875037a2e0a70c181bde322eb31912c870 python: 3.12.6 platform: Linux-6.8.0-138-generic-x86_64-with-glibc2.35 host: lab relflow_path: /home/grantham/relflow-proof-scripts.RB4VJH/src/relflow packages: relflow: 0.2.1 torch: 2.12.0 lightning: 2.6.4 numpy: 2.4.6 pyarrow: 24.0.0 polars: 1.44.1 torchmetrics: 1.9.0 tensordict: 0.12.4 pydantic: 2.13.4 cuda: '13.0' gpu: NVIDIA GeForce RTX 3090 code_sha256: b5ee5a1268175a2973329686c3cffc200b555c22b1c4e6f8c5fc3c2cc6ebc214 metrics: aligned_nrmse: 0.014522491722698478 permuted_target_nrmse: 1.15685115020635 baseline_rmse: 1.0618151651706766 checks: Target permutation preserves visible inputs: true Output retains the fixed branch length: true Inferred flags match real and padded coordinates: true Aligned subtotal nRMSE <= 0.25: true Permuting targets increases nRMSE by >= 0.50: true outcome: met duration_seconds: 24.979 P031: script: proofs/relational/fixed_length.py page: docs/proofs/relational/fixed-length.py historical: status: Provisional pass source_revision: 810b3fe3b0b25c2c7a718571d83f6006f23110c3 sources: - https://github.com/relflow/relflow/blob/810b3fe/proofs/relational/operation_conditioned_reduction/test_fixed_length.py - https://github.com/relflow/relflow/blob/810b3fe/proofs/relational/operation_conditioned_reduction/__init__.py summary: |- The checked-in report records passing accuracy, permutation, and hidden-request gates for one deterministic CPU seed. Only the fixed-length stage is implemented; variable-length behavior is not established. insights: |- **The model changes its answer when the operation changes and approximately preserves it when item order changes.** Hiding the operation makes the four requests for a bag identical while their original labels remain different. Their predictions collapse and lose skill, confirming that the request supplies necessary information. The operation is an ordinary Category that can condition the scalar decoder; it does not select an exact Python arithmetic function. The permutation control checks prediction drift as well as accuracy. However, all bags have eight values, where sum is always eight times mean. Success therefore does not establish cardinality-aware reduction at unseen lengths or exact numerical invariance. evidence: |- The protocol uses 768 training bags, 96 validation bags, and 192 test bags, each expanded to four requests, with 800 CPU steps. nRMSE uses a separate training-mean baseline for each operation. | Check | Required gate | Recorded result | | --- | --- | --- | | Each intact operation | nRMSE ≤ 0.20 | Gate reported passing | | Each permuted operation | nRMSE ≤ 0.25 | Gate reported passing | | Permutation prediction drift | ≤ 0.05 target standard deviations | Gate reported passing | | Hidden operation | At least three operations have nRMSE ≥ 0.75 | Gate reported passing | | Hidden-request predictions within a bag | Range ≤ 0.000001 | Gate reported passing | runs: - started_at: '2026-09-15T02:16:03.996588+00:00' mode: smoke seed: 15 steps_override: 2 accelerator: gpu threads: 1 provenance: revision: null source_sha256: 62ced399961d080505c59329596ad3767dc9bb626979be359aa15da5b22d2163 script_sha256: 7f81b9118213f10591d191b93e52f855717cd5a3b08be9a5712abf5fa9f24c93 python: 3.12.6 platform: Linux-6.8.0-138-generic-x86_64-with-glibc2.35 host: lab packages: relflow: 0.2.1 torch: 2.12.0 lightning: 2.6.4 numpy: 2.4.6 cuda: '13.0' gpu: NVIDIA GeForce RTX 3090 code_sha256: 31fcdcf26561155d7b818d649ea6a8403986927026e2d4b3a0798fceb2f8a7c3 metrics: intact/max/rmse: 0.24276711326097464 intact/max/baseline_rmse: 0.23385342033136094 intact/max/nrmse: 1.0381165813909556 intact/mean/rmse: 0.874801602685532 intact/mean/baseline_rmse: 0.07489171086510313 intact/mean/nrmse: 11.680886877604479 intact/min/rmse: 1.6565280672176041 intact/min/baseline_rmse: 0.16915466453487707 intact/min/nrmse: 9.792978938964191 intact/sum/rmse: 0.6963204996029917 intact/sum/baseline_rmse: 0.599133686920825 intact/sum/nrmse: 1.1622122321007295 permuted/max/rmse: 0.24370657285819872 permuted/max/baseline_rmse: 0.23385342033136094 permuted/max/nrmse: 1.042133882467386 permuted/mean/rmse: 0.8748468615042658 permuted/mean/baseline_rmse: 0.07489171086510313 permuted/mean/nrmse: 11.681491201076478 permuted/min/rmse: 1.6564830683180556 permuted/min/baseline_rmse: 0.16915466453487707 permuted/min/nrmse: 9.792712916743213 permuted/sum/rmse: 0.6965404203539123 permuted/sum/baseline_rmse: 0.599133686920825 permuted/sum/nrmse: 1.1625792966736646 hidden/max/rmse: 0.3520049073429206 hidden/max/baseline_rmse: 0.23385342033136094 hidden/max/nrmse: 1.505237369819709 hidden/mean/rmse: 1.3600422330501925 hidden/mean/baseline_rmse: 0.07489171086510313 hidden/mean/nrmse: 18.160117018823826 hidden/min/rmse: 2.5565300579880046 hidden/min/baseline_rmse: 0.16915466453487707 hidden/min/nrmse: 15.113565239348677 hidden/sum/rmse: 0.6036948735460839 hidden/sum/baseline_rmse: 0.599133686920825 hidden/sum/nrmse: 1.007612969734185 permutation_delta: 0.008511913246519127 hidden_request_spread: 0.0 checks: max intact nRMSE <= 0.20: false mean intact nRMSE <= 0.20: false min intact nRMSE <= 0.20: false sum intact nRMSE <= 0.20: false max permuted nRMSE <= 0.25: false mean permuted nRMSE <= 0.25: false min permuted nRMSE <= 0.25: false sum permuted nRMSE <= 0.25: false Permutation changes predictions by <= 0.05 target SD: true Hidden operations yield identical predictions within each bag: true At least three hidden operations have nRMSE >= 0.75: true outcome: not_met duration_seconds: 0.872 - started_at: '2026-09-15T02:28:42.814023+00:00' mode: full seed: 15 steps_override: null accelerator: gpu threads: 1 provenance: revision: null source_sha256: 2d5872818918750d5dd4a892f068fcd0cf560af376de7e361367382beb0a5b72 script_sha256: 7f81b9118213f10591d191b93e52f855717cd5a3b08be9a5712abf5fa9f24c93 python: 3.12.6 platform: Linux-6.8.0-138-generic-x86_64-with-glibc2.35 host: lab relflow_path: /home/grantham/relflow-proof-scripts.RB4VJH/src/relflow packages: relflow: 0.2.1 torch: 2.12.0 lightning: 2.6.4 numpy: 2.4.6 pyarrow: 24.0.0 polars: 1.44.1 torchmetrics: 1.9.0 tensordict: 0.12.4 pydantic: 2.13.4 cuda: '13.0' gpu: NVIDIA GeForce RTX 3090 code_sha256: 31fcdcf26561155d7b818d649ea6a8403986927026e2d4b3a0798fceb2f8a7c3 metrics: intact/max/rmse: 0.016075673436341642 intact/max/baseline_rmse: 0.23385342033136094 intact/max/nrmse: 0.06874252005193277 intact/mean/rmse: 0.008308030323814258 intact/mean/baseline_rmse: 0.07489171086510313 intact/mean/nrmse: 0.11093391014632975 intact/min/rmse: 0.015463228388389033 intact/min/baseline_rmse: 0.16915466453487707 intact/min/nrmse: 0.09141473237470643 intact/sum/rmse: 0.07823626348240274 intact/sum/baseline_rmse: 0.599133686920825 intact/sum/nrmse: 0.13058231441548301 permuted/max/rmse: 0.016185715502419738 permuted/max/baseline_rmse: 0.23385342033136094 permuted/max/nrmse: 0.06921308005452828 permuted/mean/rmse: 0.008212482576849044 permuted/mean/baseline_rmse: 0.07489171086510313 permuted/mean/nrmse: 0.10965809809902165 permuted/min/rmse: 0.015627105714010044 permuted/min/baseline_rmse: 0.16915466453487707 permuted/min/nrmse: 0.09238353406912983 permuted/sum/rmse: 0.07855264479525767 permuted/sum/baseline_rmse: 0.599133686920825 permuted/sum/nrmse: 0.1311103790524106 hidden/max/rmse: 0.43109935007945926 hidden/max/baseline_rmse: 0.23385342033136094 hidden/max/nrmse: 1.8434596743062759 hidden/mean/rmse: 1.0185653080899941 hidden/mean/baseline_rmse: 0.07489171086510313 hidden/mean/nrmse: 13.600507937716367 hidden/min/rmse: 2.18709472988387 hidden/min/baseline_rmse: 0.16915466453487707 hidden/min/nrmse: 12.929556130761767 hidden/sum/rmse: 0.4999611200004406 hidden/sum/baseline_rmse: 0.599133686920825 hidden/sum/nrmse: 0.8344733920236236 permutation_delta: 0.0035254994047517196 hidden_request_spread: 0.0 checks: max intact nRMSE <= 0.20: true mean intact nRMSE <= 0.20: true min intact nRMSE <= 0.20: true sum intact nRMSE <= 0.20: true max permuted nRMSE <= 0.25: true mean permuted nRMSE <= 0.25: true min permuted nRMSE <= 0.25: true sum permuted nRMSE <= 0.25: true Permutation changes predictions by <= 0.05 target SD: true Hidden operations yield identical predictions within each bag: true At least three hidden operations have nRMSE >= 0.75: true outcome: met duration_seconds: 22.958 P032: script: proofs/relational/grouped_peer_deviation.py page: docs/proofs/relational/grouped-peer-deviation.py historical: status: Provisional pass source_revision: 810b3fe3b0b25c2c7a718571d83f6006f23110c3 sources: - https://github.com/relflow/relflow/blob/810b3fe/proofs/relational/peer_relative_inference/test_grouped_peer_deviation.py - https://github.com/relflow/relflow/blob/810b3fe/proofs/relational/peer_relative_inference/support.py summary: |- The checked-in report records **0.054 nRMSE** on one deterministic CPU seed. Corrupting only group labels raises error to **1.424 nRMSE**, supporting dependence on the group/value relationship in this bounded task. insights: |- **The model uses an item’s group to compare its value with the right peers.** Rotating labels preserves value and label marginals while retaining original targets, and accuracy deteriorates. Independent group translations and whole-item permutations preserve the true relationship and pass their respective controls. The repeated decoder can combine retained collection evidence with its own aligned group and value. That provides a route for membership-dependent peer context without a supplied mean. However, every group has exactly two members, where deviation is half the difference between their values. This is narrower than arbitrary group averaging. Approximate invariance, broader cardinalities, and missing membership still need separate evidence. evidence: |- The protocol trains on 2,048 records, validates on 512, and tests 768 over 900 CPU steps. Every record has three two-member groups. Errors and drift are normalized by the training-target mean baseline's test RMSE. | Check | Required gate | Recorded result | | --- | --- | --- | | Raw grouped deviation | nRMSE < 0.40 | 0.054 | | Independent group translations | nRMSE < 0.50 | 0.358 | | Corrupted labels | At least 0.25 nRMSE worse | 1.424 nRMSE | | Complete-item permutation drift | < 0.20 | 0.056 | | Corruption changes the true answer | Oracle nRMSE > 0.75 | 1.801 | runs: - started_at: '2026-09-15T02:16:08.255738+00:00' mode: smoke seed: 3720 steps_override: 2 accelerator: gpu threads: 1 provenance: revision: null source_sha256: 214aba157d7c9f15e48d1db4be220c0504db173508c31bbe643d2f90b2cebd2f script_sha256: c52208c4ec14675db96c2b233454bb5c949bd9ff072c7cbbd137ea5126d8d645 python: 3.12.6 platform: Linux-6.8.0-138-generic-x86_64-with-glibc2.35 host: lab packages: relflow: 0.2.1 torch: 2.12.0 lightning: 2.6.4 numpy: 2.4.6 cuda: '13.0' gpu: NVIDIA GeForce RTX 3090 code_sha256: b5a56c905a4c6c8288cc2cf6c596fc212632505595b40fede51184b46345a0c4 metrics: intact_rmse: 0.8757277449836613 baseline_rmse: 0.7508689291187257 intact_nrmse: 1.1662857670931717 corrupted_rmse: 0.8761525686649027 corrupted_nrmse: 1.166851543175743 oracle_corruption_nrmse: 1.8011994379885266 translated_rmse: 0.8786807056759145 translated_nrmse: 1.1702184916709737 permutation_drift: 0.02689185364140891 checks: Baseline RMSE > 0.000001: true Label corruption changes the oracle by > 0.75 nRMSE: true Training deviations are exactly centered: true Translation preserves target shape: true Translation preserves all targets: true Permutation preserves the target multiset: true Normalized permutation drift < 0.20: true Grouped deviation nRMSE < 0.40: false Group-translated nRMSE < 0.50: false Label corruption increases nRMSE by >= 0.25: false Finite calibration and model metrics: true outcome: not_met duration_seconds: 1.312 - started_at: '2026-09-15T02:29:09.246487+00:00' mode: full seed: 3720 steps_override: null accelerator: gpu threads: 1 provenance: revision: null source_sha256: a10f24b8624a527c95bf8f3ce7018e8cbaf81185e82f903f2f6f38558f3bfe9a script_sha256: c52208c4ec14675db96c2b233454bb5c949bd9ff072c7cbbd137ea5126d8d645 python: 3.12.6 platform: Linux-6.8.0-138-generic-x86_64-with-glibc2.35 host: lab relflow_path: /home/grantham/relflow-proof-scripts.RB4VJH/src/relflow packages: relflow: 0.2.1 torch: 2.12.0 lightning: 2.6.4 numpy: 2.4.6 pyarrow: 24.0.0 polars: 1.44.1 torchmetrics: 1.9.0 tensordict: 0.12.4 pydantic: 2.13.4 cuda: '13.0' gpu: NVIDIA GeForce RTX 3090 code_sha256: b5a56c905a4c6c8288cc2cf6c596fc212632505595b40fede51184b46345a0c4 metrics: intact_rmse: 0.03094794110295926 baseline_rmse: 0.7508689291187257 intact_nrmse: 0.04121616956408359 corrupted_rmse: 1.1910543200116956 corrupted_nrmse: 1.5862346593693835 oracle_corruption_nrmse: 1.8011994379885266 translated_rmse: 0.05739833194248641 translated_nrmse: 0.07644254505224135 permutation_drift: 0.021679712617358064 checks: Baseline RMSE > 0.000001: true Label corruption changes the oracle by > 0.75 nRMSE: true Training deviations are exactly centered: true Translation preserves target shape: true Translation preserves all targets: true Permutation preserves the target multiset: true Normalized permutation drift < 0.20: true Grouped deviation nRMSE < 0.40: true Group-translated nRMSE < 0.50: true Label corruption increases nRMSE by >= 0.25: true Finite calibration and model metrics: true outcome: met duration_seconds: 26.181 P033: script: proofs/relational/single_query_preserves_group_membership.py page: docs/proofs/relational/single-query-preserves-group-membership.py historical: status: Provisional pass source_revision: 810b3fe3b0b25c2c7a718571d83f6006f23110c3 sources: - https://github.com/relflow/relflow/blob/810b3fe/proofs/relational/peer_relative_inference/test_single_query_preserves_group_membership.py - https://github.com/relflow/relflow/blob/810b3fe/proofs/relational/peer_relative_inference/support.py summary: |- The checked-in report records **0.074 nRMSE**, worsening to **1.561** after group-label corruption, on one deterministic CPU seed. This supports the specific grouped-deviation route; it does not show that one vector preserves arbitrary collection information. insights: |- **A single collection summary can support per-item predictions when the decoder also sees each item’s group and value.** The result is not evidence that one vector losslessly stores an arbitrary collection or that all useful information passes through that vector alone. Corrupting group labels while retaining original targets removes accuracy, supporting use of the membership relationship in this compressed route. Translation and permutation controls from the retained-token case have not been repeated here. This model also trains for more steps than that case, so the recorded accuracy comparison is not an efficiency comparison. Validate summary capacity for the actual target and collection sizes before generalizing from these three two-member groups. evidence: |- The case uses the same 2,048 training, 512 validation, and 768 test records as the retained-token case, but trains for 1,200 CPU steps. nRMSE uses the constant training-target mean baseline; these results are not an efficiency comparison. | Check | Required gate | Recorded result | | --- | --- | --- | | Grouped deviation | nRMSE ≤ 0.20 | 0.074 | | Corrupted group labels | nRMSE ≥ 0.90 | 1.561 | | Corruption gap | At least 0.75 nRMSE | Approximately 1.487 | | Corruption changes the true answer | Oracle nRMSE > 0.75 | Gate reported passing | runs: - started_at: '2026-09-15T02:16:12.968141+00:00' mode: smoke seed: 3720 steps_override: 2 accelerator: gpu threads: 1 provenance: revision: null source_sha256: 687b97acb5e8b32582ea68f03372e000e2e509b8b60aa7ae003132cee5754158 script_sha256: fe5e6cda6001a993912ca7eb2ef0339ee3fa3c49b0f7f9618bf77aba6853fcdd python: 3.12.6 platform: Linux-6.8.0-138-generic-x86_64-with-glibc2.35 host: lab packages: relflow: 0.2.1 torch: 2.12.0 lightning: 2.6.4 numpy: 2.4.6 cuda: '13.0' gpu: NVIDIA GeForce RTX 3090 code_sha256: ed0e37033289539d08a0f6e5c37cbbdaad64a7afa96e7e77512abc64e174a3cd metrics: intact_rmse: 0.8341561749725644 baseline_rmse: 0.7508689291187257 intact_nrmse: 1.110921150981158 corrupted_rmse: 0.8346241720007317 corrupted_nrmse: 1.1115444249108926 oracle_corruption_nrmse: 1.8011994379885266 checks: Baseline RMSE > 0.000001: true Label corruption changes the oracle by > 0.75 nRMSE: true Single-summary grouped nRMSE <= 0.20: false Corrupted labels nRMSE >= 0.90: true Label corruption increases nRMSE by >= 0.75: false Finite calibration and model metrics: true outcome: not_met duration_seconds: 0.905 - started_at: '2026-09-15T02:29:38.906782+00:00' mode: full seed: 3720 steps_override: null accelerator: gpu threads: 1 provenance: revision: null source_sha256: bfba4071bb16e91ed6bfc14b1e8f66ea54f3b7777ec49ec9b53ad05ee59f07c3 script_sha256: fe5e6cda6001a993912ca7eb2ef0339ee3fa3c49b0f7f9618bf77aba6853fcdd python: 3.12.6 platform: Linux-6.8.0-138-generic-x86_64-with-glibc2.35 host: lab relflow_path: /home/grantham/relflow-proof-scripts.RB4VJH/src/relflow packages: relflow: 0.2.1 torch: 2.12.0 lightning: 2.6.4 numpy: 2.4.6 pyarrow: 24.0.0 polars: 1.44.1 torchmetrics: 1.9.0 tensordict: 0.12.4 pydantic: 2.13.4 cuda: '13.0' gpu: NVIDIA GeForce RTX 3090 code_sha256: ed0e37033289539d08a0f6e5c37cbbdaad64a7afa96e7e77512abc64e174a3cd metrics: intact_rmse: 0.04267593323879911 baseline_rmse: 0.7508689291187257 intact_nrmse: 0.0568353964105116 corrupted_rmse: 1.1652676743201078 corrupted_nrmse: 1.5518922532694894 oracle_corruption_nrmse: 1.8011994379885266 checks: Baseline RMSE > 0.000001: true Label corruption changes the oracle by > 0.75 nRMSE: true Single-summary grouped nRMSE <= 0.20: true Corrupted labels nRMSE >= 0.90: true Label corruption increases nRMSE by >= 0.75: true Finite calibration and model metrics: true outcome: met duration_seconds: 37.115 P034: script: proofs/relational/supplied_peer_mean_control.py page: docs/proofs/relational/supplied-peer-mean-control.py historical: status: Provisional pass source_revision: 810b3fe3b0b25c2c7a718571d83f6006f23110c3 sources: - https://github.com/relflow/relflow/blob/810b3fe/proofs/relational/peer_relative_inference/test_supplied_peer_mean_control.py - https://github.com/relflow/relflow/blob/810b3fe/proofs/relational/peer_relative_inference/support.py summary: |- The checked-in report records **0.030 nRMSE** on one deterministic CPU seed. Because the mean is supplied, success establishes local arithmetic and coordinate decoding; it does not establish learned averaging or peer selection. insights: |- **Supplying the peer mean makes this a test of local subtraction, not learned aggregation.** Each hidden deviation has both operands beside it at the same coordinate. The repeated decoder can use those visible siblings together with ancestor context to return an item-specific Number. This control helps localize a failure before asking a model to infer a mean or select peers. Its low error cannot establish either of those harder capabilities, because the input already supplies the statistic. The recorded failed compression contrast was exploratory and is not a general impossibility result. Compare the raw-value proof to assess aggregation; use this case to check whether local arithmetic and repeated decoding work. evidence: |- Training uses 1,024 records, validation 256, and evaluation 512, with six items per record and 600 CPU steps. nRMSE is relative to the constant training-target mean baseline. | Check | Required gate | Recorded result | | --- | --- | --- | | Local subtraction | nRMSE < 0.25 | 0.030 | | Numerical calibration | Finite errors; nondegenerate baseline | Gate reported passing | The historical source also records an exploratory compressed contrast at 1.000 nRMSE. It was not retained as a behavioral gate and does not establish that compression always fails. runs: - started_at: '2026-09-15T02:16:17.273254+00:00' mode: smoke seed: 3700 steps_override: 2 accelerator: gpu threads: 1 provenance: revision: null source_sha256: 71bddf8132d18f65ccb25c6a7f9a19944e81eb12a126b584c42d684f16e78113 script_sha256: 4cace2eb41ab88a71f1bc9b585123a421d4970358090b99fe7de405bcc5d0469 python: 3.12.6 platform: Linux-6.8.0-138-generic-x86_64-with-glibc2.35 host: lab packages: relflow: 0.2.1 torch: 2.12.0 lightning: 2.6.4 numpy: 2.4.6 cuda: '13.0' gpu: NVIDIA GeForce RTX 3090 code_sha256: b9daad4e5b1827cb5da87393719da162e3431161105edf2eeda124d237114638 metrics: intact_rmse: 0.9265824077626078 baseline_rmse: 0.7110739767923704 intact_nrmse: 1.3030745576464327 checks: Baseline RMSE > 0.000001: true Supplied-mean subtraction nRMSE < 0.25: false Finite calibration and model metrics: true outcome: not_met duration_seconds: 0.788 - started_at: '2026-09-15T02:30:19.496159+00:00' mode: full seed: 3700 steps_override: null accelerator: gpu threads: 1 provenance: revision: null source_sha256: f3022f2bb476ae3edea531b988bc93f235d22fc7907aad02be265fe63caac4c6 script_sha256: 4cace2eb41ab88a71f1bc9b585123a421d4970358090b99fe7de405bcc5d0469 python: 3.12.6 platform: Linux-6.8.0-138-generic-x86_64-with-glibc2.35 host: lab relflow_path: /home/grantham/relflow-proof-scripts.RB4VJH/src/relflow packages: relflow: 0.2.1 torch: 2.12.0 lightning: 2.6.4 numpy: 2.4.6 pyarrow: 24.0.0 polars: 1.44.1 torchmetrics: 1.9.0 tensordict: 0.12.4 pydantic: 2.13.4 cuda: '13.0' gpu: NVIDIA GeForce RTX 3090 code_sha256: b9daad4e5b1827cb5da87393719da162e3431161105edf2eeda124d237114638 metrics: intact_rmse: 0.011463659567015207 baseline_rmse: 0.7110739767923704 intact_nrmse: 0.016121613138941424 checks: Baseline RMSE > 0.000001: true Supplied-mean subtraction nRMSE < 0.25: true Finite calibration and model metrics: true outcome: met duration_seconds: 17.817 P035: script: proofs/relational/ungrouped_peer_deviation.py page: docs/proofs/relational/ungrouped-peer-deviation.py historical: status: Provisional pass source_revision: 810b3fe3b0b25c2c7a718571d83f6006f23110c3 sources: - https://github.com/relflow/relflow/blob/810b3fe/proofs/relational/peer_relative_inference/test_ungrouped_peer_deviation.py - https://github.com/relflow/relflow/blob/810b3fe/proofs/relational/peer_relative_inference/support.py summary: |- The checked-in report records **0.091 nRMSE**, rising to **0.283** after a common translation, on one deterministic CPU seed. Permutation drift is **0.101** in baseline-error units. Invariance is approximate at these gates. insights: |- **The model can compare each value with the collection’s average without being given that average.** Independently varying collection locations make an item's raw value alone a poor predictor. One learned branch summary works together with each repeated target's aligned visible value; the summary is not the decoder's only source of information. Common translations preserve the correct deviations, while complete-item permutations should move predictions with their items. Both controls pass, although error increases after translation and permutation drift is nonzero. The result supports approximate aggregation and routing for six-item collections. It does not establish exact set equivariance, variable-length behavior, or peer selection by category. evidence: |- The protocol uses 1,536 training, 384 validation, and 768 test records, each with six items, over 900 CPU steps. nRMSE and permutation drift are normalized by the constant training-target mean baseline's test RMSE. | Check | Required gate | Recorded result | | --- | --- | --- | | Raw peer deviation | nRMSE < 0.35 | 0.091 | | Common translation | nRMSE < 0.45 | 0.283 | | Complete-item permutation drift | < 0.20 | 0.101 | runs: - started_at: '2026-09-15T02:16:21.437497+00:00' mode: smoke seed: 3710 steps_override: 2 accelerator: gpu threads: 1 provenance: revision: null source_sha256: 94205b0b7969106cc7f8b4127f4daff2fb705a02e4c5ffba78a9855fe4d2d18a script_sha256: 378e7fa709c17675af94bd36ec5c9c4714366d1ccc32c564bba1f9835ba949da python: 3.12.6 platform: Linux-6.8.0-138-generic-x86_64-with-glibc2.35 host: lab packages: relflow: 0.2.1 torch: 2.12.0 lightning: 2.6.4 numpy: 2.4.6 cuda: '13.0' gpu: NVIDIA GeForce RTX 3090 code_sha256: a889f86c2492f28476ab519d32d2c6775442cc0347d5393fb610c264313ff583 metrics: intact_rmse: 0.7623859710556313 baseline_rmse: 0.731698187191497 intact_nrmse: 1.0419404945937125 translated_rmse: 0.7647988210717267 translated_nrmse: 1.045238097428232 permutation_drift: 0.01002924168988225 checks: Baseline RMSE > 0.000001: true Training deviations are exactly centered: true Translation preserves target shape: true Translation preserves all targets: true Permutation preserves the target multiset: true Normalized permutation drift < 0.20: true Ungrouped deviation nRMSE < 0.35: false Common-translated nRMSE < 0.45: false Finite calibration and model metrics: true outcome: not_met duration_seconds: 1.012 - started_at: '2026-09-15T02:30:40.787058+00:00' mode: full seed: 3710 steps_override: null accelerator: gpu threads: 1 provenance: revision: null source_sha256: 60d95a6154cb98ceffe4e920888974f8c91f232a70dd01d4053d948651ff683c script_sha256: 378e7fa709c17675af94bd36ec5c9c4714366d1ccc32c564bba1f9835ba949da python: 3.12.6 platform: Linux-6.8.0-138-generic-x86_64-with-glibc2.35 host: lab relflow_path: /home/grantham/relflow-proof-scripts.RB4VJH/src/relflow packages: relflow: 0.2.1 torch: 2.12.0 lightning: 2.6.4 numpy: 2.4.6 pyarrow: 24.0.0 polars: 1.44.1 torchmetrics: 1.9.0 tensordict: 0.12.4 pydantic: 2.13.4 cuda: '13.0' gpu: NVIDIA GeForce RTX 3090 code_sha256: a889f86c2492f28476ab519d32d2c6775442cc0347d5393fb610c264313ff583 metrics: intact_rmse: 0.015504269163810253 baseline_rmse: 0.731698187191497 intact_nrmse: 0.0211894322484532 translated_rmse: 0.06316018432937254 translated_nrmse: 0.08631999564164905 permutation_drift: 0.0038441138616867295 checks: Baseline RMSE > 0.000001: true Training deviations are exactly centered: true Translation preserves target shape: true Translation preserves all targets: true Permutation preserves the target multiset: true Normalized permutation drift < 0.20: true Ungrouped deviation nRMSE < 0.35: true Common-translated nRMSE < 0.45: true Finite calibration and model metrics: true outcome: met duration_seconds: 26.877 P036: script: proofs/relational/category_identity.py page: docs/proofs/relational/category-identity.py historical: status: Provisional pass source_revision: 810b3fe3b0b25c2c7a718571d83f6006f23110c3 sources: - https://github.com/relflow/relflow/blob/810b3fe/proofs/relational/sibling_entity_transfer/test_category_identity.py - https://github.com/relflow/relflow/blob/810b3fe/proofs/relational/sibling_entity_transfer/__init__.py summary: |- The checked-in report records **0.0046 nRMSE** with token preservation and **0.0198** with fixed-width compression in the bounded two-category case. Both routes pass on the calibrated seed; larger category spaces are untested. insights: |- **The model transfers changing values by identity across independently ordered branches.** The two categories recur, but their source values are resampled for every row, so remembering one value per category cannot solve the task. Visible target identities can condition repeated decoder queries over source evidence routed through the root. Rotating target keys while retaining their original values removes accuracy, supporting keyed transfer rather than positional copying. Both retained and compressed routes work in this small persistent-vocabulary case. That does not establish compressed retrieval for unseen identities or larger category spaces. Treat it as evidence of a learned communication path; use an exact join when the matching rule and transferred value must be guaranteed. evidence: |- Each route trains on 2,048 records, validates on 512, and tests 1,024 over 600 CPU steps. nRMSE divides target error by the error of predicting zero. | Check | Required gate | Recorded result | | --- | --- | --- | | Retained-token transfer | nRMSE ≤ 0.25 | 0.0046 | | Compressed Category control | nRMSE ≤ 0.25 | 0.0198 | | Broken target identities | nRMSE ≥ 0.80 | Gate reported passing | | Broken-key gap | At least 0.50 nRMSE | Gate reported passing | runs: - started_at: '2026-09-15T02:16:25.845731+00:00' mode: smoke seed: 17 steps_override: 2 accelerator: gpu threads: 1 provenance: revision: null source_sha256: 74977865e512339bb5cf106f7386d69930ad8e388c1f5e5dbc6ab6fa7200b5ed script_sha256: f4fbf560e15c41cd8d90995f93b5a0e2468299c80e35daad69c6ff5f6c16c924 python: 3.12.6 platform: Linux-6.8.0-138-generic-x86_64-with-glibc2.35 host: lab packages: relflow: 0.2.1 torch: 2.12.0 lightning: 2.6.4 numpy: 2.4.6 cuda: '13.0' gpu: NVIDIA GeForce RTX 3090 code_sha256: 2ab0faeb6cf13b0c0171b7dafa5a0a3c5059897a0099f7d94950f06a62f9b335 metrics: compressed_nrmse: 1.0569129682637717 compressed_broken_nrmse: 1.0557399192276302 preserved_nrmse: 0.9541982666300821 preserved_broken_nrmse: 0.9526219724384437 checks: Finite errors: true Preserved transfer calibration nRMSE <= 1.10: true Preserved transfer nRMSE <= 0.25: false Broken identities nRMSE >= 0.80: true Identity corruption increases nRMSE by >= 0.50: false Compressed Category transfer nRMSE <= 0.25: false outcome: not_met duration_seconds: 1.25 - started_at: '2026-09-15T02:31:11.132537+00:00' mode: full seed: 17 steps_override: null accelerator: gpu threads: 1 provenance: revision: null source_sha256: cf819d1bef585c450ece5954dfc33af65a41b113031718b2e329be42f142cd90 script_sha256: f4fbf560e15c41cd8d90995f93b5a0e2468299c80e35daad69c6ff5f6c16c924 python: 3.12.6 platform: Linux-6.8.0-138-generic-x86_64-with-glibc2.35 host: lab relflow_path: /home/grantham/relflow-proof-scripts.RB4VJH/src/relflow packages: relflow: 0.2.1 torch: 2.12.0 lightning: 2.6.4 numpy: 2.4.6 pyarrow: 24.0.0 polars: 1.44.1 torchmetrics: 1.9.0 tensordict: 0.12.4 pydantic: 2.13.4 cuda: '13.0' gpu: NVIDIA GeForce RTX 3090 code_sha256: 2ab0faeb6cf13b0c0171b7dafa5a0a3c5059897a0099f7d94950f06a62f9b335 metrics: compressed_nrmse: 0.0073523894407589085 compressed_broken_nrmse: 1.4491613336984703 preserved_nrmse: 0.026091784336623156 preserved_broken_nrmse: 1.459659564323525 checks: Finite errors: true Preserved transfer calibration nRMSE <= 1.10: true Preserved transfer nRMSE <= 0.25: true Broken identities nRMSE >= 0.80: true Identity corruption increases nRMSE by >= 0.50: true Compressed Category transfer nRMSE <= 0.25: true outcome: met duration_seconds: 37.629 P037: script: proofs/relational/hash_identity.py page: docs/proofs/relational/hash-identity.py historical: status: Provisional pass source_revision: 810b3fe3b0b25c2c7a718571d83f6006f23110c3 sources: - https://github.com/relflow/relflow/blob/810b3fe/proofs/relational/sibling_entity_transfer/test_hash_identity.py - https://github.com/relflow/relflow/blob/810b3fe/proofs/relational/sibling_entity_transfer/__init__.py summary: |- The checked-in report records passing accuracy, broken-key, and compression comparison gates on one deterministic CPU seed with two entities. Exact Hash route scores are not recorded. insights: |- **Fresh identities can connect source values to targets in another branch without a learned identity vocabulary.** Shared Hash encoding preserves within-batch equality. Coordinate-local encoding binds each source key to its value, and visible target keys can condition repeated queries over source context carried through the root. Independent branch orders remove reliable positional copying. Rotating only target keys while retaining target values breaks accuracy, showing that the association matters. Retaining tokens also outperforms the tested compressed route, unlike the small persistent-Category control. This is evidence for two entities under one protocol, not an exact join or unlimited memory. Missing sources, duplicate identities, and collisions require explicit conventions and further tests. evidence: |- Both routes use 2,048 training, 512 validation, and 1,024 test records with 600 CPU steps each. nRMSE is relative to predicting zero for every value. | Check | Required gate | Recorded result | | --- | --- | --- | | Retained-token transfer | nRMSE ≤ 0.25 | Gate reported passing | | Improvement over compression | At least 0.15 nRMSE | Gate reported passing | | Broken target identities | nRMSE ≥ 0.80 | Gate reported passing | | Broken-key gap | At least 0.50 nRMSE | Gate reported passing | runs: - started_at: '2026-09-15T02:16:30.528071+00:00' mode: smoke seed: 17 steps_override: 2 accelerator: gpu threads: 1 provenance: revision: null source_sha256: 3b6916162868132177eb181a2fc1857a4fabf105cb8aa04fd7c302b6fbac7298 script_sha256: 6041f0d886b9eed53cb50017003dbb7b7537de816d04575d65cbd9282c9ab544 python: 3.12.6 platform: Linux-6.8.0-138-generic-x86_64-with-glibc2.35 host: lab packages: relflow: 0.2.1 torch: 2.12.0 lightning: 2.6.4 numpy: 2.4.6 cuda: '13.0' gpu: NVIDIA GeForce RTX 3090 code_sha256: d6ded589e78dd1d78616e27fbe906f2c5f43a66bb0872bd212b8f24663d3a259 metrics: compressed_nrmse: 1.0592964517256729 compressed_broken_nrmse: 1.054979311079157 preserved_nrmse: 1.0359586409059693 preserved_broken_nrmse: 1.0312150015201824 checks: Finite errors: true Preserved transfer calibration nRMSE <= 1.10: true Preserved transfer nRMSE <= 0.25: false Broken identities nRMSE >= 0.80: true Identity corruption increases nRMSE by >= 0.50: false Preserving Hash tokens improves nRMSE by >= 0.15: false outcome: not_met duration_seconds: 1.295 - started_at: '2026-09-15T02:31:52.200292+00:00' mode: full seed: 17 steps_override: null accelerator: gpu threads: 1 provenance: revision: null source_sha256: 8c2c1cfe503e109ba3f9592952c0b0242c5ed9aed61bafc84f4dcae324ade526 script_sha256: 6041f0d886b9eed53cb50017003dbb7b7537de816d04575d65cbd9282c9ab544 python: 3.12.6 platform: Linux-6.8.0-138-generic-x86_64-with-glibc2.35 host: lab relflow_path: /home/grantham/relflow-proof-scripts.RB4VJH/src/relflow packages: relflow: 0.2.1 torch: 2.12.0 lightning: 2.6.4 numpy: 2.4.6 pyarrow: 24.0.0 polars: 1.44.1 torchmetrics: 1.9.0 tensordict: 0.12.4 pydantic: 2.13.4 cuda: '13.0' gpu: NVIDIA GeForce RTX 3090 code_sha256: d6ded589e78dd1d78616e27fbe906f2c5f43a66bb0872bd212b8f24663d3a259 metrics: compressed_nrmse: 0.7247225475245465 compressed_broken_nrmse: 0.7231541658372164 preserved_nrmse: 0.08587741211423401 preserved_broken_nrmse: 1.4392457107919758 checks: Finite errors: true Preserved transfer calibration nRMSE <= 1.10: true Preserved transfer nRMSE <= 0.25: true Broken identities nRMSE >= 0.80: true Identity corruption increases nRMSE by >= 0.50: true Preserving Hash tokens improves nRMSE by >= 0.15: true outcome: met duration_seconds: 37.497 P038: script: proofs/state/null_versus_zero.py page: docs/proofs/state/null-versus-zero.py historical: status: Provisional pass source_revision: 810b3fe3b0b25c2c7a718571d83f6006f23110c3 sources: - https://github.com/relflow/relflow/blob/810b3fe/proofs/state/missingness/test_null_versus_zero.py summary: |- The source records one seed passing the signal, independent-validity, and null-to-zero controls. It does not record the individual metric values. insights: |- **A null and a real zero carry different information, even when every present number is zero.** Number encodes value state separately from continuous content, giving the model a route to recognize missingness. The recorded signal model clears both AUC and accuracy gates. Independent validity stays at chance, and replacing nulls with zeros destroys the signal while retaining the original labels. The latter creates identical inputs that require different answers. Preserve nulls when that distinction matters. Imputation can erase useful information before training begins. This proof establishes the representation's capability; it does not show that missingness is predictive in every application or cover every kind of absent data. evidence: |- | Evaluation | Required gate | Recorded result | | --- | --- | --- | | Validity signal | AUC ≥ 0.99; accuracy ≥ 0.98 | Gates reported passing | | Independent validity | AUC 0.42–0.58; accuracy 0.45–0.55 | Gates reported passing | | Nulls replaced with zero | AUC 0.42–0.58; accuracy 0.45–0.55 | Gates reported passing | | Signal versus independent validity | AUC improvement ≥ 0.40 | Gate reported passing | Accuracy uses a fixed probability threshold of 0.5. The experiment supports preserving nulls when missingness is meaningful; filling them with a real value can erase the only available signal. runs: - started_at: '2026-09-15T02:16:35.227282+00:00' mode: smoke seed: 11 steps_override: 2 accelerator: gpu threads: 1 provenance: revision: null source_sha256: f8d8f0b81af72508d75a3cfa5e48b475bc8c949e8a455dcceea3a47fb988770b script_sha256: 7baf07e28cf2b94670a1d7887577adbb24ac6f9707745ef9ed186c4d9e0b2028 python: 3.12.6 platform: Linux-6.8.0-138-generic-x86_64-with-glibc2.35 host: lab packages: relflow: 0.2.1 torch: 2.12.0 lightning: 2.6.4 numpy: 2.4.6 cuda: '13.0' gpu: NVIDIA GeForce RTX 3090 code_sha256: b3c605573856a1df5e385fe8d903293dadb821731e05695d736b3f7345b4e80e metrics: state_auc: 1.0 state_accuracy: 1.0 independent_auc: 0.49365234375 independent_accuracy: 0.49365234375 prefilled_auc: 0.5 prefilled_accuracy: 0.5 auc_gap: 0.50634765625 checks: Null state reaches 0.99 AUC: true Null state reaches 0.98 accuracy: true Independent validity AUC remains between 0.42 and 0.58: true Independent validity accuracy remains between 0.45 and 0.55: true Null-to-zero AUC remains between 0.42 and 0.58: true Null-to-zero accuracy remains between 0.45 and 0.55: true State exceeds its control by at least 0.40 AUC: true outcome: met duration_seconds: 1.18 - started_at: '2026-09-15T02:25:41.822592+00:00' mode: full seed: 11 steps_override: null accelerator: cpu threads: 1 provenance: revision: null source_sha256: 96266262609686640fc4fd3293e42c70865a5abcb7a1de6fc3a127d4835c9f9c script_sha256: 7baf07e28cf2b94670a1d7887577adbb24ac6f9707745ef9ed186c4d9e0b2028 python: 3.12.6 platform: Linux-6.8.0-138-generic-x86_64-with-glibc2.35 host: lab relflow_path: /home/grantham/relflow-proof-scripts.RB4VJH/src/relflow packages: relflow: 0.2.1 torch: 2.12.0 lightning: 2.6.4 numpy: 2.4.6 pyarrow: 24.0.0 polars: 1.44.1 torchmetrics: 1.9.0 tensordict: 0.12.4 pydantic: 2.13.4 cuda: '13.0' gpu: NVIDIA GeForce RTX 3090 code_sha256: b3c605573856a1df5e385fe8d903293dadb821731e05695d736b3f7345b4e80e metrics: state_auc: 1.0 state_accuracy: 1.0 independent_auc: 0.50634765625 independent_accuracy: 0.5 prefilled_auc: 0.5 prefilled_accuracy: 0.5 auc_gap: 0.49365234375 checks: Null state reaches 0.99 AUC: true Null state reaches 0.98 accuracy: true Independent validity AUC remains between 0.42 and 0.58: true Independent validity accuracy remains between 0.45 and 0.55: true Null-to-zero AUC remains between 0.42 and 0.58: true Null-to-zero accuracy remains between 0.45 and 0.55: true State exceeds its control by at least 0.40 AUC: true outcome: met duration_seconds: 2.408 - started_at: '2026-09-15T02:32:33.163989+00:00' mode: full seed: 11 steps_override: null accelerator: gpu threads: 1 provenance: revision: null source_sha256: 96266262609686640fc4fd3293e42c70865a5abcb7a1de6fc3a127d4835c9f9c script_sha256: 7baf07e28cf2b94670a1d7887577adbb24ac6f9707745ef9ed186c4d9e0b2028 python: 3.12.6 platform: Linux-6.8.0-138-generic-x86_64-with-glibc2.35 host: lab relflow_path: /home/grantham/relflow-proof-scripts.RB4VJH/src/relflow packages: relflow: 0.2.1 torch: 2.12.0 lightning: 2.6.4 numpy: 2.4.6 pyarrow: 24.0.0 polars: 1.44.1 torchmetrics: 1.9.0 tensordict: 0.12.4 pydantic: 2.13.4 cuda: '13.0' gpu: NVIDIA GeForce RTX 3090 code_sha256: b3c605573856a1df5e385fe8d903293dadb821731e05695d736b3f7345b4e80e metrics: state_auc: 1.0 state_accuracy: 1.0 independent_auc: 0.50634765625 independent_accuracy: 0.5 prefilled_auc: 0.5 prefilled_accuracy: 0.5 auc_gap: 0.49365234375 checks: Null state reaches 0.99 AUC: true Null state reaches 0.98 accuracy: true Independent validity AUC remains between 0.42 and 0.58: true Independent validity accuracy remains between 0.45 and 0.55: true Null-to-zero AUC remains between 0.42 and 0.58: true Null-to-zero accuracy remains between 0.45 and 0.55: true State exceeds its control by at least 0.40 AUC: true outcome: met duration_seconds: 3.926 P039: script: proofs/structure/attention_learns_fixed_width_sum.py page: docs/proofs/structure/attention-learns-fixed-width-sum.py historical: status: Provisional pass source_revision: 810b3fe3b0b25c2c7a718571d83f6006f23110c3 sources: - https://github.com/relflow/relflow/blob/810b3fe/proofs/structure/hierarchical_statistics/test_attention_learns_fixed_width_sum.py - https://github.com/relflow/relflow/blob/810b3fe/proofs/structure/hierarchical_statistics/support.py summary: |- The source records one seed clearing the normalized-error gate. It does not record the exact error. This is a fixed-length positive control, with no nested grouping or variable-length generalization claim. insights: |- **The model can learn to approximately add six visible numbers.** All six independently drawn values are visible, making this a useful positive control before testing more complicated nested statistics. The recorded model clears normalized RMSE below 0.25 against a constant training-mean baseline. That demonstrates substantially lower held-out error, but no exact error value or prediction-level permutation control is reported. The task always contains six items. Do not extend this result to variable cardinality, arbitrary hierarchy depth, or exact arithmetic. The nested-cardinality case asks whether local structure survives successive reductions; this control only establishes that the model can learn a numerical sum at one level. evidence: |- | Evaluation | Required gate | Recorded result | | --- | --- | --- | | Held-out total prediction | Normalized RMSE < 0.25 | Gate reported passing | Normalized RMSE is model RMSE divided by the RMSE of always predicting the training target mean. Clearing 0.25 means at least a 75% reduction in this error measure against that baseline. The test evaluates the sum directly. Although the target is mathematically unchanged by transaction reordering, this implementation does not assert a prediction-level permutation check. runs: - started_at: '2026-09-15T02:16:39.941841+00:00' mode: smoke seed: 2600 steps_override: 2 accelerator: gpu threads: 1 provenance: revision: null source_sha256: 4f1b50be78a216f337ac47c38eee09faae0b211abc4a8ef7c5af075296ee988a script_sha256: ce6942ec566a85121547afac7ec3726d3f39bebfdeab5164141e0c45bb6314c3 python: 3.12.6 platform: Linux-6.8.0-138-generic-x86_64-with-glibc2.35 host: lab packages: relflow: 0.2.1 torch: 2.12.0 lightning: 2.6.4 numpy: 2.4.6 cuda: '13.0' gpu: NVIDIA GeForce RTX 3090 code_sha256: 1579dd7d142c8799060b7c2f06fbc33b2ee7755a721f37420412d8c1e3500f76 metrics: steps: 2 baseline_rmse: 1.395327666226233 validate_rmse: 1.3016960558014405 test_rmse: 1.219479366755207 test_nrmse: 0.8739734732368486 target_std: 1.3947314823294805 checks: Test normalized RMSE is below 0.25: false outcome: not_met duration_seconds: 0.669 - started_at: '2026-09-15T02:32:40.570072+00:00' mode: full seed: 2600 steps_override: null accelerator: gpu threads: 1 provenance: revision: null source_sha256: 4d533a8f738c4ada1fb839506ccb931b32a981f27d52ecef16b4645f5b33eb0e script_sha256: ce6942ec566a85121547afac7ec3726d3f39bebfdeab5164141e0c45bb6314c3 python: 3.12.6 platform: Linux-6.8.0-138-generic-x86_64-with-glibc2.35 host: lab relflow_path: /home/grantham/relflow-proof-scripts.RB4VJH/src/relflow packages: relflow: 0.2.1 torch: 2.12.0 lightning: 2.6.4 numpy: 2.4.6 pyarrow: 24.0.0 polars: 1.44.1 torchmetrics: 1.9.0 tensordict: 0.12.4 pydantic: 2.13.4 cuda: '13.0' gpu: NVIDIA GeForce RTX 3090 code_sha256: 1579dd7d142c8799060b7c2f06fbc33b2ee7755a721f37420412d8c1e3500f76 metrics: steps: 400 baseline_rmse: 1.395327666226233 validate_rmse: 0.023647580528284976 test_rmse: 0.023375447649324754 test_nrmse: 0.0167526583290256 target_std: 1.3947314823294805 checks: Test normalized RMSE is below 0.25: true outcome: met duration_seconds: 10.593 P040: script: proofs/structure/attention_preserves_nested_cardinality.py page: docs/proofs/structure/attention-preserves-nested-cardinality.py historical: status: Provisional pass source_revision: 810b3fe3b0b25c2c7a718571d83f6006f23110c3 sources: - https://github.com/relflow/relflow/blob/810b3fe/proofs/structure/hierarchical_statistics/test_attention_preserves_nested_cardinality.py - https://github.com/relflow/relflow/blob/810b3fe/proofs/structure/hierarchical_statistics/support.py summary: |- The source records one seed beating the flattened-information oracle and recovering about **84%** of the true difference between regrouped pairs. Evidence is limited to two partitions of repeated equal values. insights: |- **How values are grouped can change the answer even when the values themselves stay the same.** The matched observations contain the same six amounts but divide them into different session sizes, changing the largest session total. Attention reduction includes an additive sum and count contribution alongside normalized attention. The recorded model beats the best flattened-information prediction and recovers about 84% of the true pair difference. The difference metric uses absolute magnitudes; the separate RMSE gate checks closeness to actual targets. This supports the tested route through two partitions of equal values. It does not establish exact nested arithmetic or arbitrary capacity. Varying amounts within sessions and testing more partitions remain necessary extensions. evidence: |- | Evaluation | Required gate | Recorded result | | --- | --- | --- | | Held-out RMSE versus flattened oracle | Less than 75% of oracle RMSE | Gate reported passing | | Mean absolute predicted pair difference | At least 75% of true pair difference | Approximately 84% recovered | The difference metric uses absolute magnitudes. The separate RMSE gate checks whether predictions also approach their actual targets. This supports the tested learned hierarchy, not exact arithmetic for arbitrary collections. runs: - started_at: '2026-09-15T02:16:44.262970+00:00' mode: smoke seed: 2601 steps_override: 2 accelerator: gpu threads: 1 provenance: revision: null source_sha256: e1d35c792ec05ebf703fc1845e370643161bfdfdd922ecd91d71346231885d1a script_sha256: e8c9b24b1f5c60412d8fc1e63466176cc32a1dbcfb2428c80f7f4e546839ee6b python: 3.12.6 platform: Linux-6.8.0-138-generic-x86_64-with-glibc2.35 host: lab packages: relflow: 0.2.1 torch: 2.12.0 lightning: 2.6.4 numpy: 2.4.6 cuda: '13.0' gpu: NVIDIA GeForce RTX 3090 code_sha256: 6b34f02b6be45d51b86a74a95ca42dc32b49eeeca4b4810c5042b00815fb1ec9 metrics: steps: 2 flat_oracle_rmse: 1.0647465517891985 test_rmse: 1.912723549487529 predicted_pair_delta: 0.0004965085536241531 true_pair_delta: 2.010377930933031 recovered_pair_delta_fraction: 0.00024697274377346543 checks: Attention improves over the flattened oracle by at least 25%: false Attention recovers at least 75% of the pair delta: false outcome: not_met duration_seconds: 0.624 - started_at: '2026-09-15T02:32:54.629950+00:00' mode: full seed: 2601 steps_override: null accelerator: gpu threads: 1 provenance: revision: null source_sha256: ea54b5d38d1dc4dfbc86298f368ee6648c70ca67934a913cf4e0fd9b4d3ff914 script_sha256: e8c9b24b1f5c60412d8fc1e63466176cc32a1dbcfb2428c80f7f4e546839ee6b python: 3.12.6 platform: Linux-6.8.0-138-generic-x86_64-with-glibc2.35 host: lab relflow_path: /home/grantham/relflow-proof-scripts.RB4VJH/src/relflow packages: relflow: 0.2.1 torch: 2.12.0 lightning: 2.6.4 numpy: 2.4.6 pyarrow: 24.0.0 polars: 1.44.1 torchmetrics: 1.9.0 tensordict: 0.12.4 pydantic: 2.13.4 cuda: '13.0' gpu: NVIDIA GeForce RTX 3090 code_sha256: 6b34f02b6be45d51b86a74a95ca42dc32b49eeeca4b4810c5042b00815fb1ec9 metrics: steps: 80 flat_oracle_rmse: 1.0647465517891985 test_rmse: 0.2992299183830936 predicted_pair_delta: 1.8223806042224169 true_pair_delta: 2.010377930933031 recovered_pair_delta_fraction: 0.9064865745798536 checks: Attention improves over the flattened oracle by at least 25%: true Attention recovers at least 75% of the pair delta: true outcome: met duration_seconds: 2.859 P041: script: proofs/structure/mean_erases_cardinality.py page: docs/proofs/structure/mean-erases-cardinality.py historical: status: Expected limitation source_revision: 810b3fe3b0b25c2c7a718571d83f6006f23110c3 sources: - https://github.com/relflow/relflow/blob/810b3fe/proofs/structure/hierarchical_statistics/test_mean_erases_cardinality.py - https://github.com/relflow/relflow/blob/810b3fe/proofs/structure/hierarchical_statistics/support.py summary: |- The source records one seed learning the average while producing identical total predictions for one and six equal values. Passing demonstrates both a useful average representation and deliberate count erasure in this route. insights: |- **With branch attention disabled, averaging identical encoded values erases how many copies were present.** One amount and six copies yield the same branch summary, so a downstream decoder cannot recover their different totals from that summary. The recorded model learns the average but produces identical total predictions for the explicit one-versus-six pair. This is an expected information boundary, not simply insufficient training. Mean reduces encoded tokens rather than calculating a raw numerical average directly. Choose a summary that retains what the parent needs. The result specifically uses `attention=None`; preceding attention can change the available information. Exact business totals belong in preprocessing, while learned alternatives need their own behavioral checks. evidence: |- | Evaluation | Required gate | Recorded result | | --- | --- | --- | | Held-out mean prediction | Normalized RMSE < 0.25 | Gate reported passing | | Total predictions for one versus six equal amounts | Equal within absolute tolerance 0.000001 | Gate reported passing | Normalized RMSE divides model error by the error of predicting the training target mean. `Mean` reduces encoded tokens; it is not an exact raw-number averaging operator. This count-erasure result specifically disables branch attention and does not characterize every architecture that uses Mean. runs: - started_at: '2026-09-15T02:16:48.536440+00:00' mode: smoke seed: 2599 steps_override: 2 accelerator: gpu threads: 1 provenance: revision: null source_sha256: 75235cae5c0e5c0258784aed316466a0a0c1d98394714a0e1396b453cd6b5cb6 script_sha256: a20f560bbcd9e75aa15ba4dbf243a45a13ad4a00c0dfd1dc6efee4bb6d0fadb3 python: 3.12.6 platform: Linux-6.8.0-138-generic-x86_64-with-glibc2.35 host: lab packages: relflow: 0.2.1 torch: 2.12.0 lightning: 2.6.4 numpy: 2.4.6 cuda: '13.0' gpu: NVIDIA GeForce RTX 3090 code_sha256: 6e8c37bb90f8785db9e235c6d6bf104191faf7329c8b31ed085290ebbcdcb1b2 metrics: steps: 2 mean_baseline_rmse: 0.2930153390602116 mean_rmse: 0.1510877962918501 mean_nrmse: 0.5156310136405627 paired_total_predictions: - 0.8707736730575562 - 0.8707737326622009 checks: Mean normalized RMSE is below 0.25: false One and six equal values produce the same total prediction: true outcome: not_met duration_seconds: 0.655 - started_at: '2026-09-15T02:33:00.987438+00:00' mode: full seed: 2599 steps_override: null accelerator: gpu threads: 1 provenance: revision: null source_sha256: d0d1e7ff37eb0e1e36df063b960ad31ae3dadad9f8fc9d91ca5b1677b84a5262 script_sha256: a20f560bbcd9e75aa15ba4dbf243a45a13ad4a00c0dfd1dc6efee4bb6d0fadb3 python: 3.12.6 platform: Linux-6.8.0-138-generic-x86_64-with-glibc2.35 host: lab relflow_path: /home/grantham/relflow-proof-scripts.RB4VJH/src/relflow packages: relflow: 0.2.1 torch: 2.12.0 lightning: 2.6.4 numpy: 2.4.6 pyarrow: 24.0.0 polars: 1.44.1 torchmetrics: 1.9.0 tensordict: 0.12.4 pydantic: 2.13.4 cuda: '13.0' gpu: NVIDIA GeForce RTX 3090 code_sha256: 6e8c37bb90f8785db9e235c6d6bf104191faf7329c8b31ed085290ebbcdcb1b2 metrics: steps: 300 mean_baseline_rmse: 0.2930153390602116 mean_rmse: 0.004598747738179377 mean_nrmse: 0.015694563134233674 paired_total_predictions: - 4.13906192779541 - 4.139062881469727 checks: Mean normalized RMSE is below 0.25: true One and six equal values produce the same total prediction: true outcome: met duration_seconds: 7.367 P042: script: proofs/temporal/business_window_requires_composition.py page: docs/proofs/temporal/business-window-requires-composition.py historical: status: Provisional pass source_revision: 810b3fe3b0b25c2c7a718571d83f6006f23110c3 sources: - https://github.com/relflow/relflow/blob/810b3fe/proofs/temporal/datepart_periodicity/test_business_window_requires_composition.py - https://github.com/relflow/relflow/blob/810b3fe/proofs/temporal/datepart_periodicity/support.py summary: |- The source records accuracy **1.000** with both coordinates, compared with **0.500** from weekday and **0.750** from hour alone. Evidence covers one deterministic model seed. insights: |- **Weekday and hour together identify a business window that neither coordinate resolves alone.** A weekday evening and a weekend morning each share one coordinate with an open observation, so either single-coordinate representation loses a necessary distinction. The recorded combined accuracy is 1.000, compared with 0.500 from weekday and 0.750 from hour alone. The weekday fit does not reach all the skill available from its marginal signal; these are observed fits, not three equally optimized solutions. This supports learned coordinate composition in the constructed calendar. It does not cover holidays, timezones, or daylight-saving rules. Derive fixed business logic exactly in preprocessing when learning an approximation is unnecessary. evidence: |- | Visible coordinates | Required gate | Recorded accuracy | | --- | --- | --- | | Weekday | At most 0.82 | 0.500 | | Hour | At most 0.82 | 0.750 | | Weekday and hour | At least 0.95 | 1.000 | | Improvement over stronger control | At least 0.15 | 0.250 | The weekday-only fit does not reach the approximately 0.75 accuracy available from its marginal signal. The result still establishes that the combined representation clears the target gate; it does not claim equally complete optimization of every control. For a fixed business-calendar rule, derive the result exactly in [preprocessing](../../guides/preprocessors.qmd). Here, learning the rule serves as a controlled test of coordinate composition. runs: - started_at: '2026-09-15T02:16:52.837717+00:00' mode: smoke seed: 59 steps_override: 2 accelerator: gpu threads: 1 provenance: revision: null source_sha256: f9d64288f0f466605e863191ccaaa604439c5c5c9111328694622446bd14eee4 script_sha256: 561c2cce30d0fa236db8be35d37ffa372273d89af735ae9dcaf6d24692dd5124 python: 3.12.6 platform: Linux-6.8.0-138-generic-x86_64-with-glibc2.35 host: lab packages: relflow: 0.2.1 torch: 2.12.0 lightning: 2.6.4 numpy: 2.4.6 cuda: '13.0' gpu: NVIDIA GeForce RTX 3090 code_sha256: 6417616c6292f48eaa814977e23b11536760703c6530296ab922189ebe603d79 metrics: day_only_accuracy: 0.5 hour_only_accuracy: 0.5 composed_accuracy: 0.900390625 accuracy_gap: 0.400390625 checks: Weekday-only accuracy is at most 0.82: true Hour-only accuracy is at most 0.82: true Combined coordinates reach 0.95 accuracy: false Composition improves over either coordinate by at least 0.15: true outcome: not_met duration_seconds: 1.127 - started_at: '2026-09-15T02:33:11.826682+00:00' mode: full seed: 59 steps_override: null accelerator: gpu threads: 1 provenance: revision: null source_sha256: 7cbac86d045a9eb1a03079a4157637b249bc961a9ea22c63256c06bbcce68823 script_sha256: 561c2cce30d0fa236db8be35d37ffa372273d89af735ae9dcaf6d24692dd5124 python: 3.12.6 platform: Linux-6.8.0-138-generic-x86_64-with-glibc2.35 host: lab relflow_path: /home/grantham/relflow-proof-scripts.RB4VJH/src/relflow packages: relflow: 0.2.1 torch: 2.12.0 lightning: 2.6.4 numpy: 2.4.6 pyarrow: 24.0.0 polars: 1.44.1 torchmetrics: 1.9.0 tensordict: 0.12.4 pydantic: 2.13.4 cuda: '13.0' gpu: NVIDIA GeForce RTX 3090 code_sha256: 6417616c6292f48eaa814977e23b11536760703c6530296ab922189ebe603d79 metrics: day_only_accuracy: 0.75 hour_only_accuracy: 0.75 composed_accuracy: 1.0 accuracy_gap: 0.25 checks: Weekday-only accuracy is at most 0.82: true Hour-only accuracy is at most 0.82: true Combined coordinates reach 0.95 accuracy: true Composition improves over either coordinate by at least 0.15: true outcome: met duration_seconds: 11.27 P043: script: proofs/temporal/leap_boundary_requires_context.py page: docs/proofs/temporal/leap-boundary-requires-context.py historical: status: Expected limitation source_revision: 810b3fe3b0b25c2c7a718571d83f6006f23110c3 sources: - https://github.com/relflow/relflow/blob/810b3fe/proofs/temporal/datepart_periodicity/test_leap_boundary_requires_context.py - https://github.com/relflow/relflow/blob/810b3fe/proofs/temporal/datepart_periodicity/support.py summary: |- The recorded single-seed accuracy is **0.500** from day-of-year alone and **1.000** after adding week-of-month. Passing confirms an expected boundary and a successful way to restore the missing distinction. insights: |- **Day 60 can be February 29 or March 1, so its number alone cannot identify the month.** With balanced labels and only day-of-year visible, the model has no information that can resolve the two answers. The recorded accuracy is 0.500 for that route and 1.000 after adding week-of-month, which separates the dates. The models use different training budgets, but additional optimization alone cannot resolve identical inputs with conflicting labels. This is a representation boundary with a positive control. Supply the needed coordinate or calculate exact calendar results in preprocessing. Success with week-of-month does not establish general leap-year reasoning, timezone semantics, or reliability for every alternative coordinate choice. evidence: |- | Visible coordinates | Required gate | Recorded accuracy | | --- | --- | --- | | Day-of-year | 0.49–0.51 | 0.500 | | Day-of-year and week-of-month | At least 0.95 | 1.000 | | Improvement from added context | At least 0.40 | 0.500 | For exact month extraction, use a month coordinate or [preprocessing](../../guides/preprocessors.qmd). The proof examines what a learned representation can distinguish when that direct answer is omitted. runs: - started_at: '2026-09-15T02:16:57.612852+00:00' mode: smoke seed: 53 steps_override: 2 accelerator: gpu threads: 1 provenance: revision: null source_sha256: b2f7e05a951aff1161e8460b1a9ff353ad869b084e68addee4b991fd78758323 script_sha256: cdeb97aaf83051e0807a6ec177335ad911e0845c5148ebfa1967edfec648aebc python: 3.12.6 platform: Linux-6.8.0-138-generic-x86_64-with-glibc2.35 host: lab packages: relflow: 0.2.1 torch: 2.12.0 lightning: 2.6.4 numpy: 2.4.6 cuda: '13.0' gpu: NVIDIA GeForce RTX 3090 code_sha256: be99626909e5bdc858197826e8739bbd8ce65e40403f3fbb6e0bf1b80b206bca metrics: day_only_accuracy: 0.5 with_week_of_month_accuracy: 1.0 accuracy_gap: 0.5 checks: Day-of-year accuracy remains between 0.49 and 0.51: true Visible week-of-month accuracy reaches 0.95: true Visible week-of-month improves accuracy by at least 0.40: true outcome: met duration_seconds: 0.755 - started_at: '2026-09-15T02:33:26.590070+00:00' mode: full seed: 53 steps_override: null accelerator: gpu threads: 1 provenance: revision: null source_sha256: 861d768941ab50df484595ff65b2015661898a445630cbc59a9a5c6d36124ddd script_sha256: cdeb97aaf83051e0807a6ec177335ad911e0845c5148ebfa1967edfec648aebc python: 3.12.6 platform: Linux-6.8.0-138-generic-x86_64-with-glibc2.35 host: lab relflow_path: /home/grantham/relflow-proof-scripts.RB4VJH/src/relflow packages: relflow: 0.2.1 torch: 2.12.0 lightning: 2.6.4 numpy: 2.4.6 pyarrow: 24.0.0 polars: 1.44.1 torchmetrics: 1.9.0 tensordict: 0.12.4 pydantic: 2.13.4 cuda: '13.0' gpu: NVIDIA GeForce RTX 3090 code_sha256: be99626909e5bdc858197826e8739bbd8ce65e40403f3fbb6e0bf1b80b206bca metrics: day_only_accuracy: 0.5 with_week_of_month_accuracy: 1.0 accuracy_gap: 0.5 checks: Day-of-year accuracy remains between 0.49 and 0.51: true Visible week-of-month accuracy reaches 0.95: true Visible week-of-month improves accuracy by at least 0.40: true outcome: met duration_seconds: 3.024 P044: script: proofs/temporal/month_from_day_of_year.py page: docs/proofs/temporal/month-from-day-of-year.py historical: status: Provisional pass source_revision: 810b3fe3b0b25c2c7a718571d83f6006f23110c3 sources: - https://github.com/relflow/relflow/blob/810b3fe/proofs/temporal/datepart_periodicity/test_month_from_day_of_year.py - https://github.com/relflow/relflow/blob/810b3fe/proofs/temporal/datepart_periodicity/support.py summary: |- The source records held-out accuracy of **0.967**, falling to **0.110** after timestamp permutation. These are results from one deterministic model seed. insights: |- **In non-leap years, the day number is enough to learn which month a date belongs to.** DateParts supplies only day-of-year, and the model learns its month intervals from odd dates before testing on even dates in unseen years. The recorded accuracy is 0.967, dropping to 0.110 when timestamps are permuted while labels stay fixed. That supports use of the date–month relationship rather than success from label frequencies alone. Evidence covers one model seed and the deliberately restricted calendar. For exact month extraction, expose the month coordinate or derive it in preprocessing. This proof tests learned periodicity, not general calendar reasoning; leap-year ambiguity requires additional information. evidence: |- | Evaluation | Required gate | Recorded accuracy | | --- | --- | --- | | Held-out dates and years | At least 0.90 | 0.967 | | Permuted timestamps | At most 0.20 | 0.110 | | Original minus permuted | At least 0.65 | 0.857 | This is a learned periodicity experiment. For an application that simply needs the month, expose `month_of_year` or calculate it exactly in [preprocessing](../../guides/preprocessors.qmd). The [leap boundary](leap-boundary-requires-context.qmd) explains why this experiment deliberately excludes leap years. runs: - started_at: '2026-09-15T02:17:02.045083+00:00' mode: smoke seed: 41 steps_override: 2 accelerator: gpu threads: 1 provenance: revision: null source_sha256: 0fe6738a407c52e3e325562948f4799ea000018eec419d7c20a96975ae240aa6 script_sha256: 03b24a65c18f22b9fe7716f6cecd09449ab96557cca25763fa21a90a77708d1b python: 3.12.6 platform: Linux-6.8.0-138-generic-x86_64-with-glibc2.35 host: lab packages: relflow: 0.2.1 torch: 2.12.0 lightning: 2.6.4 numpy: 2.4.6 cuda: '13.0' gpu: NVIDIA GeForce RTX 3090 code_sha256: 6ea6c4b6783b52e864c4bb7aa57dced26d1d13671aa74c77a5c3fdd10f789f26 metrics: periodic_accuracy: 0.16033495962619781 permuted_accuracy: 0.0983966514468193 accuracy_gap: 0.06193830817937851 checks: Unseen-date month accuracy reaches 0.90: false Timestamp permutation accuracy is at most 0.20: true Original dates exceed permutation by at least 0.65 accuracy: false outcome: not_met duration_seconds: 0.668 - started_at: '2026-09-15T02:33:33.068764+00:00' mode: full seed: 41 steps_override: null accelerator: gpu threads: 1 provenance: revision: null source_sha256: cae4dc037ec67b465f2ab85f206e430dced41943a92a823f44372077271946b9 script_sha256: 03b24a65c18f22b9fe7716f6cecd09449ab96557cca25763fa21a90a77708d1b python: 3.12.6 platform: Linux-6.8.0-138-generic-x86_64-with-glibc2.35 host: lab relflow_path: /home/grantham/relflow-proof-scripts.RB4VJH/src/relflow packages: relflow: 0.2.1 torch: 2.12.0 lightning: 2.6.4 numpy: 2.4.6 pyarrow: 24.0.0 polars: 1.44.1 torchmetrics: 1.9.0 tensordict: 0.12.4 pydantic: 2.13.4 cuda: '13.0' gpu: NVIDIA GeForce RTX 3090 code_sha256: 6ea6c4b6783b52e864c4bb7aa57dced26d1d13671aa74c77a5c3fdd10f789f26 metrics: periodic_accuracy: 0.915900707244873 permuted_accuracy: 0.10437091439962387 accuracy_gap: 0.8115297928452492 checks: Unseen-date month accuracy reaches 0.90: true Timestamp permutation accuracy is at most 0.20: true Original dates exceed permutation by at least 0.65 accuracy: true outcome: met duration_seconds: 8.037 P045: script: proofs/temporal/weekday_requires_year_context.py page: docs/proofs/temporal/weekday-requires-year-context.py historical: status: Expected limitation source_revision: 810b3fe3b0b25c2c7a718571d83f6006f23110c3 sources: - https://github.com/relflow/relflow/blob/810b3fe/proofs/temporal/datepart_periodicity/test_weekday_requires_year_context.py - https://github.com/relflow/relflow/blob/810b3fe/proofs/temporal/datepart_periodicity/support.py summary: |- The source records accuracy **0.143** with day-of-year alone and **1.000** after explicitly exposing day-of-week. The test passes by confirming both the information boundary and the positive control, for one model seed. insights: |- **The same day number can fall on different weekdays in different years.** Every represented ordinal is balanced across all seven labels, making identical visible inputs require different answers. The recorded day-only accuracy is 0.143, matching the one-seventh boundary. Adding `day_of_week` raises it to 1.000 by exposing the missing distinction directly. Despite the source filename, this control does not test learning weekday arithmetic from a year coordinate. Select DateParts coordinates according to what the target requires. A complete source timestamp does not mean all its information reaches the model. The passing test confirms an expected limitation and a successful added-coordinate control, not universal failure or general calendar reasoning. evidence: |- | Visible coordinates | Required gate | Recorded accuracy | | --- | --- | --- | | Day-of-year | At most 0.20 | 0.143 | | Day-of-year and day-of-week | At least 0.95 | 1.000 | | Improvement from weekday | At least 0.70 | 0.857 | A complete source timestamp does not imply a complete encoded timestamp. [DateParts](../../data-types/dateparts.qmd) retains the coordinates requested by the schema; omitted distinctions need to be supplied explicitly. runs: - started_at: '2026-09-15T02:17:06.356064+00:00' mode: smoke seed: 47 steps_override: 2 accelerator: gpu threads: 1 provenance: revision: null source_sha256: f932ea49261b3d9c86dd0947bad78be64e7b6df8f7b90bfd98394434176f5954 script_sha256: 4b3a05db80bcfd8f2912738c433ed97bc9210457a3de47c4d6458ef4b422f608 python: 3.12.6 platform: Linux-6.8.0-138-generic-x86_64-with-glibc2.35 host: lab packages: relflow: 0.2.1 torch: 2.12.0 lightning: 2.6.4 numpy: 2.4.6 cuda: '13.0' gpu: NVIDIA GeForce RTX 3090 code_sha256: df3bf9be7aebd8408418a9dfc723090bac4ec1356f74c45f232a56d39f4cde1f metrics: day_only_accuracy: 0.1428549438714981 with_weekday_accuracy: 0.27595964074134827 accuracy_gap: 0.13310469686985016 checks: Day-of-year accuracy is at most 0.20: true Visible weekday accuracy reaches 0.95: false Visible weekday improves accuracy by at least 0.70: false outcome: not_met duration_seconds: 0.827 - started_at: '2026-09-15T02:33:44.589034+00:00' mode: full seed: 47 steps_override: null accelerator: gpu threads: 1 provenance: revision: null source_sha256: 902065690b27c7df4f8149bc457b5b5fcaa2380ca4bf9806c00aa25c3f443d66 script_sha256: 4b3a05db80bcfd8f2912738c433ed97bc9210457a3de47c4d6458ef4b422f608 python: 3.12.6 platform: Linux-6.8.0-138-generic-x86_64-with-glibc2.35 host: lab relflow_path: /home/grantham/relflow-proof-scripts.RB4VJH/src/relflow packages: relflow: 0.2.1 torch: 2.12.0 lightning: 2.6.4 numpy: 2.4.6 pyarrow: 24.0.0 polars: 1.44.1 torchmetrics: 1.9.0 tensordict: 0.12.4 pydantic: 2.13.4 cuda: '13.0' gpu: NVIDIA GeForce RTX 3090 code_sha256: df3bf9be7aebd8408418a9dfc723090bac4ec1356f74c45f232a56d39f4cde1f metrics: day_only_accuracy: 0.1428549438714981 with_weekday_accuracy: 1.0 accuracy_gap: 0.8571450561285019 checks: Day-of-year accuracy is at most 0.20: true Visible weekday accuracy reaches 0.95: true Visible weekday improves accuracy by at least 0.70: true outcome: met duration_seconds: 3.722 P046: script: proofs/mutations/neutral_edits_preserve_learning.py page: docs/proofs/mutations/neutral-edits-preserve-learning.py historical: status: Provisional pass source_revision: null sources: [] summary: New experiment; no historical run predates registration. insights: 'Initial hypothesis: Neutral edits preserve a learned numerical and categorical relationship.' evidence: No historical measurements. See recorded runs for actual outcomes. runs: - started_at: '2026-09-15T17:53:00.092179+00:00' mode: smoke seed: 4601 steps_override: 2 accelerator: cpu threads: 1 provenance: revision: 5d985953fd0416bc5982a3518d8f43cadb52110c source_sha256: 82f4b5e5a25603c42ab071ce91f000d3515c904c046a425b2a229bc5f562ae61 script_sha256: e006fbd5d1e18fe5c35bc5d8cd0ab3867855985bbc54aeae57fe8d84c4123e9d code_sha256: f3a2f9c73440410deac5725a9f49e5ec28b78e148071179f73fc9681a0c21f69 python: 3.12.10 platform: macOS-26.5.2-arm64-arm-64bit host: Granthams-MacBook-Pro.local relflow_path: /Users/grantham/Desktop/relflow/src/relflow packages: relflow: 0.2.2 torch: 2.12.0 lightning: 2.6.4 numpy: 2.4.6 pyarrow: 24.0.0 polars: 1.44.1 torchmetrics: 1.9.0 tensordict: 0.12.4 pydantic: 2.13.4 cuda: null gpu: null metrics: source: rmse: 0.6351351040185311 baseline_rmse: 0.8103651362027426 nrmse: 0.7837641029261021 source_steps: 2 edits: - edit: cycle 1 metadata max_prediction_drift: 0.0 nrmse: 0.7837641029261021 changed_or_missing_state_entries: [] vocabulary_preserved: true - edit: cycle 1 inactive extension max_prediction_drift: 0.0 nrmse: 0.7837641029261021 changed_or_missing_state_entries: [] vocabulary_preserved: true - edit: cycle 1 inactive deletion max_prediction_drift: 0.0 nrmse: 0.7837641029261021 changed_or_missing_state_entries: [] vocabulary_preserved: true - edit: cycle 2 metadata max_prediction_drift: 0.0 nrmse: 0.7837641029261021 changed_or_missing_state_entries: [] vocabulary_preserved: true - edit: cycle 2 inactive extension max_prediction_drift: 0.0 nrmse: 0.7837641029261021 changed_or_missing_state_entries: [] vocabulary_preserved: true - edit: cycle 2 inactive deletion max_prediction_drift: 0.0 nrmse: 0.7837641029261021 changed_or_missing_state_entries: [] vocabulary_preserved: true - edit: cycle 3 metadata max_prediction_drift: 0.0 nrmse: 0.7837641029261021 changed_or_missing_state_entries: [] vocabulary_preserved: true - edit: cycle 3 inactive extension max_prediction_drift: 0.0 nrmse: 0.7837641029261021 changed_or_missing_state_entries: [] vocabulary_preserved: true - edit: cycle 3 inactive deletion max_prediction_drift: 0.0 nrmse: 0.7837641029261021 changed_or_missing_state_entries: [] vocabulary_preserved: true - edit: equivalent source rebind max_prediction_drift: 0.0 nrmse: 0.7837641029261021 changed_or_missing_state_entries: [] vocabulary_preserved: true - edit: source rebind restoration max_prediction_drift: 0.0 nrmse: 0.7837641029261021 changed_or_missing_state_entries: [] vocabulary_preserved: true - edit: rejected duplicate extension max_prediction_drift: 0.0 nrmse: 0.7837641029261021 changed_or_missing_state_entries: [] vocabulary_preserved: true controls: complete_reset: rmse: 0.9009977527204428 baseline_rmse: 0.8103651362027426 nrmse: 1.1118417025470664 source_vocabulary: - code-1 - code-2 - code-3 - code-0 test_rows: 1024 optimizer_policy: Fresh AdamW for source training; no post-edit fitting checks: Source learns the relationship below 0.25 nRMSE: false 'cycle 1 metadata: predictions survive': true 'cycle 1 metadata: all original state and vocabulary survive': true 'cycle 1 inactive extension: predictions survive': true 'cycle 1 inactive extension: all original state and vocabulary survive': true 'cycle 1 inactive deletion: predictions survive': true 'cycle 1 inactive deletion: all original state and vocabulary survive': true 'cycle 2 metadata: predictions survive': true 'cycle 2 metadata: all original state and vocabulary survive': true 'cycle 2 inactive extension: predictions survive': true 'cycle 2 inactive extension: all original state and vocabulary survive': true 'cycle 2 inactive deletion: predictions survive': true 'cycle 2 inactive deletion: all original state and vocabulary survive': true 'cycle 3 metadata: predictions survive': true 'cycle 3 metadata: all original state and vocabulary survive': true 'cycle 3 inactive extension: predictions survive': true 'cycle 3 inactive extension: all original state and vocabulary survive': true 'cycle 3 inactive deletion: predictions survive': true 'cycle 3 inactive deletion: all original state and vocabulary survive': true 'equivalent source rebind: predictions survive': true 'equivalent source rebind: all original state and vocabulary survive': true 'source rebind restoration: predictions survive': true 'source rebind restoration: all original state and vocabulary survive': true Duplicate-name edit is rejected: true 'rejected duplicate extension: predictions survive': true 'rejected duplicate extension: all original state and vocabulary survive': true Edited checkpoint preserves schema: true Edited checkpoint preserves learned state: true Edited checkpoint preserves predictions: true Complete reset loses the learned relationship: false outcome: not_met duration_seconds: 0.575 - started_at: '2026-09-15T17:53:35.853115+00:00' mode: full seed: 4601 steps_override: null accelerator: cpu threads: 1 provenance: revision: 5d985953fd0416bc5982a3518d8f43cadb52110c source_sha256: 82f4b5e5a25603c42ab071ce91f000d3515c904c046a425b2a229bc5f562ae61 script_sha256: e006fbd5d1e18fe5c35bc5d8cd0ab3867855985bbc54aeae57fe8d84c4123e9d code_sha256: f3a2f9c73440410deac5725a9f49e5ec28b78e148071179f73fc9681a0c21f69 python: 3.12.10 platform: macOS-26.5.2-arm64-arm-64bit host: Granthams-MacBook-Pro.local relflow_path: /Users/grantham/Desktop/relflow/src/relflow packages: relflow: 0.2.2 torch: 2.12.0 lightning: 2.6.4 numpy: 2.4.6 pyarrow: 24.0.0 polars: 1.44.1 torchmetrics: 1.9.0 tensordict: 0.12.4 pydantic: 2.13.4 cuda: null gpu: null metrics: source: rmse: 0.020029435105787627 baseline_rmse: 0.8103651362027426 nrmse: 0.02471655579809708 source_steps: 512 edits: - edit: cycle 1 metadata max_prediction_drift: 0.0 nrmse: 0.02471655579809708 changed_or_missing_state_entries: [] vocabulary_preserved: true - edit: cycle 1 inactive extension max_prediction_drift: 0.0 nrmse: 0.02471655579809708 changed_or_missing_state_entries: [] vocabulary_preserved: true - edit: cycle 1 inactive deletion max_prediction_drift: 0.0 nrmse: 0.02471655579809708 changed_or_missing_state_entries: [] vocabulary_preserved: true - edit: cycle 2 metadata max_prediction_drift: 0.0 nrmse: 0.02471655579809708 changed_or_missing_state_entries: [] vocabulary_preserved: true - edit: cycle 2 inactive extension max_prediction_drift: 0.0 nrmse: 0.02471655579809708 changed_or_missing_state_entries: [] vocabulary_preserved: true - edit: cycle 2 inactive deletion max_prediction_drift: 0.0 nrmse: 0.02471655579809708 changed_or_missing_state_entries: [] vocabulary_preserved: true - edit: cycle 3 metadata max_prediction_drift: 0.0 nrmse: 0.02471655579809708 changed_or_missing_state_entries: [] vocabulary_preserved: true - edit: cycle 3 inactive extension max_prediction_drift: 0.0 nrmse: 0.02471655579809708 changed_or_missing_state_entries: [] vocabulary_preserved: true - edit: cycle 3 inactive deletion max_prediction_drift: 0.0 nrmse: 0.02471655579809708 changed_or_missing_state_entries: [] vocabulary_preserved: true - edit: equivalent source rebind max_prediction_drift: 0.0 nrmse: 0.02471655579809708 changed_or_missing_state_entries: [] vocabulary_preserved: true - edit: source rebind restoration max_prediction_drift: 0.0 nrmse: 0.02471655579809708 changed_or_missing_state_entries: [] vocabulary_preserved: true - edit: rejected duplicate extension max_prediction_drift: 0.0 nrmse: 0.02471655579809708 changed_or_missing_state_entries: [] vocabulary_preserved: true controls: complete_reset: rmse: 0.9009977527204428 baseline_rmse: 0.8103651362027426 nrmse: 1.1118417025470664 source_vocabulary: - code-1 - code-2 - code-3 - code-0 test_rows: 1024 optimizer_policy: Fresh AdamW for source training; no post-edit fitting checks: Source learns the relationship below 0.25 nRMSE: true 'cycle 1 metadata: predictions survive': true 'cycle 1 metadata: all original state and vocabulary survive': true 'cycle 1 inactive extension: predictions survive': true 'cycle 1 inactive extension: all original state and vocabulary survive': true 'cycle 1 inactive deletion: predictions survive': true 'cycle 1 inactive deletion: all original state and vocabulary survive': true 'cycle 2 metadata: predictions survive': true 'cycle 2 metadata: all original state and vocabulary survive': true 'cycle 2 inactive extension: predictions survive': true 'cycle 2 inactive extension: all original state and vocabulary survive': true 'cycle 2 inactive deletion: predictions survive': true 'cycle 2 inactive deletion: all original state and vocabulary survive': true 'cycle 3 metadata: predictions survive': true 'cycle 3 metadata: all original state and vocabulary survive': true 'cycle 3 inactive extension: predictions survive': true 'cycle 3 inactive extension: all original state and vocabulary survive': true 'cycle 3 inactive deletion: predictions survive': true 'cycle 3 inactive deletion: all original state and vocabulary survive': true 'equivalent source rebind: predictions survive': true 'equivalent source rebind: all original state and vocabulary survive': true 'source rebind restoration: predictions survive': true 'source rebind restoration: all original state and vocabulary survive': true Duplicate-name edit is rejected: true 'rejected duplicate extension: predictions survive': true 'rejected duplicate extension: all original state and vocabulary survive': true Edited checkpoint preserves schema: true Edited checkpoint preserves learned state: true Edited checkpoint preserves predictions: true Complete reset loses the learned relationship: true outcome: met duration_seconds: 4.439 - started_at: '2026-09-15T17:55:28.371185+00:00' mode: full seed: 4601 steps_override: null accelerator: cpu threads: 1 provenance: revision: 5d985953fd0416bc5982a3518d8f43cadb52110c source_sha256: b3208281c58d1be68c68809b5446640c4c4f4b7b3667d374ef8756cfa7e57ca3 script_sha256: 0c049cfe501c455e9ac72a81d889e994bd73c86179e17629285e989c51373285 code_sha256: 8769fd40d8bec0dbf288d51fab6c613b6b9bf8f5a573d2fc8c74f92b56c22988 python: 3.12.10 platform: macOS-26.5.2-arm64-arm-64bit host: Granthams-MacBook-Pro.local relflow_path: /Users/grantham/Desktop/relflow/src/relflow packages: relflow: 0.2.2 torch: 2.12.0 lightning: 2.6.4 numpy: 2.4.6 pyarrow: 24.0.0 polars: 1.44.1 torchmetrics: 1.9.0 tensordict: 0.12.4 pydantic: 2.13.4 cuda: null gpu: null metrics: source: rmse: 0.020029435105787627 baseline_rmse: 0.8103651362027426 nrmse: 0.02471655579809708 source_prerequisite_met: true downstream_interpretable: true source_steps: 512 edits: - edit: cycle 1 metadata max_prediction_drift: 0.0 nrmse: 0.02471655579809708 changed_or_missing_state_entries: [] vocabulary_preserved: true - edit: cycle 1 inactive extension max_prediction_drift: 0.0 nrmse: 0.02471655579809708 changed_or_missing_state_entries: [] vocabulary_preserved: true - edit: cycle 1 inactive deletion max_prediction_drift: 0.0 nrmse: 0.02471655579809708 changed_or_missing_state_entries: [] vocabulary_preserved: true - edit: cycle 2 metadata max_prediction_drift: 0.0 nrmse: 0.02471655579809708 changed_or_missing_state_entries: [] vocabulary_preserved: true - edit: cycle 2 inactive extension max_prediction_drift: 0.0 nrmse: 0.02471655579809708 changed_or_missing_state_entries: [] vocabulary_preserved: true - edit: cycle 2 inactive deletion max_prediction_drift: 0.0 nrmse: 0.02471655579809708 changed_or_missing_state_entries: [] vocabulary_preserved: true - edit: cycle 3 metadata max_prediction_drift: 0.0 nrmse: 0.02471655579809708 changed_or_missing_state_entries: [] vocabulary_preserved: true - edit: cycle 3 inactive extension max_prediction_drift: 0.0 nrmse: 0.02471655579809708 changed_or_missing_state_entries: [] vocabulary_preserved: true - edit: cycle 3 inactive deletion max_prediction_drift: 0.0 nrmse: 0.02471655579809708 changed_or_missing_state_entries: [] vocabulary_preserved: true - edit: equivalent source rebind max_prediction_drift: 0.0 nrmse: 0.02471655579809708 changed_or_missing_state_entries: [] vocabulary_preserved: true - edit: source rebind restoration max_prediction_drift: 0.0 nrmse: 0.02471655579809708 changed_or_missing_state_entries: [] vocabulary_preserved: true - edit: rejected duplicate extension max_prediction_drift: 0.0 nrmse: 0.02471655579809708 changed_or_missing_state_entries: [] vocabulary_preserved: true controls: complete_reset: rmse: 0.9009977527204428 baseline_rmse: 0.8103651362027426 nrmse: 1.1118417025470664 source_vocabulary: - code-1 - code-2 - code-3 - code-0 test_rows: 1024 optimizer_policy: Fresh AdamW for source training; no post-edit fitting checks: Source learns the relationship below 0.25 nRMSE: true 'cycle 1 metadata: predictions survive': true 'cycle 1 metadata: all original state and vocabulary survive': true 'cycle 1 inactive extension: predictions survive': true 'cycle 1 inactive extension: all original state and vocabulary survive': true 'cycle 1 inactive deletion: predictions survive': true 'cycle 1 inactive deletion: all original state and vocabulary survive': true 'cycle 2 metadata: predictions survive': true 'cycle 2 metadata: all original state and vocabulary survive': true 'cycle 2 inactive extension: predictions survive': true 'cycle 2 inactive extension: all original state and vocabulary survive': true 'cycle 2 inactive deletion: predictions survive': true 'cycle 2 inactive deletion: all original state and vocabulary survive': true 'cycle 3 metadata: predictions survive': true 'cycle 3 metadata: all original state and vocabulary survive': true 'cycle 3 inactive extension: predictions survive': true 'cycle 3 inactive extension: all original state and vocabulary survive': true 'cycle 3 inactive deletion: predictions survive': true 'cycle 3 inactive deletion: all original state and vocabulary survive': true 'equivalent source rebind: predictions survive': true 'equivalent source rebind: all original state and vocabulary survive': true 'source rebind restoration: predictions survive': true 'source rebind restoration: all original state and vocabulary survive': true Duplicate-name edit is rejected: true 'rejected duplicate extension: predictions survive': true 'rejected duplicate extension: all original state and vocabulary survive': true Edited checkpoint preserves schema: true Edited checkpoint preserves learned state: true Edited checkpoint preserves predictions: true Complete reset loses the learned relationship: true outcome: met duration_seconds: 4.967 - started_at: '2026-09-15T17:58:06.195510+00:00' mode: full seed: 4601 steps_override: null accelerator: gpu threads: 1 provenance: revision: null source_sha256: b3208281c58d1be68c68809b5446640c4c4f4b7b3667d374ef8756cfa7e57ca3 script_sha256: 0c049cfe501c455e9ac72a81d889e994bd73c86179e17629285e989c51373285 code_sha256: 8769fd40d8bec0dbf288d51fab6c613b6b9bf8f5a573d2fc8c74f92b56c22988 python: 3.12.6 platform: Linux-6.8.0-138-generic-x86_64-with-glibc2.35 host: lab relflow_path: /home/grantham/relflow-mutations-v3_tgit_/src/relflow packages: relflow: 0.2.2 torch: 2.12.0 lightning: 2.6.4 numpy: 2.4.6 pyarrow: 24.0.0 polars: 1.44.1 torchmetrics: 1.9.0 tensordict: 0.12.4 pydantic: 2.13.4 cuda: '13.0' gpu: NVIDIA GeForce RTX 3090 metrics: source: rmse: 0.015390001694667008 baseline_rmse: 0.8103651362027426 nrmse: 0.01899144102716758 source_prerequisite_met: true downstream_interpretable: true source_steps: 512 edits: - edit: cycle 1 metadata max_prediction_drift: 0.0 nrmse: 0.01899144102716758 changed_or_missing_state_entries: [] vocabulary_preserved: true - edit: cycle 1 inactive extension max_prediction_drift: 0.0 nrmse: 0.01899144102716758 changed_or_missing_state_entries: [] vocabulary_preserved: true - edit: cycle 1 inactive deletion max_prediction_drift: 0.0 nrmse: 0.01899144102716758 changed_or_missing_state_entries: [] vocabulary_preserved: true - edit: cycle 2 metadata max_prediction_drift: 0.0 nrmse: 0.01899144102716758 changed_or_missing_state_entries: [] vocabulary_preserved: true - edit: cycle 2 inactive extension max_prediction_drift: 0.0 nrmse: 0.01899144102716758 changed_or_missing_state_entries: [] vocabulary_preserved: true - edit: cycle 2 inactive deletion max_prediction_drift: 0.0 nrmse: 0.01899144102716758 changed_or_missing_state_entries: [] vocabulary_preserved: true - edit: cycle 3 metadata max_prediction_drift: 0.0 nrmse: 0.01899144102716758 changed_or_missing_state_entries: [] vocabulary_preserved: true - edit: cycle 3 inactive extension max_prediction_drift: 0.0 nrmse: 0.01899144102716758 changed_or_missing_state_entries: [] vocabulary_preserved: true - edit: cycle 3 inactive deletion max_prediction_drift: 0.0 nrmse: 0.01899144102716758 changed_or_missing_state_entries: [] vocabulary_preserved: true - edit: equivalent source rebind max_prediction_drift: 0.0 nrmse: 0.01899144102716758 changed_or_missing_state_entries: [] vocabulary_preserved: true - edit: source rebind restoration max_prediction_drift: 0.0 nrmse: 0.01899144102716758 changed_or_missing_state_entries: [] vocabulary_preserved: true - edit: rejected duplicate extension max_prediction_drift: 0.0 nrmse: 0.01899144102716758 changed_or_missing_state_entries: [] vocabulary_preserved: true controls: complete_reset: rmse: 0.9009977688127063 baseline_rmse: 0.8103651362027426 nrmse: 1.1118417224051067 source_vocabulary: - code-1 - code-2 - code-3 - code-0 test_rows: 1024 optimizer_policy: Fresh AdamW for source training; no post-edit fitting checks: Source learns the relationship below 0.25 nRMSE: true 'cycle 1 metadata: predictions survive': true 'cycle 1 metadata: all original state and vocabulary survive': true 'cycle 1 inactive extension: predictions survive': true 'cycle 1 inactive extension: all original state and vocabulary survive': true 'cycle 1 inactive deletion: predictions survive': true 'cycle 1 inactive deletion: all original state and vocabulary survive': true 'cycle 2 metadata: predictions survive': true 'cycle 2 metadata: all original state and vocabulary survive': true 'cycle 2 inactive extension: predictions survive': true 'cycle 2 inactive extension: all original state and vocabulary survive': true 'cycle 2 inactive deletion: predictions survive': true 'cycle 2 inactive deletion: all original state and vocabulary survive': true 'cycle 3 metadata: predictions survive': true 'cycle 3 metadata: all original state and vocabulary survive': true 'cycle 3 inactive extension: predictions survive': true 'cycle 3 inactive extension: all original state and vocabulary survive': true 'cycle 3 inactive deletion: predictions survive': true 'cycle 3 inactive deletion: all original state and vocabulary survive': true 'equivalent source rebind: predictions survive': true 'equivalent source rebind: all original state and vocabulary survive': true 'source rebind restoration: predictions survive': true 'source rebind restoration: all original state and vocabulary survive': true Duplicate-name edit is rejected: true 'rejected duplicate extension: predictions survive': true 'rejected duplicate extension: all original state and vocabulary survive': true Edited checkpoint preserves schema: true Edited checkpoint preserves learned state: true Edited checkpoint preserves predictions: true Complete reset loses the learned relationship: true outcome: met duration_seconds: 11.213 - started_at: '2026-09-15T18:00:08.222756+00:00' mode: full seed: 7101 steps_override: null accelerator: gpu threads: 1 provenance: revision: null source_sha256: b3208281c58d1be68c68809b5446640c4c4f4b7b3667d374ef8756cfa7e57ca3 script_sha256: 0c049cfe501c455e9ac72a81d889e994bd73c86179e17629285e989c51373285 code_sha256: 8769fd40d8bec0dbf288d51fab6c613b6b9bf8f5a573d2fc8c74f92b56c22988 python: 3.12.6 platform: Linux-6.8.0-138-generic-x86_64-with-glibc2.35 host: lab relflow_path: /home/grantham/relflow-mutations-v3_tgit_/src/relflow packages: relflow: 0.2.2 torch: 2.12.0 lightning: 2.6.4 numpy: 2.4.6 pyarrow: 24.0.0 polars: 1.44.1 torchmetrics: 1.9.0 tensordict: 0.12.4 pydantic: 2.13.4 cuda: '13.0' gpu: NVIDIA GeForce RTX 3090 metrics: source: rmse: 0.008368666828375167 baseline_rmse: 0.7822472327744479 nrmse: 0.0106982376897563 source_prerequisite_met: true downstream_interpretable: true source_steps: 512 edits: - edit: cycle 1 metadata max_prediction_drift: 0.0 nrmse: 0.0106982376897563 changed_or_missing_state_entries: [] vocabulary_preserved: true - edit: cycle 1 inactive extension max_prediction_drift: 0.0 nrmse: 0.0106982376897563 changed_or_missing_state_entries: [] vocabulary_preserved: true - edit: cycle 1 inactive deletion max_prediction_drift: 0.0 nrmse: 0.0106982376897563 changed_or_missing_state_entries: [] vocabulary_preserved: true - edit: cycle 2 metadata max_prediction_drift: 0.0 nrmse: 0.0106982376897563 changed_or_missing_state_entries: [] vocabulary_preserved: true - edit: cycle 2 inactive extension max_prediction_drift: 0.0 nrmse: 0.0106982376897563 changed_or_missing_state_entries: [] vocabulary_preserved: true - edit: cycle 2 inactive deletion max_prediction_drift: 0.0 nrmse: 0.0106982376897563 changed_or_missing_state_entries: [] vocabulary_preserved: true - edit: cycle 3 metadata max_prediction_drift: 0.0 nrmse: 0.0106982376897563 changed_or_missing_state_entries: [] vocabulary_preserved: true - edit: cycle 3 inactive extension max_prediction_drift: 0.0 nrmse: 0.0106982376897563 changed_or_missing_state_entries: [] vocabulary_preserved: true - edit: cycle 3 inactive deletion max_prediction_drift: 0.0 nrmse: 0.0106982376897563 changed_or_missing_state_entries: [] vocabulary_preserved: true - edit: equivalent source rebind max_prediction_drift: 0.0 nrmse: 0.0106982376897563 changed_or_missing_state_entries: [] vocabulary_preserved: true - edit: source rebind restoration max_prediction_drift: 0.0 nrmse: 0.0106982376897563 changed_or_missing_state_entries: [] vocabulary_preserved: true - edit: rejected duplicate extension max_prediction_drift: 0.0 nrmse: 0.0106982376897563 changed_or_missing_state_entries: [] vocabulary_preserved: true controls: complete_reset: rmse: 0.9555554631912822 baseline_rmse: 0.7822472327744479 nrmse: 1.2215517334617478 source_vocabulary: - code-1 - code-2 - code-3 - code-0 test_rows: 1024 optimizer_policy: Fresh AdamW for source training; no post-edit fitting checks: Source learns the relationship below 0.25 nRMSE: true 'cycle 1 metadata: predictions survive': true 'cycle 1 metadata: all original state and vocabulary survive': true 'cycle 1 inactive extension: predictions survive': true 'cycle 1 inactive extension: all original state and vocabulary survive': true 'cycle 1 inactive deletion: predictions survive': true 'cycle 1 inactive deletion: all original state and vocabulary survive': true 'cycle 2 metadata: predictions survive': true 'cycle 2 metadata: all original state and vocabulary survive': true 'cycle 2 inactive extension: predictions survive': true 'cycle 2 inactive extension: all original state and vocabulary survive': true 'cycle 2 inactive deletion: predictions survive': true 'cycle 2 inactive deletion: all original state and vocabulary survive': true 'cycle 3 metadata: predictions survive': true 'cycle 3 metadata: all original state and vocabulary survive': true 'cycle 3 inactive extension: predictions survive': true 'cycle 3 inactive extension: all original state and vocabulary survive': true 'cycle 3 inactive deletion: predictions survive': true 'cycle 3 inactive deletion: all original state and vocabulary survive': true 'equivalent source rebind: predictions survive': true 'equivalent source rebind: all original state and vocabulary survive': true 'source rebind restoration: predictions survive': true 'source rebind restoration: all original state and vocabulary survive': true Duplicate-name edit is rejected: true 'rejected duplicate extension: predictions survive': true 'rejected duplicate extension: all original state and vocabulary survive': true Edited checkpoint preserves schema: true Edited checkpoint preserves learned state: true Edited checkpoint preserves predictions: true Complete reset loses the learned relationship: true outcome: met duration_seconds: 11.319 - started_at: '2026-09-15T18:02:10.645831+00:00' mode: full seed: 7102 steps_override: null accelerator: gpu threads: 1 provenance: revision: null source_sha256: b3208281c58d1be68c68809b5446640c4c4f4b7b3667d374ef8756cfa7e57ca3 script_sha256: 0c049cfe501c455e9ac72a81d889e994bd73c86179e17629285e989c51373285 code_sha256: 8769fd40d8bec0dbf288d51fab6c613b6b9bf8f5a573d2fc8c74f92b56c22988 python: 3.12.6 platform: Linux-6.8.0-138-generic-x86_64-with-glibc2.35 host: lab relflow_path: /home/grantham/relflow-mutations-v3_tgit_/src/relflow packages: relflow: 0.2.2 torch: 2.12.0 lightning: 2.6.4 numpy: 2.4.6 pyarrow: 24.0.0 polars: 1.44.1 torchmetrics: 1.9.0 tensordict: 0.12.4 pydantic: 2.13.4 cuda: '13.0' gpu: NVIDIA GeForce RTX 3090 metrics: source: rmse: 0.013073767926484693 baseline_rmse: 0.8037941573669484 nrmse: 0.016265069615971656 source_prerequisite_met: true downstream_interpretable: true source_steps: 512 edits: - edit: cycle 1 metadata max_prediction_drift: 0.0 nrmse: 0.016265069615971656 changed_or_missing_state_entries: [] vocabulary_preserved: true - edit: cycle 1 inactive extension max_prediction_drift: 0.0 nrmse: 0.016265069615971656 changed_or_missing_state_entries: [] vocabulary_preserved: true - edit: cycle 1 inactive deletion max_prediction_drift: 0.0 nrmse: 0.016265069615971656 changed_or_missing_state_entries: [] vocabulary_preserved: true - edit: cycle 2 metadata max_prediction_drift: 0.0 nrmse: 0.016265069615971656 changed_or_missing_state_entries: [] vocabulary_preserved: true - edit: cycle 2 inactive extension max_prediction_drift: 0.0 nrmse: 0.016265069615971656 changed_or_missing_state_entries: [] vocabulary_preserved: true - edit: cycle 2 inactive deletion max_prediction_drift: 0.0 nrmse: 0.016265069615971656 changed_or_missing_state_entries: [] vocabulary_preserved: true - edit: cycle 3 metadata max_prediction_drift: 0.0 nrmse: 0.016265069615971656 changed_or_missing_state_entries: [] vocabulary_preserved: true - edit: cycle 3 inactive extension max_prediction_drift: 0.0 nrmse: 0.016265069615971656 changed_or_missing_state_entries: [] vocabulary_preserved: true - edit: cycle 3 inactive deletion max_prediction_drift: 0.0 nrmse: 0.016265069615971656 changed_or_missing_state_entries: [] vocabulary_preserved: true - edit: equivalent source rebind max_prediction_drift: 0.0 nrmse: 0.016265069615971656 changed_or_missing_state_entries: [] vocabulary_preserved: true - edit: source rebind restoration max_prediction_drift: 0.0 nrmse: 0.016265069615971656 changed_or_missing_state_entries: [] vocabulary_preserved: true - edit: rejected duplicate extension max_prediction_drift: 0.0 nrmse: 0.016265069615971656 changed_or_missing_state_entries: [] vocabulary_preserved: true controls: complete_reset: rmse: 0.754748018777057 baseline_rmse: 0.8037941573669484 nrmse: 0.9389817179679986 source_vocabulary: - code-0 - code-1 - code-2 - code-3 test_rows: 1024 optimizer_policy: Fresh AdamW for source training; no post-edit fitting checks: Source learns the relationship below 0.25 nRMSE: true 'cycle 1 metadata: predictions survive': true 'cycle 1 metadata: all original state and vocabulary survive': true 'cycle 1 inactive extension: predictions survive': true 'cycle 1 inactive extension: all original state and vocabulary survive': true 'cycle 1 inactive deletion: predictions survive': true 'cycle 1 inactive deletion: all original state and vocabulary survive': true 'cycle 2 metadata: predictions survive': true 'cycle 2 metadata: all original state and vocabulary survive': true 'cycle 2 inactive extension: predictions survive': true 'cycle 2 inactive extension: all original state and vocabulary survive': true 'cycle 2 inactive deletion: predictions survive': true 'cycle 2 inactive deletion: all original state and vocabulary survive': true 'cycle 3 metadata: predictions survive': true 'cycle 3 metadata: all original state and vocabulary survive': true 'cycle 3 inactive extension: predictions survive': true 'cycle 3 inactive extension: all original state and vocabulary survive': true 'cycle 3 inactive deletion: predictions survive': true 'cycle 3 inactive deletion: all original state and vocabulary survive': true 'equivalent source rebind: predictions survive': true 'equivalent source rebind: all original state and vocabulary survive': true 'source rebind restoration: predictions survive': true 'source rebind restoration: all original state and vocabulary survive': true Duplicate-name edit is rejected: true 'rejected duplicate extension: predictions survive': true 'rejected duplicate extension: all original state and vocabulary survive': true Edited checkpoint preserves schema: true Edited checkpoint preserves learned state: true Edited checkpoint preserves predictions: true Complete reset loses the learned relationship: true outcome: met duration_seconds: 11.269 P047: script: proofs/mutations/extend_prediction_target.py page: docs/proofs/mutations/extend-prediction-target.py historical: status: Provisional pass source_revision: null sources: [] summary: New experiment; no historical run predates registration. insights: 'Initial hypothesis: A new hidden target learns while old targets retain skill with rehearsal.' evidence: No historical measurements. See recorded runs for actual outcomes. runs: - started_at: '2026-09-15T17:53:04.610375+00:00' mode: smoke seed: 4701 steps_override: 2 accelerator: cpu threads: 1 provenance: revision: 5d985953fd0416bc5982a3518d8f43cadb52110c source_sha256: 749fbc77b2bff4cb5eb3842d361be641dfe6894e42a0593c89ee25357a5d65eb script_sha256: b4d50df5f8885ae16d5cefacb11986dfa357b7fed7200efc42f4825b1b1a7ed9 code_sha256: 8c67ea6056f22b2dad82d1c0d0fba89bc36149117988a447d1b5340440a41a9c python: 3.12.10 platform: macOS-26.5.2-arm64-arm-64bit host: Granthams-MacBook-Pro.local relflow_path: /Users/grantham/Desktop/relflow/src/relflow packages: relflow: 0.2.2 torch: 2.12.0 lightning: 2.6.4 numpy: 2.4.6 pyarrow: 24.0.0 polars: 1.44.1 torchmetrics: 1.9.0 tensordict: 0.12.4 pydantic: 2.13.4 cuda: null gpu: null metrics: source: u: rmse: 0.7184877635587731 baseline_rmse: 1.2663362526593396 nrmse: 0.5673751833685009 v: rmse: 0.9520061498610833 baseline_rmse: 1.2961683303676375 nrmse: 0.7344772492559373 source_fit: steps: 2 optimizer_covers_current_parameters: true validation_curve: - step: 0 scores: u: rmse: 1.3509519179467873 baseline_rmse: 1.295883063514218 nrmse: 1.0424952343178495 v: rmse: 1.7025466041136943 baseline_rmse: 1.2898120364298942 nrmse: 1.3199959032993827 - step: 2 scores: u: rmse: 0.7289427421825075 baseline_rmse: 1.295883063514218 nrmse: 0.5625065738614847 v: rmse: 0.9725274664061965 baseline_rmse: 1.2898120364298942 nrmse: 0.7540071257964701 source_prerequisite_met: false downstream_interpretable: false arms: extended: before: u: rmse: 0.7196009950687224 baseline_rmse: 1.2663362526593396 nrmse: 0.56825427966509 v: rmse: 0.952411532418295 baseline_rmse: 1.2961683303676375 nrmse: 0.7347900038169878 w: rmse: 1.6819942223834303 baseline_rmse: 1.8289556685792812 nrmse: 0.9196473437160949 after: u: rmse: 0.3613070045248406 baseline_rmse: 1.2663362526593396 nrmse: 0.2853167977826477 v: rmse: 0.3062141048508276 baseline_rmse: 1.2961683303676375 nrmse: 0.2362456308155399 w: rmse: 0.9172763753206038 baseline_rmse: 1.8289556685792812 nrmse: 0.5015301305980462 initial_old_prediction_drift: u: 0.0033138394355773926 v: 0.006249785423278809 steps: 2 optimizer_covers_current_parameters: true validation_curve: - step: 0 scores: u: rmse: 0.7446389398139777 baseline_rmse: 1.3104433241303899 nrmse: 0.5682343723702206 v: rmse: 0.939437151882393 baseline_rmse: 1.2705869192092003 nrmse: 0.7393725983477688 w: rmse: 1.6629773847810996 baseline_rmse: 1.8209322963213432 nrmse: 0.9132560217316454 - step: 2 scores: u: rmse: 0.36930718684363484 baseline_rmse: 1.3104433241303899 nrmse: 0.28181851137187264 v: rmse: 0.3046839697242332 baseline_rmse: 1.2705869192092003 nrmse: 0.23979781714883797 w: rmse: 0.9176497254254723 baseline_rmse: 1.8209322963213432 nrmse: 0.5039449996462323 changed_existing_state_entries: [] new_head_updated_parameters: - embedder.embeddings.weight - embedder.linear.weight - embedder.scalar.weight - decoder.context_projection.0.weight - decoder.context_projection.0.bias - decoder.context_projection.1.weight - decoder.context_projection.1.bias - decoder.context_projection.3.weight - decoder.context_projection.3.bias - decoder.pool.queries - decoder.pool.blocks.0.ffn_norm.weight - decoder.pool.blocks.0.ffn_norm.bias - decoder.pool.blocks.0.attention.q_proj.weight - decoder.pool.blocks.0.attention.k_proj.weight - decoder.pool.blocks.0.attention.v_proj.weight - decoder.pool.blocks.0.attention.v_proj.bias - decoder.pool.blocks.0.attention.out_proj.weight - decoder.pool.blocks.0.attention.out_proj.bias - decoder.pool.blocks.0.ffn.0.weight - decoder.pool.blocks.0.ffn.0.bias - decoder.pool.blocks.0.ffn.3.weight - decoder.pool.blocks.0.ffn.3.bias - decoder.pool.norm.weight - decoder.pool.norm.bias - decoder.classification.weight - decoder.classification.bias - decoder.regression.weight - decoder.regression.bias continuation: before: u: rmse: 0.7184877635587731 baseline_rmse: 1.2663362526593396 nrmse: 0.5673751833685009 v: rmse: 0.9520061498610833 baseline_rmse: 1.2961683303676375 nrmse: 0.7344772492559373 after: u: rmse: 0.28971781770269434 baseline_rmse: 1.2663362526593396 nrmse: 0.22878427202433735 v: rmse: 0.3071165254487839 baseline_rmse: 1.2961683303676375 nrmse: 0.2369418525768757 initial_old_prediction_drift: u: 0.0 v: 0.0 steps: 2 optimizer_covers_current_parameters: true validation_curve: - step: 0 scores: u: rmse: 0.7435094022931849 baseline_rmse: 1.3104433241303899 nrmse: 0.5673724216852931 v: rmse: 0.9390350848147172 baseline_rmse: 1.2705869192092003 nrmse: 0.739056156346362 - step: 2 scores: u: rmse: 0.2995876155239647 baseline_rmse: 1.3104433241303899 nrmse: 0.22861546928996032 v: rmse: 0.30069907949272984 baseline_rmse: 1.2705869192092003 nrmse: 0.23666155769955646 scratch_adaptation: before: u: rmse: 1.2899407966264593 baseline_rmse: 1.2663362526593396 nrmse: 1.0186400286002628 v: rmse: 1.10920508178733 baseline_rmse: 1.2961683303676375 nrmse: 0.8557569690602777 w: rmse: 1.7488962582222345 baseline_rmse: 1.8289556685792812 nrmse: 0.9562267080977221 after: u: rmse: 0.5574314651650658 baseline_rmse: 1.2663362526593396 nrmse: 0.44019229805231036 v: rmse: 0.5670848387680588 baseline_rmse: 1.2961683303676375 nrmse: 0.4375086364031239 w: rmse: 1.2267252746736357 baseline_rmse: 1.8289556685792812 nrmse: 0.670724444418353 initial_old_prediction_drift: u: 1.4069195687770844 v: 1.0155301690101624 steps: 2 optimizer_covers_current_parameters: true validation_curve: - step: 0 scores: u: rmse: 1.341623916158934 baseline_rmse: 1.3104433241303899 nrmse: 1.0237939264174096 v: rmse: 1.0918173893993468 baseline_rmse: 1.2705869192092003 nrmse: 0.8593016132094939 w: rmse: 1.7316059452084105 baseline_rmse: 1.8209322963213432 nrmse: 0.9509447159054786 - step: 2 scores: u: rmse: 0.5915161306289228 baseline_rmse: 1.3104433241303899 nrmse: 0.45138627496267564 v: rmse: 0.5553606677147325 baseline_rmse: 1.2705869192092003 nrmse: 0.43708986714610837 w: rmse: 1.2266355350396705 baseline_rmse: 1.8209322963213432 nrmse: 0.6736305009899192 scratch_total: before: u: rmse: 1.5094414653137729 baseline_rmse: 1.2663362526593396 nrmse: 1.1919752452351466 v: rmse: 1.5644460083090737 baseline_rmse: 1.2961683303676375 nrmse: 1.2069774979499333 w: rmse: 1.9603643402199182 baseline_rmse: 1.8289556685792812 nrmse: 1.0718490195788695 after: u: rmse: 0.8671169965970984 baseline_rmse: 1.2663362526593396 nrmse: 0.6847446677579748 v: rmse: 0.967181049149102 baseline_rmse: 1.2961683303676375 nrmse: 0.7461847558602026 w: rmse: 1.2764903003450534 baseline_rmse: 1.8289556685792812 nrmse: 0.6979339752595651 initial_old_prediction_drift: u: 1.7696527242660522 v: 1.7364290952682495 steps: 2 optimizer_covers_current_parameters: true validation_curve: - step: 0 scores: u: rmse: 1.5127669768475775 baseline_rmse: 1.3104433241303899 nrmse: 1.1543932873643732 v: rmse: 1.5470830375162308 baseline_rmse: 1.2705869192092003 nrmse: 1.2176129111097087 w: rmse: 1.9448778758458212 baseline_rmse: 1.8209322963213432 nrmse: 1.0680670993506312 - step: 2 scores: u: rmse: 0.902403166266992 baseline_rmse: 1.3104433241303899 nrmse: 0.6886243377719725 v: rmse: 0.9551300270125085 baseline_rmse: 1.2705869192092003 nrmse: 0.7517234850859091 w: rmse: 1.2728006365564104 baseline_rmse: 1.8209322963213432 nrmse: 0.6989829545709793 test_rows: 1024 baseline: Source-training mean for test comparisons; phase-training mean for validation curves optimizer_policy: New AdamW factory and Trainer for every fit; all available labels rehearsed mutation: extend record with hidden Number record/w checks: Source learns both original tasks below 0.25 nRMSE: false Extension preserves every existing state entry: true Extension immediately preserves old predictions: false 'extended: fresh optimizer covers current parameters': true 'extended: all final targets below 0.25 nRMSE': false Added head parameters actually learn: true Hidden target placeholders cannot affect predictions: true Extended checkpoint preserves schema and learned state: true Extended checkpoint preserves predictions: true 'continuation: fresh optimizer covers current parameters': true 'scratch_adaptation: fresh optimizer covers current parameters': true 'scratch_total: fresh optimizer covers current parameters': true 'scratch_total: all final targets below 0.25 nRMSE': false outcome: not_met duration_seconds: 0.876 - started_at: '2026-09-15T17:53:43.745091+00:00' mode: full seed: 4701 steps_override: null accelerator: cpu threads: 1 provenance: revision: 5d985953fd0416bc5982a3518d8f43cadb52110c source_sha256: 749fbc77b2bff4cb5eb3842d361be641dfe6894e42a0593c89ee25357a5d65eb script_sha256: b4d50df5f8885ae16d5cefacb11986dfa357b7fed7200efc42f4825b1b1a7ed9 code_sha256: 8c67ea6056f22b2dad82d1c0d0fba89bc36149117988a447d1b5340440a41a9c python: 3.12.10 platform: macOS-26.5.2-arm64-arm-64bit host: Granthams-MacBook-Pro.local relflow_path: /Users/grantham/Desktop/relflow/src/relflow packages: relflow: 0.2.2 torch: 2.12.0 lightning: 2.6.4 numpy: 2.4.6 pyarrow: 24.0.0 polars: 1.44.1 torchmetrics: 1.9.0 tensordict: 0.12.4 pydantic: 2.13.4 cuda: null gpu: null metrics: source: u: rmse: 0.04339685848153591 baseline_rmse: 1.2663362526593396 nrmse: 0.034269617086616105 v: rmse: 0.03464917023016499 baseline_rmse: 1.2961683303676375 nrmse: 0.02673199878316523 source_fit: steps: 512 optimizer_covers_current_parameters: true validation_curve: - step: 0 scores: u: rmse: 1.3509519179467873 baseline_rmse: 1.295883063514218 nrmse: 1.0424952343178495 v: rmse: 1.7025466041136943 baseline_rmse: 1.2898120364298942 nrmse: 1.3199959032993827 - step: 32 scores: u: rmse: 0.08500438323886007 baseline_rmse: 1.295883063514218 nrmse: 0.06559572050300774 v: rmse: 0.11706025505541162 baseline_rmse: 1.2898120364298942 nrmse: 0.09075760789101169 - step: 128 scores: u: rmse: 0.02709548454351126 baseline_rmse: 1.295883063514218 nrmse: 0.02090889626262484 v: rmse: 0.04247795145991918 baseline_rmse: 1.2898120364298942 nrmse: 0.032933443215102144 - step: 512 scores: u: rmse: 0.046377564578228725 baseline_rmse: 1.295883063514218 nrmse: 0.03578838699570664 v: rmse: 0.03358961215940764 baseline_rmse: 1.2898120364298942 nrmse: 0.026042253608038298 source_prerequisite_met: true downstream_interpretable: true arms: extended: before: u: rmse: 0.04338463782724802 baseline_rmse: 1.2663362526593396 nrmse: 0.034259966684314004 v: rmse: 0.04017069929827001 baseline_rmse: 1.2961683303676375 nrmse: 0.030991884585604892 w: rmse: 1.73040449423147 baseline_rmse: 1.8289556685792812 nrmse: 0.9461161492096936 after: u: rmse: 0.020053658093265946 baseline_rmse: 1.2663362526593396 nrmse: 0.015835966198671746 v: rmse: 0.022480374165581508 baseline_rmse: 1.2961683303676375 nrmse: 0.01734371503985544 w: rmse: 0.047069585087884114 baseline_rmse: 1.8289556685792812 nrmse: 0.025735771455000552 initial_old_prediction_drift: u: 0.015998125076293945 v: 0.020020246505737305 steps: 256 optimizer_covers_current_parameters: true validation_curve: - step: 0 scores: u: rmse: 0.043597498705624405 baseline_rmse: 1.3104433241303899 nrmse: 0.03326927452933205 v: rmse: 0.03898050824242802 baseline_rmse: 1.2705869192092003 nrmse: 0.030679135486999246 w: rmse: 1.7157907679756903 baseline_rmse: 1.8209322963213432 nrmse: 0.9422595070898241 - step: 32 scores: u: rmse: 0.05030544976923204 baseline_rmse: 1.3104433241303899 nrmse: 0.03838811556586374 v: rmse: 0.04795291787662376 baseline_rmse: 1.2705869192092003 nrmse: 0.037740761495064924 w: rmse: 0.10949826635992924 baseline_rmse: 1.8209322963213432 nrmse: 0.060133079401764794 - step: 128 scores: u: rmse: 0.007891395414127643 baseline_rmse: 1.3104433241303899 nrmse: 0.006021928051992918 v: rmse: 0.012467297531026718 baseline_rmse: 1.2705869192092003 nrmse: 0.00981223507226584 w: rmse: 0.018547623205124056 baseline_rmse: 1.8209322963213432 nrmse: 0.010185784085764232 - step: 256 scores: u: rmse: 0.020140006992766024 baseline_rmse: 1.3104433241303899 nrmse: 0.015368850084478802 v: rmse: 0.02328180518929462 baseline_rmse: 1.2705869192092003 nrmse: 0.01832366195284378 w: rmse: 0.046479505370370475 baseline_rmse: 1.8209322963213432 nrmse: 0.025525114505502818 changed_existing_state_entries: [] new_head_updated_parameters: - embedder.embeddings.weight - embedder.linear.weight - embedder.scalar.weight - decoder.context_projection.0.weight - decoder.context_projection.0.bias - decoder.context_projection.1.weight - decoder.context_projection.1.bias - decoder.context_projection.3.weight - decoder.context_projection.3.bias - decoder.pool.queries - decoder.pool.blocks.0.ffn_norm.weight - decoder.pool.blocks.0.ffn_norm.bias - decoder.pool.blocks.0.attention.q_proj.weight - decoder.pool.blocks.0.attention.k_proj.weight - decoder.pool.blocks.0.attention.v_proj.weight - decoder.pool.blocks.0.attention.v_proj.bias - decoder.pool.blocks.0.attention.out_proj.weight - decoder.pool.blocks.0.attention.out_proj.bias - decoder.pool.blocks.0.ffn.0.weight - decoder.pool.blocks.0.ffn.0.bias - decoder.pool.blocks.0.ffn.3.weight - decoder.pool.blocks.0.ffn.3.bias - decoder.pool.norm.weight - decoder.pool.norm.bias - decoder.classification.weight - decoder.classification.bias - decoder.regression.weight - decoder.regression.bias continuation: before: u: rmse: 0.04339685848153591 baseline_rmse: 1.2663362526593396 nrmse: 0.034269617086616105 v: rmse: 0.03464917023016499 baseline_rmse: 1.2961683303676375 nrmse: 0.02673199878316523 after: u: rmse: 0.024847849450170354 baseline_rmse: 1.2663362526593396 nrmse: 0.01962184166961122 v: rmse: 0.03539640971424924 baseline_rmse: 1.2961683303676375 nrmse: 0.027308497580873323 initial_old_prediction_drift: u: 0.0 v: 0.0 steps: 256 optimizer_covers_current_parameters: true validation_curve: - step: 0 scores: u: rmse: 0.043578616866279236 baseline_rmse: 1.3104433241303899 nrmse: 0.03325486578765091 v: rmse: 0.033650245498206555 baseline_rmse: 1.2705869192092003 nrmse: 0.02648401694482272 - step: 32 scores: u: rmse: 0.03494173974815033 baseline_rmse: 1.3104433241303899 nrmse: 0.02666406024948669 v: rmse: 0.08441034946932946 baseline_rmse: 1.2705869192092003 nrmse: 0.06643414015458743 - step: 128 scores: u: rmse: 0.02933732798141578 baseline_rmse: 1.3104433241303899 nrmse: 0.022387330639334613 v: rmse: 0.012658927846920236 baseline_rmse: 1.2705869192092003 nrmse: 0.009963055384513968 - step: 256 scores: u: rmse: 0.02554737302907079 baseline_rmse: 1.3104433241303899 nrmse: 0.019495213992580737 v: rmse: 0.03563651009027191 baseline_rmse: 1.2705869192092003 nrmse: 0.028047282363375576 scratch_adaptation: before: u: rmse: 1.2899407966264593 baseline_rmse: 1.2663362526593396 nrmse: 1.0186400286002628 v: rmse: 1.10920508178733 baseline_rmse: 1.2961683303676375 nrmse: 0.8557569690602777 w: rmse: 1.7488962582222345 baseline_rmse: 1.8289556685792812 nrmse: 0.9562267080977221 after: u: rmse: 0.046204841875476875 baseline_rmse: 1.2663362526593396 nrmse: 0.03648702449957149 v: rmse: 0.05890874541622356 baseline_rmse: 1.2961683303676375 nrmse: 0.0454483758290214 w: rmse: 0.03371907689280475 baseline_rmse: 1.8289556685792812 nrmse: 0.01843624614422583 initial_old_prediction_drift: u: 3.1339073181152344 v: 2.9099842980504036 steps: 256 optimizer_covers_current_parameters: true validation_curve: - step: 0 scores: u: rmse: 1.341623916158934 baseline_rmse: 1.3104433241303899 nrmse: 1.0237939264174096 v: rmse: 1.0918173893993468 baseline_rmse: 1.2705869192092003 nrmse: 0.8593016132094939 w: rmse: 1.7316059452084105 baseline_rmse: 1.8209322963213432 nrmse: 0.9509447159054786 - step: 32 scores: u: rmse: 0.0754211523406791 baseline_rmse: 1.3104433241303899 nrmse: 0.05755392160185834 v: rmse: 0.08702010311291794 baseline_rmse: 1.2705869192092003 nrmse: 0.06848811505715667 w: rmse: 0.17318271728552895 baseline_rmse: 1.8209322963213432 nrmse: 0.09510662073235428 - step: 128 scores: u: rmse: 0.02006432085463077 baseline_rmse: 1.3104433241303899 nrmse: 0.015311093952075685 v: rmse: 0.07701617202253476 baseline_rmse: 1.2705869192092003 nrmse: 0.060614642617656415 w: rmse: 0.06543257210337809 baseline_rmse: 1.8209322963213432 nrmse: 0.035933555704166106 - step: 256 scores: u: rmse: 0.04627067116592861 baseline_rmse: 1.3104433241303899 nrmse: 0.03530917386040623 v: rmse: 0.05786074070198421 baseline_rmse: 1.2705869192092003 nrmse: 0.04553859309207757 w: rmse: 0.034026255949966284 baseline_rmse: 1.8209322963213432 nrmse: 0.018686173021757208 scratch_total: before: u: rmse: 1.5094414653137729 baseline_rmse: 1.2663362526593396 nrmse: 1.1919752452351466 v: rmse: 1.5644460083090737 baseline_rmse: 1.2961683303676375 nrmse: 1.2069774979499333 w: rmse: 1.9603643402199182 baseline_rmse: 1.8289556685792812 nrmse: 1.0718490195788695 after: u: rmse: 0.009593514818976712 baseline_rmse: 1.2663362526593396 nrmse: 0.007575803661018215 v: rmse: 0.01717958501003693 baseline_rmse: 1.2961683303676375 nrmse: 0.0132541311244383 w: rmse: 0.028883275348878173 baseline_rmse: 1.8289556685792812 nrmse: 0.01579222276683966 initial_old_prediction_drift: u: 3.4934839010238647 v: 3.6688733100891113 steps: 768 optimizer_covers_current_parameters: true validation_curve: - step: 0 scores: u: rmse: 1.5127669768475775 baseline_rmse: 1.3104433241303899 nrmse: 1.1543932873643732 v: rmse: 1.5470830375162308 baseline_rmse: 1.2705869192092003 nrmse: 1.2176129111097087 w: rmse: 1.9448778758458212 baseline_rmse: 1.8209322963213432 nrmse: 1.0680670993506312 - step: 32 scores: u: rmse: 0.1180748304040611 baseline_rmse: 1.3104433241303899 nrmse: 0.09010296609539795 v: rmse: 0.09132736641290372 baseline_rmse: 1.2705869192092003 nrmse: 0.07187809431388203 w: rmse: 0.15109867945633018 baseline_rmse: 1.8209322963213432 nrmse: 0.08297874652538181 - step: 128 scores: u: rmse: 0.06649819828470592 baseline_rmse: 1.3104433241303899 nrmse: 0.050744810599751895 v: rmse: 0.036841081833149394 baseline_rmse: 1.2705869192092003 nrmse: 0.028995325920779108 w: rmse: 0.043636162065740156 baseline_rmse: 1.8209322963213432 nrmse: 0.023963637832056774 - step: 768 scores: u: rmse: 0.00949244348164829 baseline_rmse: 1.3104433241303899 nrmse: 0.007243688686763676 v: rmse: 0.016199661195887685 baseline_rmse: 1.2705869192092003 nrmse: 0.012749746554899353 w: rmse: 0.02952968655074886 baseline_rmse: 1.8209322963213432 nrmse: 0.016216795435176193 test_rows: 1024 baseline: Source-training mean for test comparisons; phase-training mean for validation curves optimizer_policy: New AdamW factory and Trainer for every fit; all available labels rehearsed mutation: extend record with hidden Number record/w checks: Source learns both original tasks below 0.25 nRMSE: true Extension preserves every existing state entry: true Extension immediately preserves old predictions: false 'extended: fresh optimizer covers current parameters': true 'extended: all final targets below 0.25 nRMSE': true Added head parameters actually learn: true Hidden target placeholders cannot affect predictions: true Extended checkpoint preserves schema and learned state: true Extended checkpoint preserves predictions: true 'continuation: fresh optimizer covers current parameters': true 'scratch_adaptation: fresh optimizer covers current parameters': true 'scratch_total: fresh optimizer covers current parameters': true 'scratch_total: all final targets below 0.25 nRMSE': true outcome: not_met duration_seconds: 26.161 - started_at: '2026-09-15T17:55:37.078997+00:00' mode: full seed: 4701 steps_override: null accelerator: cpu threads: 1 provenance: revision: 5d985953fd0416bc5982a3518d8f43cadb52110c source_sha256: 04f09ccb89861dbe7977f303d13c1e11bf180f1706f6ca837c89f4ba1050cf3b script_sha256: 61b3e662a7ec74440893deec64e6fb16806b8409e0f466c4c9feb98be4e8d144 code_sha256: 51ba353c87e01417438f385202d12cd43919a7e1f61896c4ab0391420ba86ace python: 3.12.10 platform: macOS-26.5.2-arm64-arm-64bit host: Granthams-MacBook-Pro.local relflow_path: /Users/grantham/Desktop/relflow/src/relflow packages: relflow: 0.2.2 torch: 2.12.0 lightning: 2.6.4 numpy: 2.4.6 pyarrow: 24.0.0 polars: 1.44.1 torchmetrics: 1.9.0 tensordict: 0.12.4 pydantic: 2.13.4 cuda: null gpu: null metrics: source: u: rmse: 0.04339685848153591 baseline_rmse: 1.2663362526593396 nrmse: 0.034269617086616105 v: rmse: 0.03464917023016499 baseline_rmse: 1.2961683303676375 nrmse: 0.02673199878316523 source_fit: steps: 512 optimizer_covers_current_parameters: true validation_curve: - step: 0 scores: u: rmse: 1.3509519179467873 baseline_rmse: 1.295883063514218 nrmse: 1.0424952343178495 v: rmse: 1.7025466041136943 baseline_rmse: 1.2898120364298942 nrmse: 1.3199959032993827 - step: 32 scores: u: rmse: 0.08500438323886007 baseline_rmse: 1.295883063514218 nrmse: 0.06559572050300774 v: rmse: 0.11706025505541162 baseline_rmse: 1.2898120364298942 nrmse: 0.09075760789101169 - step: 128 scores: u: rmse: 0.02709548454351126 baseline_rmse: 1.295883063514218 nrmse: 0.02090889626262484 v: rmse: 0.04247795145991918 baseline_rmse: 1.2898120364298942 nrmse: 0.032933443215102144 - step: 512 scores: u: rmse: 0.046377564578228725 baseline_rmse: 1.295883063514218 nrmse: 0.03578838699570664 v: rmse: 0.03358961215940764 baseline_rmse: 1.2898120364298942 nrmse: 0.026042253608038298 source_prerequisite_met: true downstream_interpretable: true arms: extended: before: u: rmse: 0.04338463782724802 baseline_rmse: 1.2663362526593396 nrmse: 0.034259966684314004 v: rmse: 0.04017069929827001 baseline_rmse: 1.2961683303676375 nrmse: 0.030991884585604892 w: rmse: 1.73040449423147 baseline_rmse: 1.8289556685792812 nrmse: 0.9461161492096936 after: u: rmse: 0.020053658093265946 baseline_rmse: 1.2663362526593396 nrmse: 0.015835966198671746 v: rmse: 0.022480374165581508 baseline_rmse: 1.2961683303676375 nrmse: 0.01734371503985544 w: rmse: 0.047069585087884114 baseline_rmse: 1.8289556685792812 nrmse: 0.025735771455000552 initial_old_prediction_drift: u: 0.015998125076293945 v: 0.020020246505737305 steps: 256 optimizer_covers_current_parameters: true validation_curve: - step: 0 scores: u: rmse: 0.043597498705624405 baseline_rmse: 1.3104433241303899 nrmse: 0.03326927452933205 v: rmse: 0.03898050824242802 baseline_rmse: 1.2705869192092003 nrmse: 0.030679135486999246 w: rmse: 1.7157907679756903 baseline_rmse: 1.8209322963213432 nrmse: 0.9422595070898241 - step: 32 scores: u: rmse: 0.05030544976923204 baseline_rmse: 1.3104433241303899 nrmse: 0.03838811556586374 v: rmse: 0.04795291787662376 baseline_rmse: 1.2705869192092003 nrmse: 0.037740761495064924 w: rmse: 0.10949826635992924 baseline_rmse: 1.8209322963213432 nrmse: 0.060133079401764794 - step: 128 scores: u: rmse: 0.007891395414127643 baseline_rmse: 1.3104433241303899 nrmse: 0.006021928051992918 v: rmse: 0.012467297531026718 baseline_rmse: 1.2705869192092003 nrmse: 0.00981223507226584 w: rmse: 0.018547623205124056 baseline_rmse: 1.8209322963213432 nrmse: 0.010185784085764232 - step: 256 scores: u: rmse: 0.020140006992766024 baseline_rmse: 1.3104433241303899 nrmse: 0.015368850084478802 v: rmse: 0.02328180518929462 baseline_rmse: 1.2705869192092003 nrmse: 0.01832366195284378 w: rmse: 0.046479505370370475 baseline_rmse: 1.8209322963213432 nrmse: 0.025525114505502818 changed_existing_state_entries: [] pooling_scale_control: source_capacity: 4 extended_capacity: 5 old_prediction_drift_with_source_capacity: u: 0.0 v: 0.0 new_head_updated_parameters: - embedder.embeddings.weight - embedder.linear.weight - embedder.scalar.weight - decoder.context_projection.0.weight - decoder.context_projection.0.bias - decoder.context_projection.1.weight - decoder.context_projection.1.bias - decoder.context_projection.3.weight - decoder.context_projection.3.bias - decoder.pool.queries - decoder.pool.blocks.0.ffn_norm.weight - decoder.pool.blocks.0.ffn_norm.bias - decoder.pool.blocks.0.attention.q_proj.weight - decoder.pool.blocks.0.attention.k_proj.weight - decoder.pool.blocks.0.attention.v_proj.weight - decoder.pool.blocks.0.attention.v_proj.bias - decoder.pool.blocks.0.attention.out_proj.weight - decoder.pool.blocks.0.attention.out_proj.bias - decoder.pool.blocks.0.ffn.0.weight - decoder.pool.blocks.0.ffn.0.bias - decoder.pool.blocks.0.ffn.3.weight - decoder.pool.blocks.0.ffn.3.bias - decoder.pool.norm.weight - decoder.pool.norm.bias - decoder.classification.weight - decoder.classification.bias - decoder.regression.weight - decoder.regression.bias continuation: before: u: rmse: 0.04339685848153591 baseline_rmse: 1.2663362526593396 nrmse: 0.034269617086616105 v: rmse: 0.03464917023016499 baseline_rmse: 1.2961683303676375 nrmse: 0.02673199878316523 after: u: rmse: 0.024847849450170354 baseline_rmse: 1.2663362526593396 nrmse: 0.01962184166961122 v: rmse: 0.03539640971424924 baseline_rmse: 1.2961683303676375 nrmse: 0.027308497580873323 initial_old_prediction_drift: u: 0.0 v: 0.0 steps: 256 optimizer_covers_current_parameters: true validation_curve: - step: 0 scores: u: rmse: 0.043578616866279236 baseline_rmse: 1.3104433241303899 nrmse: 0.03325486578765091 v: rmse: 0.033650245498206555 baseline_rmse: 1.2705869192092003 nrmse: 0.02648401694482272 - step: 32 scores: u: rmse: 0.03494173974815033 baseline_rmse: 1.3104433241303899 nrmse: 0.02666406024948669 v: rmse: 0.08441034946932946 baseline_rmse: 1.2705869192092003 nrmse: 0.06643414015458743 - step: 128 scores: u: rmse: 0.02933732798141578 baseline_rmse: 1.3104433241303899 nrmse: 0.022387330639334613 v: rmse: 0.012658927846920236 baseline_rmse: 1.2705869192092003 nrmse: 0.009963055384513968 - step: 256 scores: u: rmse: 0.02554737302907079 baseline_rmse: 1.3104433241303899 nrmse: 0.019495213992580737 v: rmse: 0.03563651009027191 baseline_rmse: 1.2705869192092003 nrmse: 0.028047282363375576 scratch_adaptation: before: u: rmse: 1.2899407966264593 baseline_rmse: 1.2663362526593396 nrmse: 1.0186400286002628 v: rmse: 1.10920508178733 baseline_rmse: 1.2961683303676375 nrmse: 0.8557569690602777 w: rmse: 1.7488962582222345 baseline_rmse: 1.8289556685792812 nrmse: 0.9562267080977221 after: u: rmse: 0.046204841875476875 baseline_rmse: 1.2663362526593396 nrmse: 0.03648702449957149 v: rmse: 0.05890874541622356 baseline_rmse: 1.2961683303676375 nrmse: 0.0454483758290214 w: rmse: 0.03371907689280475 baseline_rmse: 1.8289556685792812 nrmse: 0.01843624614422583 initial_old_prediction_drift: u: 3.1339073181152344 v: 2.9099842980504036 steps: 256 optimizer_covers_current_parameters: true validation_curve: - step: 0 scores: u: rmse: 1.341623916158934 baseline_rmse: 1.3104433241303899 nrmse: 1.0237939264174096 v: rmse: 1.0918173893993468 baseline_rmse: 1.2705869192092003 nrmse: 0.8593016132094939 w: rmse: 1.7316059452084105 baseline_rmse: 1.8209322963213432 nrmse: 0.9509447159054786 - step: 32 scores: u: rmse: 0.0754211523406791 baseline_rmse: 1.3104433241303899 nrmse: 0.05755392160185834 v: rmse: 0.08702010311291794 baseline_rmse: 1.2705869192092003 nrmse: 0.06848811505715667 w: rmse: 0.17318271728552895 baseline_rmse: 1.8209322963213432 nrmse: 0.09510662073235428 - step: 128 scores: u: rmse: 0.02006432085463077 baseline_rmse: 1.3104433241303899 nrmse: 0.015311093952075685 v: rmse: 0.07701617202253476 baseline_rmse: 1.2705869192092003 nrmse: 0.060614642617656415 w: rmse: 0.06543257210337809 baseline_rmse: 1.8209322963213432 nrmse: 0.035933555704166106 - step: 256 scores: u: rmse: 0.04627067116592861 baseline_rmse: 1.3104433241303899 nrmse: 0.03530917386040623 v: rmse: 0.05786074070198421 baseline_rmse: 1.2705869192092003 nrmse: 0.04553859309207757 w: rmse: 0.034026255949966284 baseline_rmse: 1.8209322963213432 nrmse: 0.018686173021757208 scratch_total: before: u: rmse: 1.5094414653137729 baseline_rmse: 1.2663362526593396 nrmse: 1.1919752452351466 v: rmse: 1.5644460083090737 baseline_rmse: 1.2961683303676375 nrmse: 1.2069774979499333 w: rmse: 1.9603643402199182 baseline_rmse: 1.8289556685792812 nrmse: 1.0718490195788695 after: u: rmse: 0.009593514818976712 baseline_rmse: 1.2663362526593396 nrmse: 0.007575803661018215 v: rmse: 0.01717958501003693 baseline_rmse: 1.2961683303676375 nrmse: 0.0132541311244383 w: rmse: 0.028883275348878173 baseline_rmse: 1.8289556685792812 nrmse: 0.01579222276683966 initial_old_prediction_drift: u: 3.4934839010238647 v: 3.6688733100891113 steps: 768 optimizer_covers_current_parameters: true validation_curve: - step: 0 scores: u: rmse: 1.5127669768475775 baseline_rmse: 1.3104433241303899 nrmse: 1.1543932873643732 v: rmse: 1.5470830375162308 baseline_rmse: 1.2705869192092003 nrmse: 1.2176129111097087 w: rmse: 1.9448778758458212 baseline_rmse: 1.8209322963213432 nrmse: 1.0680670993506312 - step: 32 scores: u: rmse: 0.1180748304040611 baseline_rmse: 1.3104433241303899 nrmse: 0.09010296609539795 v: rmse: 0.09132736641290372 baseline_rmse: 1.2705869192092003 nrmse: 0.07187809431388203 w: rmse: 0.15109867945633018 baseline_rmse: 1.8209322963213432 nrmse: 0.08297874652538181 - step: 128 scores: u: rmse: 0.06649819828470592 baseline_rmse: 1.3104433241303899 nrmse: 0.050744810599751895 v: rmse: 0.036841081833149394 baseline_rmse: 1.2705869192092003 nrmse: 0.028995325920779108 w: rmse: 0.043636162065740156 baseline_rmse: 1.8209322963213432 nrmse: 0.023963637832056774 - step: 768 scores: u: rmse: 0.00949244348164829 baseline_rmse: 1.3104433241303899 nrmse: 0.007243688686763676 v: rmse: 0.016199661195887685 baseline_rmse: 1.2705869192092003 nrmse: 0.012749746554899353 w: rmse: 0.02952968655074886 baseline_rmse: 1.8209322963213432 nrmse: 0.016216795435176193 test_rows: 1024 baseline: Source-training mean for test comparisons; phase-training mean for validation curves optimizer_policy: New AdamW factory and Trainer for every fit; all available labels rehearsed mutation: extend record with hidden Number record/w checks: Source learns both original tasks below 0.25 nRMSE: true Extension preserves every existing state entry: true Extension immediately preserves old predictions: false Diagnostic source pooling scale restores old predictions: true 'extended: fresh optimizer covers current parameters': true 'extended: all final targets below 0.25 nRMSE': true Added head parameters actually learn: true Hidden target placeholders cannot affect predictions: true Extended checkpoint preserves schema and learned state: true Extended checkpoint preserves predictions: true 'continuation: fresh optimizer covers current parameters': true 'scratch_adaptation: fresh optimizer covers current parameters': true 'scratch_total: fresh optimizer covers current parameters': true 'scratch_total: all final targets below 0.25 nRMSE': true outcome: not_met duration_seconds: 27.232 - started_at: '2026-09-15T17:58:20.909761+00:00' mode: full seed: 4701 steps_override: null accelerator: gpu threads: 1 provenance: revision: null source_sha256: 04f09ccb89861dbe7977f303d13c1e11bf180f1706f6ca837c89f4ba1050cf3b script_sha256: 61b3e662a7ec74440893deec64e6fb16806b8409e0f466c4c9feb98be4e8d144 code_sha256: 51ba353c87e01417438f385202d12cd43919a7e1f61896c4ab0391420ba86ace python: 3.12.6 platform: Linux-6.8.0-138-generic-x86_64-with-glibc2.35 host: lab relflow_path: /home/grantham/relflow-mutations-v3_tgit_/src/relflow packages: relflow: 0.2.2 torch: 2.12.0 lightning: 2.6.4 numpy: 2.4.6 pyarrow: 24.0.0 polars: 1.44.1 torchmetrics: 1.9.0 tensordict: 0.12.4 pydantic: 2.13.4 cuda: '13.0' gpu: NVIDIA GeForce RTX 3090 metrics: source: u: rmse: 0.01819302748771275 baseline_rmse: 1.2663362526593396 nrmse: 0.014366664027431033 v: rmse: 0.0131784870577723 baseline_rmse: 1.2961683303676375 nrmse: 0.010167265122142301 source_fit: steps: 512 optimizer_covers_current_parameters: true validation_curve: - step: 0 scores: u: rmse: 1.3509519123451783 baseline_rmse: 1.295883063514218 nrmse: 1.0424952299952304 v: rmse: 1.7025465773867603 baseline_rmse: 1.2898120364298942 nrmse: 1.3199958825778098 - step: 32 scores: u: rmse: 0.08500439261560418 baseline_rmse: 1.295883063514218 nrmse: 0.06559572773880268 v: rmse: 0.11706019494570107 baseline_rmse: 1.2898120364298942 nrmse: 0.09075756128754633 - step: 128 scores: u: rmse: 0.027080408434786946 baseline_rmse: 1.295883063514218 nrmse: 0.020897262412975296 v: rmse: 0.04247991507908372 baseline_rmse: 1.2898120364298942 nrmse: 0.03293496562232822 - step: 512 scores: u: rmse: 0.018074103467788448 baseline_rmse: 1.295883063514218 nrmse: 0.013947325940640434 v: rmse: 0.012383353951690268 baseline_rmse: 1.2898120364298942 nrmse: 0.009600898116881038 source_prerequisite_met: true downstream_interpretable: true arms: extended: before: u: rmse: 0.014853532158426426 baseline_rmse: 1.2663362526593396 nrmse: 0.011729532442298494 v: rmse: 0.018807925601864606 baseline_rmse: 1.2961683303676375 nrmse: 0.014510403595904891 w: rmse: 1.7267224752213859 baseline_rmse: 1.8289556685792812 nrmse: 0.9441029680958264 after: u: rmse: 0.019240178824238174 baseline_rmse: 1.2663362526593396 nrmse: 0.015193578154169787 v: rmse: 0.03134749681726191 baseline_rmse: 1.2961683303676375 nrmse: 0.02418474212247625 w: rmse: 0.02621196407253994 baseline_rmse: 1.8289556685792812 nrmse: 0.014331656323250959 initial_old_prediction_drift: u: 0.015697479248046875 v: 0.019600391387939453 steps: 256 optimizer_covers_current_parameters: true validation_curve: - step: 0 scores: u: rmse: 0.015278067071696535 baseline_rmse: 1.3104433241303899 nrmse: 0.011658701135995377 v: rmse: 0.01805821366500976 baseline_rmse: 1.2705869192092003 nrmse: 0.014212497698503773 w: rmse: 1.712543069857594 baseline_rmse: 1.8209322963213432 nrmse: 0.9404759711919451 - step: 32 scores: u: rmse: 0.07303491092190062 baseline_rmse: 1.3104433241303899 nrmse: 0.05573297950170152 v: rmse: 0.07692725262793854 baseline_rmse: 1.2705869192092003 nrmse: 0.060544659688309424 w: rmse: 0.09891990919687285 baseline_rmse: 1.8209322963213432 nrmse: 0.05432377106865058 - step: 128 scores: u: rmse: 0.02054482042767095 baseline_rmse: 1.3104433241303899 nrmse: 0.015677763432695183 v: rmse: 0.012109868563221179 baseline_rmse: 1.2705869192092003 nrmse: 0.0095309249451098 w: rmse: 0.03621440731686269 baseline_rmse: 1.8209322963213432 nrmse: 0.01988783843859721 - step: 256 scores: u: rmse: 0.019279549956777758 baseline_rmse: 1.3104433241303899 nrmse: 0.014712234861108293 v: rmse: 0.030539371964238324 baseline_rmse: 1.2705869192092003 nrmse: 0.024035641720005822 w: rmse: 0.026819529271701184 baseline_rmse: 1.8209322963213432 nrmse: 0.014728460429793097 changed_existing_state_entries: [] pooling_scale_control: source_capacity: 4 extended_capacity: 5 old_prediction_drift_with_source_capacity: u: 0.0 v: 0.0 new_head_updated_parameters: - embedder.embeddings.weight - embedder.linear.weight - embedder.scalar.weight - decoder.context_projection.0.weight - decoder.context_projection.0.bias - decoder.context_projection.1.weight - decoder.context_projection.1.bias - decoder.context_projection.3.weight - decoder.context_projection.3.bias - decoder.pool.queries - decoder.pool.blocks.0.attention_norm.weight - decoder.pool.blocks.0.attention_norm.bias - decoder.pool.blocks.0.ffn_norm.weight - decoder.pool.blocks.0.ffn_norm.bias - decoder.pool.blocks.0.attention.q_proj.weight - decoder.pool.blocks.0.attention.q_proj.bias - decoder.pool.blocks.0.attention.k_proj.weight - decoder.pool.blocks.0.attention.k_proj.bias - decoder.pool.blocks.0.attention.v_proj.weight - decoder.pool.blocks.0.attention.v_proj.bias - decoder.pool.blocks.0.attention.out_proj.weight - decoder.pool.blocks.0.attention.out_proj.bias - decoder.pool.blocks.0.ffn.0.weight - decoder.pool.blocks.0.ffn.0.bias - decoder.pool.blocks.0.ffn.3.weight - decoder.pool.blocks.0.ffn.3.bias - decoder.pool.norm.weight - decoder.pool.norm.bias - decoder.classification.weight - decoder.classification.bias - decoder.regression.weight - decoder.regression.bias continuation: before: u: rmse: 0.01819302748771275 baseline_rmse: 1.2663362526593396 nrmse: 0.014366664027431033 v: rmse: 0.0131784870577723 baseline_rmse: 1.2961683303676375 nrmse: 0.010167265122142301 after: u: rmse: 0.04389199309315439 baseline_rmse: 1.2663362526593396 nrmse: 0.034660614825627906 v: rmse: 0.014499176546597178 baseline_rmse: 1.2961683303676375 nrmse: 0.011186183311919616 initial_old_prediction_drift: u: 0.0 v: 0.0 steps: 256 optimizer_covers_current_parameters: true validation_curve: - step: 0 scores: u: rmse: 0.018735863047899252 baseline_rmse: 1.3104433241303899 nrmse: 0.014297347090788814 v: rmse: 0.012682751132877708 baseline_rmse: 1.2705869192092003 nrmse: 0.009981805212327636 - step: 32 scores: u: rmse: 0.047761862787583105 baseline_rmse: 1.3104433241303899 nrmse: 0.036447102982708444 v: rmse: 0.0863394126802456 baseline_rmse: 1.2705869192092003 nrmse: 0.06795238592097448 - step: 128 scores: u: rmse: 0.010687123759535412 baseline_rmse: 1.3104433241303899 nrmse: 0.00815534984439513 v: rmse: 0.03350553736899516 baseline_rmse: 1.2705869192092003 nrmse: 0.026370126169604085 - step: 256 scores: u: rmse: 0.04446787523680337 baseline_rmse: 1.3104433241303899 nrmse: 0.033933459324776406 v: rmse: 0.014381186909938976 baseline_rmse: 1.2705869192092003 nrmse: 0.011318538458502054 scratch_adaptation: before: u: rmse: 1.2899407763604933 baseline_rmse: 1.2663362526593396 nrmse: 1.0186400125966413 v: rmse: 1.109205002962587 baseline_rmse: 1.2961683303676375 nrmse: 0.8557569082466154 w: rmse: 1.748896289350403 baseline_rmse: 1.8289556685792812 nrmse: 0.956226725117363 after: u: rmse: 0.04540419300395933 baseline_rmse: 1.2663362526593396 nrmse: 0.03585476835920106 v: rmse: 0.056871312722615336 baseline_rmse: 1.2961683303676375 nrmse: 0.043876486865316866 w: rmse: 0.033184504097638984 baseline_rmse: 1.8289556685792812 nrmse: 0.018143963064679667 initial_old_prediction_drift: u: 3.0743343234062195 v: 2.957523114979267 steps: 256 optimizer_covers_current_parameters: true validation_curve: - step: 0 scores: u: rmse: 1.341623887228593 baseline_rmse: 1.3104433241303899 nrmse: 1.023793904340651 v: rmse: 1.0918173067057397 baseline_rmse: 1.2705869192092003 nrmse: 0.8593015481264951 w: rmse: 1.731605942968808 baseline_rmse: 1.8209322963213432 nrmse: 0.9509447146755577 - step: 32 scores: u: rmse: 0.07542124900297918 baseline_rmse: 1.3104433241303899 nrmse: 0.05755399536491112 v: rmse: 0.08702007920715463 baseline_rmse: 1.2705869192092003 nrmse: 0.06848809624241606 w: rmse: 0.17318279451152083 baseline_rmse: 1.8209322963213432 nrmse: 0.09510666314249333 - step: 128 scores: u: rmse: 0.02006893947490001 baseline_rmse: 1.3104433241303899 nrmse: 0.015314618423668002 v: rmse: 0.07701768016981209 baseline_rmse: 1.2705869192092003 nrmse: 0.060615829586650447 w: rmse: 0.0654289532399687 baseline_rmse: 1.8209322963213432 nrmse: 0.03593156833570836 - step: 256 scores: u: rmse: 0.04544946245375731 baseline_rmse: 1.3104433241303899 nrmse: 0.03468250905388646 v: rmse: 0.055825085656781126 baseline_rmse: 1.2705869192092003 nrmse: 0.043936455517365205 w: rmse: 0.033380081026084765 baseline_rmse: 1.8209322963213432 nrmse: 0.018331313631769493 scratch_total: before: u: rmse: 1.509441428360601 baseline_rmse: 1.2663362526593396 nrmse: 1.1919752160539778 v: rmse: 1.564446056193685 baseline_rmse: 1.2961683303676375 nrmse: 1.2069775348931375 w: rmse: 1.9603643690797954 baseline_rmse: 1.8289556685792812 nrmse: 1.071849035358299 after: u: rmse: 0.014952189544917813 baseline_rmse: 1.2663362526593396 nrmse: 0.011807440175165025 v: rmse: 0.014238709417792353 baseline_rmse: 1.2961683303676375 nrmse: 0.010985231689586004 w: rmse: 0.03583276384072388 baseline_rmse: 1.8289556685792812 nrmse: 0.01959192584944309 initial_old_prediction_drift: u: 3.5350125432014465 v: 3.7103660702705383 steps: 768 optimizer_covers_current_parameters: true validation_curve: - step: 0 scores: u: rmse: 1.5127669405502646 baseline_rmse: 1.3104433241303899 nrmse: 1.1543932596658744 v: rmse: 1.5470830785671061 baseline_rmse: 1.2705869192092003 nrmse: 1.2176129434183016 w: rmse: 1.9448779062015815 baseline_rmse: 1.8209322963213432 nrmse: 1.068067116021081 - step: 32 scores: u: rmse: 0.11807480780524139 baseline_rmse: 1.3104433241303899 nrmse: 0.09010294885022657 v: rmse: 0.09132753090093358 baseline_rmse: 1.2705869192092003 nrmse: 0.07187822377218779 w: rmse: 0.1510987225721377 baseline_rmse: 1.8209322963213432 nrmse: 0.08297877020325693 - step: 128 scores: u: rmse: 0.06653331588804018 baseline_rmse: 1.3104433241303899 nrmse: 0.050771608861597796 v: rmse: 0.03689510605571201 baseline_rmse: 1.2705869192092003 nrmse: 0.02903784502887463 w: rmse: 0.04361173320269421 baseline_rmse: 1.8209322963213432 nrmse: 0.02395022225197436 - step: 768 scores: u: rmse: 0.015076530816569185 baseline_rmse: 1.3104433241303899 nrmse: 0.011504908712152026 v: rmse: 0.0148685561383759 baseline_rmse: 1.2705869192092003 nrmse: 0.011702116489307107 w: rmse: 0.03509559328366416 baseline_rmse: 1.8209322963213432 nrmse: 0.019273420189517457 test_rows: 1024 baseline: Source-training mean for test comparisons; phase-training mean for validation curves optimizer_policy: New AdamW factory and Trainer for every fit; all available labels rehearsed mutation: extend record with hidden Number record/w checks: Source learns both original tasks below 0.25 nRMSE: true Extension preserves every existing state entry: true Extension immediately preserves old predictions: false Diagnostic source pooling scale restores old predictions: true 'extended: fresh optimizer covers current parameters': true 'extended: all final targets below 0.25 nRMSE': true Added head parameters actually learn: true Hidden target placeholders cannot affect predictions: true Extended checkpoint preserves schema and learned state: true Extended checkpoint preserves predictions: true 'continuation: fresh optimizer covers current parameters': true 'scratch_adaptation: fresh optimizer covers current parameters': true 'scratch_total: fresh optimizer covers current parameters': true 'scratch_total: all final targets below 0.25 nRMSE': true outcome: not_met duration_seconds: 64.293 - started_at: '2026-09-15T17:59:46.523871+00:00' mode: full seed: 4701 steps_override: null accelerator: cpu threads: 1 provenance: revision: 5d985953fd0416bc5982a3518d8f43cadb52110c source_sha256: 2db77891e89376c6926a1f11a18aa46fb4f8bebed0df14438ba8eff89c4db6c1 script_sha256: 624c767c1b4024e120b601bc1efaadbe9a478f8c3ab6d8b0e375600c396a001e code_sha256: 5f613c009ea56a099f0f766dcac921baef4eabf53cf14a0a335ebfb2b4ca271d python: 3.12.10 platform: macOS-26.5.2-arm64-arm-64bit host: Granthams-MacBook-Pro.local relflow_path: /Users/grantham/Desktop/relflow/src/relflow packages: relflow: 0.2.2 torch: 2.12.0 lightning: 2.6.4 numpy: 2.4.6 pyarrow: 24.0.0 polars: 1.44.1 torchmetrics: 1.9.0 tensordict: 0.12.4 pydantic: 2.13.4 cuda: null gpu: null metrics: source: u: rmse: 0.02812596173874027 baseline_rmse: 1.2663362526593396 nrmse: 0.02221050031512168 v: rmse: 0.02869417926503313 baseline_rmse: 1.2961683303676375 nrmse: 0.022137695076143764 source_fit: steps: 512 optimizer_covers_current_parameters: true validation_curve: - step: 0 scores: u: rmse: 1.1675138399845537 baseline_rmse: 1.295883063514218 nrmse: 0.9009407352068107 v: rmse: 1.6476149821345185 baseline_rmse: 1.2898120364298942 nrmse: 1.2774070450567332 - step: 32 scores: u: rmse: 0.09449283629063597 baseline_rmse: 1.295883063514218 nrmse: 0.07291771838917874 v: rmse: 0.10256670686191162 baseline_rmse: 1.2898120364298942 nrmse: 0.07952066189877464 - step: 128 scores: u: rmse: 0.07973298418747886 baseline_rmse: 1.295883063514218 nrmse: 0.06152791593035899 v: rmse: 0.08406665221429643 baseline_rmse: 1.2898120364298942 nrmse: 0.06517744434063959 - step: 512 scores: u: rmse: 0.027936281602329575 baseline_rmse: 1.295883063514218 nrmse: 0.02155771796767762 v: rmse: 0.02875704011542896 baseline_rmse: 1.2898120364298942 nrmse: 0.022295527800334655 source_prerequisite_met: true downstream_interpretable: true arms: extended: before: u: rmse: 0.03620714967850646 baseline_rmse: 1.2663362526593396 nrmse: 0.028592050178197527 v: rmse: 0.02858616864862413 baseline_rmse: 1.2961683303676375 nrmse: 0.02205436437450691 w: rmse: 1.8046207412478505 baseline_rmse: 1.8289556685792812 nrmse: 0.9866946324891878 after: u: rmse: 0.029878410352648474 baseline_rmse: 1.2663362526593396 nrmse: 0.023594373366396977 v: rmse: 0.01723088652453224 baseline_rmse: 1.2961683303676375 nrmse: 0.013293710485616459 w: rmse: 0.017570523323954566 baseline_rmse: 1.8289556685792812 nrmse: 0.009606861240985252 initial_old_prediction_drift: u: 0.03130531311035156 v: 0.028259754180908203 steps: 256 optimizer_covers_current_parameters: true validation_curve: - step: 0 scores: u: rmse: 0.036365910348353285 baseline_rmse: 1.3104433241303899 nrmse: 0.027750845594551522 v: rmse: 0.028621911686710133 baseline_rmse: 1.2705869192092003 nrmse: 0.0225265279013923 w: rmse: 1.8242640005983064 baseline_rmse: 1.8209322963213432 nrmse: 1.0018296694960565 - step: 32 scores: u: rmse: 0.0829061445372226 baseline_rmse: 1.3104433241303899 nrmse: 0.06326572314162393 v: rmse: 0.09543959500564209 baseline_rmse: 1.2705869192092003 nrmse: 0.07511457387349987 w: rmse: 0.1167811929072492 baseline_rmse: 1.8209322963213432 nrmse: 0.06413263861768566 - step: 128 scores: u: rmse: 0.016524041520787628 baseline_rmse: 1.3104433241303899 nrmse: 0.012609504903047204 v: rmse: 0.026544684355906265 baseline_rmse: 1.2705869192092003 nrmse: 0.020891671364307286 w: rmse: 0.0275912369630025 baseline_rmse: 1.8209322963213432 nrmse: 0.015152258553897065 - step: 256 scores: u: rmse: 0.02991676615019795 baseline_rmse: 1.3104433241303899 nrmse: 0.02282950021516628 v: rmse: 0.01665090444553678 baseline_rmse: 1.2705869192092003 nrmse: 0.013104892072948559 w: rmse: 0.016667696211426185 baseline_rmse: 1.8209322963213432 nrmse: 0.00915338601281242 changed_existing_state_entries: [] pooling_scale_control: source_capacity: 4 extended_capacity: 5 old_prediction_drift_with_source_capacity: u: 0.0 v: 0.0 new_head_updated_parameters: - embedder.embeddings.weight - embedder.linear.weight - embedder.scalar.weight - decoder.context_projection.0.weight - decoder.context_projection.0.bias - decoder.context_projection.1.weight - decoder.context_projection.1.bias - decoder.context_projection.3.weight - decoder.context_projection.3.bias - decoder.pool.queries - decoder.pool.blocks.0.ffn_norm.weight - decoder.pool.blocks.0.ffn_norm.bias - decoder.pool.blocks.0.attention.q_proj.weight - decoder.pool.blocks.0.attention.k_proj.weight - decoder.pool.blocks.0.attention.v_proj.weight - decoder.pool.blocks.0.attention.v_proj.bias - decoder.pool.blocks.0.attention.out_proj.weight - decoder.pool.blocks.0.attention.out_proj.bias - decoder.pool.blocks.0.ffn.0.weight - decoder.pool.blocks.0.ffn.0.bias - decoder.pool.blocks.0.ffn.3.weight - decoder.pool.blocks.0.ffn.3.bias - decoder.pool.norm.weight - decoder.pool.norm.bias - decoder.classification.weight - decoder.classification.bias - decoder.regression.weight - decoder.regression.bias continuation: before: u: rmse: 0.02812596173874027 baseline_rmse: 1.2663362526593396 nrmse: 0.02221050031512168 v: rmse: 0.02869417926503313 baseline_rmse: 1.2961683303676375 nrmse: 0.022137695076143764 after: u: rmse: 0.011916546676619125 baseline_rmse: 1.2663362526593396 nrmse: 0.009410254702567396 v: rmse: 0.021253849793999875 baseline_rmse: 1.2961683303676375 nrmse: 0.016397445683595403 initial_old_prediction_drift: u: 0.0 v: 0.0 steps: 256 optimizer_covers_current_parameters: true validation_curve: - step: 0 scores: u: rmse: 0.02812612461247093 baseline_rmse: 1.3104433241303899 nrmse: 0.021463060702098982 v: rmse: 0.028593799530620252 baseline_rmse: 1.2705869192092003 nrmse: 0.02250440257044101 - step: 32 scores: u: rmse: 0.04618764406803422 baseline_rmse: 1.3104433241303899 nrmse: 0.0352458158376932 v: rmse: 0.12720398522725013 baseline_rmse: 1.2705869192092003 nrmse: 0.10011435133176133 - step: 128 scores: u: rmse: 0.030859885900039748 baseline_rmse: 1.3104433241303899 nrmse: 0.023549195399593773 v: rmse: 0.015013459278503071 baseline_rmse: 1.2705869192092003 nrmse: 0.011816160745498063 - step: 256 scores: u: rmse: 0.011949764555192689 baseline_rmse: 1.3104433241303899 nrmse: 0.009118871709406092 v: rmse: 0.021453482848453262 baseline_rmse: 1.2705869192092003 nrmse: 0.016884703064474864 scratch_adaptation: before: u: rmse: 1.2522115077441496 baseline_rmse: 1.2663362526593396 nrmse: 0.9888459760308309 v: rmse: 1.3195483181509495 baseline_rmse: 1.2961683303676375 nrmse: 1.0180377712026651 w: rmse: 1.880565184544753 baseline_rmse: 1.8289556685792812 nrmse: 1.028218024554724 after: u: rmse: 0.02305263211996044 baseline_rmse: 1.2663362526593396 nrmse: 0.018204195032361512 v: rmse: 0.03041060302085751 baseline_rmse: 1.2961683303676375 nrmse: 0.02346192412541975 w: rmse: 0.030353420581652007 baseline_rmse: 1.8289556685792812 nrmse: 0.016596039533987344 initial_old_prediction_drift: u: 2.8804617300629616 v: 3.253007024526596 steps: 256 optimizer_covers_current_parameters: true validation_curve: - step: 0 scores: u: rmse: 1.3100973727454814 baseline_rmse: 1.3104433241303899 nrmse: 0.9997360043135494 v: rmse: 1.3005482675294557 baseline_rmse: 1.2705869192092003 nrmse: 1.023580715232692 w: rmse: 1.8851250732582583 baseline_rmse: 1.8209322963213432 nrmse: 1.0352526983384267 - step: 32 scores: u: rmse: 0.08970207863563875 baseline_rmse: 1.3104433241303899 nrmse: 0.0684517040789727 v: rmse: 0.08707615899438154 baseline_rmse: 1.2705869192092003 nrmse: 0.06853223315771015 w: rmse: 0.13807697597571358 baseline_rmse: 1.8209322963213432 nrmse: 0.0758276275590572 - step: 128 scores: u: rmse: 0.020001989634006886 baseline_rmse: 1.3104433241303899 nrmse: 0.015263528964352735 v: rmse: 0.04367162099246063 baseline_rmse: 1.2705869192092003 nrmse: 0.03437121879047943 w: rmse: 0.042662652300409235 baseline_rmse: 1.8209322963213432 nrmse: 0.023429016216910722 - step: 256 scores: u: rmse: 0.022727568289104348 baseline_rmse: 1.3104433241303899 nrmse: 0.017343419490641734 v: rmse: 0.031449760681353835 baseline_rmse: 1.2705869192092003 nrmse: 0.024752152100643243 w: rmse: 0.029936998382595555 baseline_rmse: 1.8209322963213432 nrmse: 0.016440478563137375 scratch_total: before: u: rmse: 1.2108697453294641 baseline_rmse: 1.2663362526593396 nrmse: 0.9561992265376638 v: rmse: 1.4984804390435351 baseline_rmse: 1.2961683303676375 nrmse: 1.1560847491301651 w: rmse: 1.7811620013140086 baseline_rmse: 1.8289556685792812 nrmse: 0.9738683292950461 after: u: rmse: 0.024418871191067566 baseline_rmse: 1.2663362526593396 nrmse: 0.019283086257530172 v: rmse: 0.01390915970708952 baseline_rmse: 1.2961683303676375 nrmse: 0.010730982528437806 w: rmse: 0.030269642862249965 baseline_rmse: 1.8289556685792812 nrmse: 0.016550233219027773 initial_old_prediction_drift: u: 2.729536585509777 v: 3.533669948577881 steps: 768 optimizer_covers_current_parameters: true validation_curve: - step: 0 scores: u: rmse: 1.256439849717246 baseline_rmse: 1.3104433241303899 nrmse: 0.958789919854809 v: rmse: 1.4780311219579572 baseline_rmse: 1.2705869192092003 nrmse: 1.163266439794507 w: rmse: 1.7783453205838398 baseline_rmse: 1.8209322963213432 nrmse: 0.9766125430234074 - step: 32 scores: u: rmse: 0.09396981368876796 baseline_rmse: 1.3104433241303899 nrmse: 0.07170841497561622 v: rmse: 0.09152444125120604 baseline_rmse: 1.2705869192092003 nrmse: 0.07203319967135335 w: rmse: 0.1645743642753917 baseline_rmse: 1.8209322963213432 nrmse: 0.09037917807700246 - step: 128 scores: u: rmse: 0.03882999888277942 baseline_rmse: 1.3104433241303899 nrmse: 0.029631192870204447 v: rmse: 0.06739170171141323 baseline_rmse: 1.2705869192092003 nrmse: 0.053039820174881944 w: rmse: 0.05676309714715544 baseline_rmse: 1.8209322963213432 nrmse: 0.031172546756311884 - step: 768 scores: u: rmse: 0.024525948593334793 baseline_rmse: 1.3104433241303899 nrmse: 0.018715764460557813 v: rmse: 0.014408844714654189 baseline_rmse: 1.2705869192092003 nrmse: 0.011340306197723254 w: rmse: 0.030040026284371683 baseline_rmse: 1.8209322963213432 nrmse: 0.016497058317356827 test_rows: 1024 baseline: Source-training mean for test comparisons; phase-training mean for validation curves optimizer_policy: New AdamW factory and Trainer for every fit; all available labels rehearsed mutation: extend record with hidden Number record/w checks: Source learns both original tasks below 0.25 nRMSE: true Source optimizer covers current parameters: true Extension preserves every existing state entry: true Extension immediately preserves old predictions: false Diagnostic source pooling scale restores old predictions: true 'extended: fresh optimizer covers current parameters': true 'extended: all final targets below 0.25 nRMSE': true Added head parameters actually learn: true Hidden target placeholders cannot affect predictions: true Extended checkpoint preserves schema and learned state: true Extended checkpoint preserves predictions: true 'continuation: fresh optimizer covers current parameters': true 'scratch_adaptation: schema and field order match extension': true 'scratch_adaptation: fresh optimizer covers current parameters': true 'scratch_total: schema and field order match extension': true 'scratch_total: fresh optimizer covers current parameters': true 'scratch_total: all final targets below 0.25 nRMSE': true outcome: not_met duration_seconds: 25.312 - started_at: '2026-09-15T18:00:23.122619+00:00' mode: full seed: 7101 steps_override: null accelerator: gpu threads: 1 provenance: revision: null source_sha256: 04f09ccb89861dbe7977f303d13c1e11bf180f1706f6ca837c89f4ba1050cf3b script_sha256: 61b3e662a7ec74440893deec64e6fb16806b8409e0f466c4c9feb98be4e8d144 code_sha256: 51ba353c87e01417438f385202d12cd43919a7e1f61896c4ab0391420ba86ace python: 3.12.6 platform: Linux-6.8.0-138-generic-x86_64-with-glibc2.35 host: lab relflow_path: /home/grantham/relflow-mutations-v3_tgit_/src/relflow packages: relflow: 0.2.2 torch: 2.12.0 lightning: 2.6.4 numpy: 2.4.6 pyarrow: 24.0.0 polars: 1.44.1 torchmetrics: 1.9.0 tensordict: 0.12.4 pydantic: 2.13.4 cuda: '13.0' gpu: NVIDIA GeForce RTX 3090 metrics: source: u: rmse: 0.012642827012693 baseline_rmse: 1.2908974940031006 nrmse: 0.009793827218214922 v: rmse: 0.020391549975120982 baseline_rmse: 1.2969480983896875 nrmse: 0.015722718588692543 source_fit: steps: 512 optimizer_covers_current_parameters: true validation_curve: - step: 0 scores: u: rmse: 1.4481777383281333 baseline_rmse: 1.214202473192484 nrmse: 1.1926987222489027 v: rmse: 1.3169269796610394 baseline_rmse: 1.3285155523136438 nrmse: 0.9912770515690068 - step: 32 scores: u: rmse: 0.10487309515287249 baseline_rmse: 1.214202473192484 nrmse: 0.08637199929030885 v: rmse: 0.07997501419402694 baseline_rmse: 1.3285155523136438 nrmse: 0.06019877904684549 - step: 128 scores: u: rmse: 0.04253704243701743 baseline_rmse: 1.214202473192484 nrmse: 0.035032907094296586 v: rmse: 0.03825585841345199 baseline_rmse: 1.3285155523136438 nrmse: 0.028795943221611848 - step: 512 scores: u: rmse: 0.012037955555919891 baseline_rmse: 1.214202473192484 nrmse: 0.009914290097160383 v: rmse: 0.020201592321032528 baseline_rmse: 1.3285155523136438 nrmse: 0.015206139127126466 source_prerequisite_met: true downstream_interpretable: true arms: extended: before: u: rmse: 0.011472294450191853 baseline_rmse: 1.2908974940031006 nrmse: 0.008887068495745564 v: rmse: 0.021527162316054067 baseline_rmse: 1.2969480983896875 nrmse: 0.016598322124672955 w: rmse: 2.435462882951994 baseline_rmse: 1.8516566868110593 nrmse: 1.3152885739020936 after: u: rmse: 0.027158872331041452 baseline_rmse: 1.2908974940031006 nrmse: 0.021038752075365187 v: rmse: 0.03049903148222875 baseline_rmse: 1.2969480983896875 nrmse: 0.0235160000003831 w: rmse: 0.021762923572475352 baseline_rmse: 1.8516566868110593 nrmse: 0.011753217390398469 initial_old_prediction_drift: u: 0.015616416931152344 v: 0.007040500640869141 steps: 256 optimizer_covers_current_parameters: true validation_curve: - step: 0 scores: u: rmse: 0.011989812926413383 baseline_rmse: 1.2520463239225392 nrmse: 0.009576173578666376 v: rmse: 0.02191870629844032 baseline_rmse: 1.3128148986756512 nrmse: 0.016695960961862634 w: rmse: 2.3604918548808027 baseline_rmse: 1.7725715949102796 nrmse: 1.3316764533848244 - step: 32 scores: u: rmse: 0.09033914836785212 baseline_rmse: 1.2520463239225392 nrmse: 0.0721531996394737 v: rmse: 0.07151185546255423 baseline_rmse: 1.3128148986756512 nrmse: 0.054472154097804926 w: rmse: 0.12841173673469872 baseline_rmse: 1.7725715949102796 nrmse: 0.07244375183683252 - step: 128 scores: u: rmse: 0.012968520732906323 baseline_rmse: 1.2520463239225392 nrmse: 0.010357860156705073 v: rmse: 0.01598330308986742 baseline_rmse: 1.3128148986756512 nrmse: 0.01217483371493662 w: rmse: 0.021488180853669117 baseline_rmse: 1.7725715949102796 nrmse: 0.012122602503261236 - step: 256 scores: u: rmse: 0.027410738175180402 baseline_rmse: 1.2520463239225392 nrmse: 0.021892750812370288 v: rmse: 0.030175315962539417 baseline_rmse: 1.3128148986756512 nrmse: 0.022985202249745828 w: rmse: 0.02093588394903854 baseline_rmse: 1.7725715949102796 nrmse: 0.011811023040848306 changed_existing_state_entries: [] pooling_scale_control: source_capacity: 4 extended_capacity: 5 old_prediction_drift_with_source_capacity: u: 0.0 v: 0.0 new_head_updated_parameters: - embedder.embeddings.weight - embedder.linear.weight - embedder.scalar.weight - decoder.context_projection.0.weight - decoder.context_projection.0.bias - decoder.context_projection.1.weight - decoder.context_projection.1.bias - decoder.context_projection.3.weight - decoder.context_projection.3.bias - decoder.pool.queries - decoder.pool.blocks.0.attention_norm.weight - decoder.pool.blocks.0.attention_norm.bias - decoder.pool.blocks.0.ffn_norm.weight - decoder.pool.blocks.0.ffn_norm.bias - decoder.pool.blocks.0.attention.q_proj.weight - decoder.pool.blocks.0.attention.q_proj.bias - decoder.pool.blocks.0.attention.k_proj.weight - decoder.pool.blocks.0.attention.k_proj.bias - decoder.pool.blocks.0.attention.v_proj.weight - decoder.pool.blocks.0.attention.v_proj.bias - decoder.pool.blocks.0.attention.out_proj.weight - decoder.pool.blocks.0.attention.out_proj.bias - decoder.pool.blocks.0.ffn.0.weight - decoder.pool.blocks.0.ffn.0.bias - decoder.pool.blocks.0.ffn.3.weight - decoder.pool.blocks.0.ffn.3.bias - decoder.pool.norm.weight - decoder.pool.norm.bias - decoder.classification.weight - decoder.classification.bias - decoder.regression.weight - decoder.regression.bias continuation: before: u: rmse: 0.012642827012693 baseline_rmse: 1.2908974940031006 nrmse: 0.009793827218214922 v: rmse: 0.020391549975120982 baseline_rmse: 1.2969480983896875 nrmse: 0.015722718588692543 after: u: rmse: 0.013533551226681172 baseline_rmse: 1.2908974940031006 nrmse: 0.010483831047431459 v: rmse: 0.02074744998911489 baseline_rmse: 1.2969480983896875 nrmse: 0.015997132047824634 initial_old_prediction_drift: u: 0.0 v: 0.0 steps: 256 optimizer_covers_current_parameters: true validation_curve: - step: 0 scores: u: rmse: 0.012602725270002058 baseline_rmse: 1.2520463239225392 nrmse: 0.010065702066453058 v: rmse: 0.020709362689774323 baseline_rmse: 1.3128148986756512 nrmse: 0.015774777320599905 - step: 32 scores: u: rmse: 0.020870263408642533 baseline_rmse: 1.2520463239225392 nrmse: 0.01666892271466285 v: rmse: 0.07169098957546331 baseline_rmse: 1.3128148986756512 nrmse: 0.05460860449388878 - step: 128 scores: u: rmse: 0.01757523696940431 baseline_rmse: 1.2520463239225392 nrmse: 0.014037209832894046 v: rmse: 0.008292946113259245 baseline_rmse: 1.3128148986756512 nrmse: 0.00631691956088025 - step: 256 scores: u: rmse: 0.01312904452612221 baseline_rmse: 1.2520463239225392 nrmse: 0.010486069305319465 v: rmse: 0.020331855190023305 baseline_rmse: 1.3128148986756512 nrmse: 0.015487221550070608 scratch_adaptation: before: u: rmse: 1.2192609324518888 baseline_rmse: 1.2908974940031006 nrmse: 0.9445063904113213 v: rmse: 1.5097844652449826 baseline_rmse: 1.2969480983896875 nrmse: 1.164105539087922 w: rmse: 2.1888246131338143 baseline_rmse: 1.8516566868110593 nrmse: 1.1820898705058704 after: u: rmse: 0.028828906762000617 baseline_rmse: 1.2908974940031006 nrmse: 0.02233245234104651 v: rmse: 0.013749167456600677 baseline_rmse: 1.2969480983896875 nrmse: 0.010601170142175985 w: rmse: 0.04253773112874039 baseline_rmse: 1.8516566868110593 nrmse: 0.02297279589230943 initial_old_prediction_drift: u: 2.8637255281209946 v: 3.57562917470932 steps: 256 optimizer_covers_current_parameters: true validation_curve: - step: 0 scores: u: rmse: 1.1908904023705609 baseline_rmse: 1.2520463239225392 nrmse: 0.9511552245444219 v: rmse: 1.5197112873402683 baseline_rmse: 1.3128148986756512 nrmse: 1.1575975325031207 w: rmse: 2.086514793810442 baseline_rmse: 1.7725715949102796 nrmse: 1.1771117171242118 - step: 32 scores: u: rmse: 0.09774674429524874 baseline_rmse: 1.2520463239225392 nrmse: 0.07806959089901538 v: rmse: 0.09318242030853932 baseline_rmse: 1.3128148986756512 nrmse: 0.07097910025437738 w: rmse: 0.1725976821797815 baseline_rmse: 1.7725715949102796 nrmse: 0.09737134605754398 - step: 128 scores: u: rmse: 0.01929405457029216 baseline_rmse: 1.2520463239225392 nrmse: 0.015410016547826892 v: rmse: 0.027658242828711484 baseline_rmse: 1.3128148986756512 nrmse: 0.021067892249404484 w: rmse: 0.0574078117716002 baseline_rmse: 1.7725715949102796 nrmse: 0.032386737966714374 - step: 256 scores: u: rmse: 0.029240613188751634 baseline_rmse: 1.2520463239225392 nrmse: 0.023354258249122635 v: rmse: 0.012891726881292757 baseline_rmse: 1.3128148986756512 nrmse: 0.009819912079225904 w: rmse: 0.04314531080519493 baseline_rmse: 1.7725715949102796 nrmse: 0.02434051799604674 scratch_total: before: u: rmse: 1.4380359347862814 baseline_rmse: 1.2908974940031006 nrmse: 1.113981506251826 v: rmse: 1.6409674251416102 baseline_rmse: 1.2969480983896875 nrmse: 1.265252963614398 w: rmse: 1.872982479969112 baseline_rmse: 1.8516566868110593 nrmse: 1.0115171420868412 after: u: rmse: 0.021905167550788005 baseline_rmse: 1.2908974940031006 nrmse: 0.016968944205522946 v: rmse: 0.00931536736181866 baseline_rmse: 1.2969480983896875 nrmse: 0.00718252902593772 w: rmse: 0.024011370962885997 baseline_rmse: 1.8516566868110593 nrmse: 0.012967506954131226 initial_old_prediction_drift: u: 3.3363696336746216 v: 3.764289975166321 steps: 768 optimizer_covers_current_parameters: true validation_curve: - step: 0 scores: u: rmse: 1.3889139176866374 baseline_rmse: 1.2520463239225392 nrmse: 1.1093151196956559 v: rmse: 1.6675407357134648 baseline_rmse: 1.3128148986756512 nrmse: 1.270202476674858 w: rmse: 1.8040810343170564 baseline_rmse: 1.7725715949102796 nrmse: 1.0177761166303534 - step: 32 scores: u: rmse: 0.08006947008944894 baseline_rmse: 1.2520463239225392 nrmse: 0.06395088469138992 v: rmse: 0.07946685057847 baseline_rmse: 1.3128148986756512 nrmse: 0.06053164894657657 w: rmse: 0.19453834067577835 baseline_rmse: 1.7725715949102796 nrmse: 0.10974921477607515 - step: 128 scores: u: rmse: 0.039602594093698616 baseline_rmse: 1.2520463239225392 nrmse: 0.031630294612125485 v: rmse: 0.06683773516488707 baseline_rmse: 1.3128148986756512 nrmse: 0.05091177380170808 w: rmse: 0.06675999568144961 baseline_rmse: 1.7725715949102796 nrmse: 0.03766279222410124 - step: 768 scores: u: rmse: 0.021776166860430823 baseline_rmse: 1.2520463239225392 nrmse: 0.017392461001129903 v: rmse: 0.00983754729978949 baseline_rmse: 1.3128148986756512 nrmse: 0.00749347627735903 w: rmse: 0.02306899778914374 baseline_rmse: 1.7725715949102796 nrmse: 0.01301442370812187 test_rows: 1024 baseline: Source-training mean for test comparisons; phase-training mean for validation curves optimizer_policy: New AdamW factory and Trainer for every fit; all available labels rehearsed mutation: extend record with hidden Number record/w checks: Source learns both original tasks below 0.25 nRMSE: true Extension preserves every existing state entry: true Extension immediately preserves old predictions: false Diagnostic source pooling scale restores old predictions: true 'extended: fresh optimizer covers current parameters': true 'extended: all final targets below 0.25 nRMSE': true Added head parameters actually learn: true Hidden target placeholders cannot affect predictions: true Extended checkpoint preserves schema and learned state: true Extended checkpoint preserves predictions: true 'continuation: fresh optimizer covers current parameters': true 'scratch_adaptation: fresh optimizer covers current parameters': true 'scratch_total: fresh optimizer covers current parameters': true 'scratch_total: all final targets below 0.25 nRMSE': true outcome: not_met duration_seconds: 64.508 - started_at: '2026-09-15T18:02:25.640505+00:00' mode: full seed: 7102 steps_override: null accelerator: gpu threads: 1 provenance: revision: null source_sha256: 04f09ccb89861dbe7977f303d13c1e11bf180f1706f6ca837c89f4ba1050cf3b script_sha256: 61b3e662a7ec74440893deec64e6fb16806b8409e0f466c4c9feb98be4e8d144 code_sha256: 51ba353c87e01417438f385202d12cd43919a7e1f61896c4ab0391420ba86ace python: 3.12.6 platform: Linux-6.8.0-138-generic-x86_64-with-glibc2.35 host: lab relflow_path: /home/grantham/relflow-mutations-v3_tgit_/src/relflow packages: relflow: 0.2.2 torch: 2.12.0 lightning: 2.6.4 numpy: 2.4.6 pyarrow: 24.0.0 polars: 1.44.1 torchmetrics: 1.9.0 tensordict: 0.12.4 pydantic: 2.13.4 cuda: '13.0' gpu: NVIDIA GeForce RTX 3090 metrics: source: u: rmse: 0.03619475777939533 baseline_rmse: 1.2967289998405087 nrmse: 0.027912353146915897 v: rmse: 0.04001551584952593 baseline_rmse: 1.2693145185962316 nrmse: 0.03152529594775308 source_fit: steps: 512 optimizer_covers_current_parameters: true validation_curve: - step: 0 scores: u: rmse: 1.3057816740747938 baseline_rmse: 1.2723746695686682 nrmse: 1.0262556346845937 v: rmse: 1.7709300687901668 baseline_rmse: 1.3468533739051283 nrmse: 1.314864782686367 - step: 32 scores: u: rmse: 0.09028085487952377 baseline_rmse: 1.2723746695686682 nrmse: 0.0709546150507136 v: rmse: 0.11234495794035862 baseline_rmse: 1.3468533739051283 nrmse: 0.08341290901965112 - step: 128 scores: u: rmse: 0.04704011706734231 baseline_rmse: 1.2723746695686682 nrmse: 0.036970334440317644 v: rmse: 0.024967212372297355 baseline_rmse: 1.3468533739051283 nrmse: 0.018537439082850035 - step: 512 scores: u: rmse: 0.03577179115478007 baseline_rmse: 1.2723746695686682 nrmse: 0.028114196242924747 v: rmse: 0.038996974405307465 baseline_rmse: 1.3468533739051283 nrmse: 0.02895413499409951 source_prerequisite_met: true downstream_interpretable: true arms: extended: before: u: rmse: 0.034872039702596086 baseline_rmse: 1.2967289998405087 nrmse: 0.02689231112043085 v: rmse: 0.03595434348491558 baseline_rmse: 1.2693145185962316 nrmse: 0.028325795504709452 w: rmse: 1.7387411370781507 baseline_rmse: 1.781881033784205 nrmse: 0.9757896874773746 after: u: rmse: 0.02617816181915077 baseline_rmse: 1.2967289998405087 nrmse: 0.02018784327517203 v: rmse: 0.034476748667049514 baseline_rmse: 1.2693145185962316 nrmse: 0.027161706702274437 w: rmse: 0.020370154209505255 baseline_rmse: 1.781881033784205 nrmse: 0.011431826156342705 initial_old_prediction_drift: u: 0.02763199806213379 v: 0.02259039878845215 steps: 256 optimizer_covers_current_parameters: true validation_curve: - step: 0 scores: u: rmse: 0.03497645538661746 baseline_rmse: 1.2761651462099894 nrmse: 0.02740746798366344 v: rmse: 0.0359159228346385 baseline_rmse: 1.3343993130174792 nrmse: 0.02691542365487417 w: rmse: 1.7619841261571572 baseline_rmse: 1.7848255975425062 nrmse: 0.9872024071053221 - step: 32 scores: u: rmse: 0.07851660846800294 baseline_rmse: 1.2761651462099894 nrmse: 0.061525429291957215 v: rmse: 0.04250877017546018 baseline_rmse: 1.3343993130174792 nrmse: 0.031856109157711594 w: rmse: 0.12655087870319198 baseline_rmse: 1.7848255975425062 nrmse: 0.0709037784293533 - step: 128 scores: u: rmse: 0.016569259863831672 baseline_rmse: 1.2761651462099894 nrmse: 0.012983632967127944 v: rmse: 0.02356812906567172 baseline_rmse: 1.3343993130174792 nrmse: 0.017661976318300913 w: rmse: 0.04250871038825767 baseline_rmse: 1.7848255975425062 nrmse: 0.023816730579607964 - step: 256 scores: u: rmse: 0.026681915213499285 baseline_rmse: 1.2761651462099894 nrmse: 0.020907885858456793 v: rmse: 0.03380502289581418 baseline_rmse: 1.3343993130174792 nrmse: 0.025333513413890202 w: rmse: 0.02048024492089371 baseline_rmse: 1.7848255975425062 nrmse: 0.011474647690560122 changed_existing_state_entries: [] pooling_scale_control: source_capacity: 4 extended_capacity: 5 old_prediction_drift_with_source_capacity: u: 0.0 v: 0.0 new_head_updated_parameters: - embedder.embeddings.weight - embedder.linear.weight - embedder.scalar.weight - decoder.context_projection.0.weight - decoder.context_projection.0.bias - decoder.context_projection.1.weight - decoder.context_projection.1.bias - decoder.context_projection.3.weight - decoder.context_projection.3.bias - decoder.pool.queries - decoder.pool.blocks.0.attention_norm.weight - decoder.pool.blocks.0.attention_norm.bias - decoder.pool.blocks.0.ffn_norm.weight - decoder.pool.blocks.0.ffn_norm.bias - decoder.pool.blocks.0.attention.q_proj.weight - decoder.pool.blocks.0.attention.q_proj.bias - decoder.pool.blocks.0.attention.k_proj.weight - decoder.pool.blocks.0.attention.k_proj.bias - decoder.pool.blocks.0.attention.v_proj.weight - decoder.pool.blocks.0.attention.v_proj.bias - decoder.pool.blocks.0.attention.out_proj.weight - decoder.pool.blocks.0.attention.out_proj.bias - decoder.pool.blocks.0.ffn.0.weight - decoder.pool.blocks.0.ffn.0.bias - decoder.pool.blocks.0.ffn.3.weight - decoder.pool.blocks.0.ffn.3.bias - decoder.pool.norm.weight - decoder.pool.norm.bias - decoder.classification.weight - decoder.classification.bias - decoder.regression.weight - decoder.regression.bias continuation: before: u: rmse: 0.03619475777939533 baseline_rmse: 1.2967289998405087 nrmse: 0.027912353146915897 v: rmse: 0.04001551584952593 baseline_rmse: 1.2693145185962316 nrmse: 0.03152529594775308 after: u: rmse: 0.01113863220749229 baseline_rmse: 1.2967289998405087 nrmse: 0.008589791860028029 v: rmse: 0.02652724758703262 baseline_rmse: 1.2693145185962316 nrmse: 0.02089887667586896 initial_old_prediction_drift: u: 0.0 v: 0.0 steps: 256 optimizer_covers_current_parameters: true validation_curve: - step: 0 scores: u: rmse: 0.036313625699826456 baseline_rmse: 1.2761651462099894 nrmse: 0.028455271488703665 v: rmse: 0.0399088880002447 baseline_rmse: 1.3343993130174792 nrmse: 0.029907755205597842 - step: 32 scores: u: rmse: 0.12632528095574588 baseline_rmse: 1.2761651462099894 nrmse: 0.09898819234400201 v: rmse: 0.08776433975828732 baseline_rmse: 1.3343993130174792 nrmse: 0.06577067216845735 - step: 128 scores: u: rmse: 0.013455144390389199 baseline_rmse: 1.2761651462099894 nrmse: 0.010543419423692043 v: rmse: 0.032192121521092104 baseline_rmse: 1.3343993130174792 nrmse: 0.02412480372782568 - step: 256 scores: u: rmse: 0.011103581102286394 baseline_rmse: 1.2761651462099894 nrmse: 0.008700739974965066 v: rmse: 0.026375210338510726 baseline_rmse: 1.3343993130174792 nrmse: 0.01976560545348935 scratch_adaptation: before: u: rmse: 1.4458342313924608 baseline_rmse: 1.2967289998405087 nrmse: 1.114985653571634 v: rmse: 1.6118714028737393 baseline_rmse: 1.2693145185962316 nrmse: 1.2698754952053573 w: rmse: 1.8029852957423909 baseline_rmse: 1.781881033784205 nrmse: 1.0118438108706767 after: u: rmse: 0.0301686763189551 baseline_rmse: 1.2967289998405087 nrmse: 0.023265212949402456 v: rmse: 0.06359981585245611 baseline_rmse: 1.2693145185962316 nrmse: 0.05010563963515743 w: rmse: 0.02979031595210785 baseline_rmse: 1.781881033784205 nrmse: 0.01671846514289551 initial_old_prediction_drift: u: 3.4850218892097473 v: 3.782611846923828 steps: 256 optimizer_covers_current_parameters: true validation_curve: - step: 0 scores: u: rmse: 1.4203856895664793 baseline_rmse: 1.2761651462099894 nrmse: 1.1130108777729921 v: rmse: 1.6633459686330494 baseline_rmse: 1.3343993130174792 nrmse: 1.2465129083975042 w: rmse: 1.8216507481098305 baseline_rmse: 1.7848255975425062 nrmse: 1.0206323523250833 - step: 32 scores: u: rmse: 0.10742258146671714 baseline_rmse: 1.2761651462099894 nrmse: 0.08417608158767334 v: rmse: 0.11297330867112705 baseline_rmse: 1.3343993130174792 nrmse: 0.08466229528825246 w: rmse: 0.1840070516908891 baseline_rmse: 1.7848255975425062 nrmse: 0.10309525588620258 - step: 128 scores: u: rmse: 0.015407758053566608 baseline_rmse: 1.2761651462099894 nrmse: 0.012073482886854602 v: rmse: 0.021865477122599692 baseline_rmse: 1.3343993130174792 nrmse: 0.016386007478642398 w: rmse: 0.04819928675177008 baseline_rmse: 1.7848255975425062 nrmse: 0.027005040054409124 - step: 256 scores: u: rmse: 0.030187696362557213 baseline_rmse: 1.2761651462099894 nrmse: 0.023655007701949816 v: rmse: 0.06318318408346218 baseline_rmse: 1.3343993130174792 nrmse: 0.04734953283255666 w: rmse: 0.029402092898951477 baseline_rmse: 1.7848255975425062 nrmse: 0.016473370249415226 scratch_total: before: u: rmse: 1.4600618508708036 baseline_rmse: 1.2967289998405087 nrmse: 1.125957583311844 v: rmse: 1.1369067410066136 baseline_rmse: 1.2693145185962316 nrmse: 0.8956856038044446 w: rmse: 1.9207764498672377 baseline_rmse: 1.781881033784205 nrmse: 1.077948759456774 after: u: rmse: 0.021576266176060177 baseline_rmse: 1.2967289998405087 nrmse: 0.016638994098777733 v: rmse: 0.011082497616507384 baseline_rmse: 1.2693145185962316 nrmse: 0.00873108867356517 w: rmse: 0.03203390226617649 baseline_rmse: 1.781881033784205 nrmse: 0.017977576313355583 initial_old_prediction_drift: u: 3.5209039449691772 v: 2.71368208527565 steps: 768 optimizer_covers_current_parameters: true validation_curve: - step: 0 scores: u: rmse: 1.4458387934119932 baseline_rmse: 1.2761651462099894 nrmse: 1.132955869940429 v: rmse: 1.1961971300684766 baseline_rmse: 1.3343993130174792 nrmse: 0.8964311644941679 w: rmse: 1.9311233553703957 baseline_rmse: 1.7848255975425062 nrmse: 1.0819675367886499 - step: 32 scores: u: rmse: 0.07357051740094268 baseline_rmse: 1.2761651462099894 nrmse: 0.0576496839922604 v: rmse: 0.07911851292948417 baseline_rmse: 1.3343993130174792 nrmse: 0.059291482060623485 w: rmse: 0.16387433636776993 baseline_rmse: 1.7848255975425062 nrmse: 0.09181532167255194 - step: 128 scores: u: rmse: 0.01636567853368862 baseline_rmse: 1.2761651462099894 nrmse: 0.01282410711677256 v: rmse: 0.04020440315094177 baseline_rmse: 1.3343993130174792 nrmse: 0.030129214515276913 w: rmse: 0.032668450699060055 baseline_rmse: 1.7848255975425062 nrmse: 0.018303441380513954 - step: 768 scores: u: rmse: 0.021918321358026984 baseline_rmse: 1.2761651462099894 nrmse: 0.017175144943521584 v: rmse: 0.011705376725132962 baseline_rmse: 1.3343993130174792 nrmse: 0.008772019447959379 w: rmse: 0.032805710450395445 baseline_rmse: 1.7848255975425062 nrmse: 0.018380345113587025 test_rows: 1024 baseline: Source-training mean for test comparisons; phase-training mean for validation curves optimizer_policy: New AdamW factory and Trainer for every fit; all available labels rehearsed mutation: extend record with hidden Number record/w checks: Source learns both original tasks below 0.25 nRMSE: true Extension preserves every existing state entry: true Extension immediately preserves old predictions: false Diagnostic source pooling scale restores old predictions: true 'extended: fresh optimizer covers current parameters': true 'extended: all final targets below 0.25 nRMSE': true Added head parameters actually learn: true Hidden target placeholders cannot affect predictions: true Extended checkpoint preserves schema and learned state: true Extended checkpoint preserves predictions: true 'continuation: fresh optimizer covers current parameters': true 'scratch_adaptation: fresh optimizer covers current parameters': true 'scratch_total: fresh optimizer covers current parameters': true 'scratch_total: all final targets below 0.25 nRMSE': true outcome: not_met duration_seconds: 77.691 - started_at: '2026-09-15T18:02:41.420090+00:00' mode: full seed: 4701 steps_override: null accelerator: gpu threads: 1 provenance: revision: null source_sha256: 2db77891e89376c6926a1f11a18aa46fb4f8bebed0df14438ba8eff89c4db6c1 script_sha256: 624c767c1b4024e120b601bc1efaadbe9a478f8c3ab6d8b0e375600c396a001e code_sha256: 5f613c009ea56a099f0f766dcac921baef4eabf53cf14a0a335ebfb2b4ca271d python: 3.12.6 platform: Linux-6.8.0-138-generic-x86_64-with-glibc2.35 host: lab relflow_path: /home/grantham/relflow-mutations-v3_tgit_/src/relflow packages: relflow: 0.2.2 torch: 2.12.0 lightning: 2.6.4 numpy: 2.4.6 pyarrow: 24.0.0 polars: 1.44.1 torchmetrics: 1.9.0 tensordict: 0.12.4 pydantic: 2.13.4 cuda: '13.0' gpu: NVIDIA GeForce RTX 3090 metrics: source: u: rmse: 0.011876638061833422 baseline_rmse: 1.2663362526593396 nrmse: 0.009378739680627612 v: rmse: 0.029771526388557422 baseline_rmse: 1.2961683303676375 nrmse: 0.02296887347966078 source_fit: steps: 512 optimizer_covers_current_parameters: true validation_curve: - step: 0 scores: u: rmse: 1.1675138300279053 baseline_rmse: 1.295883063514218 nrmse: 0.9009407275235184 v: rmse: 1.6476150162913918 baseline_rmse: 1.2898120364298942 nrmse: 1.2774070715387882 - step: 32 scores: u: rmse: 0.09449274849298497 baseline_rmse: 1.295883063514218 nrmse: 0.07291765063796456 v: rmse: 0.10256670713688135 baseline_rmse: 1.2898120364298942 nrmse: 0.07952066211196053 - step: 128 scores: u: rmse: 0.07972301215306196 baseline_rmse: 1.295883063514218 nrmse: 0.06152022076503299 v: rmse: 0.08407132185296792 baseline_rmse: 1.2898120364298942 nrmse: 0.06518106474310104 - step: 512 scores: u: rmse: 0.0117645243818427 baseline_rmse: 1.295883063514218 nrmse: 0.009078384240889205 v: rmse: 0.028888381892659666 baseline_rmse: 1.2898120364298942 nrmse: 0.022397357969011208 source_prerequisite_met: true downstream_interpretable: true arms: extended: before: u: rmse: 0.018670282761765817 baseline_rmse: 1.2663362526593396 nrmse: 0.014743542816971187 v: rmse: 0.03499104816395304 baseline_rmse: 1.2961683303676375 nrmse: 0.026995759226757524 w: rmse: 1.803691733248121 baseline_rmse: 1.8289556685792812 nrmse: 0.9861866879743537 after: u: rmse: 0.03541920265964375 baseline_rmse: 1.2663362526593396 nrmse: 0.02796982443269905 v: rmse: 0.05258085584809817 baseline_rmse: 1.2961683303676375 nrmse: 0.040566379085334114 w: rmse: 0.03525589383133793 baseline_rmse: 1.8289556685792812 nrmse: 0.019276516340455886 initial_old_prediction_drift: u: 0.030529499053955078 v: 0.027675628662109375 steps: 256 optimizer_covers_current_parameters: true validation_curve: - step: 0 scores: u: rmse: 0.01872948207341073 baseline_rmse: 1.3104433241303899 nrmse: 0.014292477765751231 v: rmse: 0.034707762389102226 baseline_rmse: 1.2705869192092003 nrmse: 0.027316322767358542 w: rmse: 1.8227888101287097 baseline_rmse: 1.8209322963213432 nrmse: 1.0010195402712758 - step: 32 scores: u: rmse: 0.08361337667316292 baseline_rmse: 1.3104433241303899 nrmse: 0.06380541236199494 v: rmse: 0.07764783231496324 baseline_rmse: 1.2705869192092003 nrmse: 0.06111178317756523 w: rmse: 0.10993914495916435 baseline_rmse: 1.8209322963213432 nrmse: 0.06037519636576493 - step: 128 scores: u: rmse: 0.02387832804820865 baseline_rmse: 1.3104433241303899 nrmse: 0.01822156487695056 v: rmse: 0.02310094682784648 baseline_rmse: 1.2705869192092003 nrmse: 0.018181319576486953 w: rmse: 0.056417158157502645 baseline_rmse: 1.8209322963213432 nrmse: 0.03098256770527761 - step: 256 scores: u: rmse: 0.03487140440785663 baseline_rmse: 1.3104433241303899 nrmse: 0.026610387313772074 v: rmse: 0.05276883469692369 baseline_rmse: 1.2705869192092003 nrmse: 0.04153107032596121 w: rmse: 0.03631950546289549 baseline_rmse: 1.8209322963213432 nrmse: 0.01994555510727573 changed_existing_state_entries: [] pooling_scale_control: source_capacity: 4 extended_capacity: 5 old_prediction_drift_with_source_capacity: u: 0.0 v: 0.0 new_head_updated_parameters: - embedder.embeddings.weight - embedder.linear.weight - embedder.scalar.weight - decoder.context_projection.0.weight - decoder.context_projection.0.bias - decoder.context_projection.1.weight - decoder.context_projection.1.bias - decoder.context_projection.3.weight - decoder.context_projection.3.bias - decoder.pool.queries - decoder.pool.blocks.0.attention_norm.weight - decoder.pool.blocks.0.attention_norm.bias - decoder.pool.blocks.0.ffn_norm.weight - decoder.pool.blocks.0.ffn_norm.bias - decoder.pool.blocks.0.attention.q_proj.weight - decoder.pool.blocks.0.attention.q_proj.bias - decoder.pool.blocks.0.attention.k_proj.weight - decoder.pool.blocks.0.attention.k_proj.bias - decoder.pool.blocks.0.attention.v_proj.weight - decoder.pool.blocks.0.attention.v_proj.bias - decoder.pool.blocks.0.attention.out_proj.weight - decoder.pool.blocks.0.attention.out_proj.bias - decoder.pool.blocks.0.ffn.0.weight - decoder.pool.blocks.0.ffn.0.bias - decoder.pool.blocks.0.ffn.3.weight - decoder.pool.blocks.0.ffn.3.bias - decoder.pool.norm.weight - decoder.pool.norm.bias - decoder.classification.weight - decoder.classification.bias - decoder.regression.weight - decoder.regression.bias continuation: before: u: rmse: 0.011876638061833422 baseline_rmse: 1.2663362526593396 nrmse: 0.009378739680627612 v: rmse: 0.029771526388557422 baseline_rmse: 1.2961683303676375 nrmse: 0.02296887347966078 after: u: rmse: 0.009253100199328049 baseline_rmse: 1.2663362526593396 nrmse: 0.007306985154927291 v: rmse: 0.016477268987651207 baseline_rmse: 1.2961683303676375 nrmse: 0.012712290989996409 initial_old_prediction_drift: u: 0.0 v: 0.0 steps: 256 optimizer_covers_current_parameters: true validation_curve: - step: 0 scores: u: rmse: 0.011535892794200158 baseline_rmse: 1.3104433241303899 nrmse: 0.008803045947718018 v: rmse: 0.029509003359319566 baseline_rmse: 1.2705869192092003 nrmse: 0.023224702626157723 - step: 32 scores: u: rmse: 0.08205461194764767 baseline_rmse: 1.3104433241303899 nrmse: 0.06261591816807423 v: rmse: 0.081501683033155 baseline_rmse: 1.2705869192092003 nrmse: 0.06414490956972922 - step: 128 scores: u: rmse: 0.024969262126465445 baseline_rmse: 1.3104433241303899 nrmse: 0.019054057254276942 v: rmse: 0.008535274276989115 baseline_rmse: 1.2705869192092003 nrmse: 0.006717583935384269 - step: 256 scores: u: rmse: 0.009532766297869321 baseline_rmse: 1.3104433241303899 nrmse: 0.007274459049341386 v: rmse: 0.016615988523582266 baseline_rmse: 1.2705869192092003 nrmse: 0.013077411920723913 scratch_adaptation: before: u: rmse: 1.2522114938581133 baseline_rmse: 1.2663362526593396 nrmse: 0.9888459650653103 v: rmse: 1.3195483233826368 baseline_rmse: 1.2961683303676375 nrmse: 1.0180377752389367 w: rmse: 1.8805651915817234 baseline_rmse: 1.8289556685792812 nrmse: 1.0282180284022586 after: u: rmse: 0.026914617420797867 baseline_rmse: 1.2663362526593396 nrmse: 0.021253926328237433 v: rmse: 0.05481134055653444 baseline_rmse: 1.2961683303676375 nrmse: 0.04228720859194891 w: rmse: 0.03336312800981138 baseline_rmse: 1.8289556685792812 nrmse: 0.01824162749429985 initial_old_prediction_drift: u: 2.8913912177085876 v: 3.190234363079071 steps: 256 optimizer_covers_current_parameters: true validation_curve: - step: 0 scores: u: rmse: 1.3100973600945236 baseline_rmse: 1.3104433241303899 nrmse: 0.9997359946595967 v: rmse: 1.3005482680231673 baseline_rmse: 1.2705869192092003 nrmse: 1.0235807156212615 w: rmse: 1.8851250617203843 baseline_rmse: 1.8209322963213432 nrmse: 1.0352526920021814 - step: 32 scores: u: rmse: 0.08970210480598517 baseline_rmse: 1.3104433241303899 nrmse: 0.06845172404957801 v: rmse: 0.08707615719809386 baseline_rmse: 1.2705869192092003 nrmse: 0.06853223174396375 w: rmse: 0.13807694966229056 baseline_rmse: 1.8209322963213432 nrmse: 0.07582761310853474 - step: 128 scores: u: rmse: 0.02000263508140492 baseline_rmse: 1.3104433241303899 nrmse: 0.015264021505606636 v: rmse: 0.043675904542061904 baseline_rmse: 1.2705869192092003 nrmse: 0.03437459010615765 w: rmse: 0.04264531821229424 baseline_rmse: 1.8209322963213432 nrmse: 0.023419496868964614 - step: 256 scores: u: rmse: 0.027012699849219925 baseline_rmse: 1.3104433241303899 nrmse: 0.020613405671050713 v: rmse: 0.055833388807019234 baseline_rmse: 1.2705869192092003 nrmse: 0.04394299041089557 w: rmse: 0.03318094317825782 baseline_rmse: 1.8209322963213432 nrmse: 0.018221953251798616 scratch_total: before: u: rmse: 1.2108697607158914 baseline_rmse: 1.2663362526593396 nrmse: 0.9561992386880127 v: rmse: 1.4984804880233138 baseline_rmse: 1.2961683303676375 nrmse: 1.1560847869182962 w: rmse: 1.781162022536913 baseline_rmse: 1.8289556685792812 nrmse: 0.9738683408988836 after: u: rmse: 0.015303161159544906 baseline_rmse: 1.2663362526593396 nrmse: 0.012084595325615817 v: rmse: 0.02876708367985974 baseline_rmse: 1.2961683303676375 nrmse: 0.022193941177146657 w: rmse: 0.028018792189193974 baseline_rmse: 1.8289556685792812 nrmse: 0.01531955786055698 initial_old_prediction_drift: u: 2.768245704472065 v: 3.4774352312088013 steps: 768 optimizer_covers_current_parameters: true validation_curve: - step: 0 scores: u: rmse: 1.2564398692105467 baseline_rmse: 1.3104433241303899 nrmse: 0.9587899347301572 v: rmse: 1.4780311568075868 baseline_rmse: 1.2705869192092003 nrmse: 1.1632664672224846 w: rmse: 1.778345330480088 baseline_rmse: 1.8209322963213432 nrmse: 0.9766125484581225 - step: 32 scores: u: rmse: 0.09396991963840232 baseline_rmse: 1.3104433241303899 nrmse: 0.07170849582583874 v: rmse: 0.0915242341058268 baseline_rmse: 1.2705869192092003 nrmse: 0.07203303664009897 w: rmse: 0.16457404207107804 baseline_rmse: 1.8209322963213432 nrmse: 0.09037900113230533 - step: 128 scores: u: rmse: 0.03882763985005823 baseline_rmse: 1.3104433241303899 nrmse: 0.029629392691075937 v: rmse: 0.06739140155497268 baseline_rmse: 1.2705869192092003 nrmse: 0.053039583940401626 w: rmse: 0.05676065906718735 baseline_rmse: 1.8209322963213432 nrmse: 0.031171207837795797 - step: 768 scores: u: rmse: 0.016408114695598202 baseline_rmse: 1.3104433241303899 nrmse: 0.01252104108087744 v: rmse: 0.028186368231612118 baseline_rmse: 1.2705869192092003 nrmse: 0.022183738715927447 w: rmse: 0.027628165942140175 baseline_rmse: 1.8209322963213432 nrmse: 0.01517253881319736 test_rows: 1024 baseline: Source-training mean for test comparisons; phase-training mean for validation curves optimizer_policy: New AdamW factory and Trainer for every fit; all available labels rehearsed mutation: extend record with hidden Number record/w checks: Source learns both original tasks below 0.25 nRMSE: true Source optimizer covers current parameters: true Extension preserves every existing state entry: true Extension immediately preserves old predictions: false Diagnostic source pooling scale restores old predictions: true 'extended: fresh optimizer covers current parameters': true 'extended: all final targets below 0.25 nRMSE': true Added head parameters actually learn: true Hidden target placeholders cannot affect predictions: true Extended checkpoint preserves schema and learned state: true Extended checkpoint preserves predictions: true 'continuation: fresh optimizer covers current parameters': true 'scratch_adaptation: schema and field order match extension': true 'scratch_adaptation: fresh optimizer covers current parameters': true 'scratch_total: schema and field order match extension': true 'scratch_total: fresh optimizer covers current parameters': true 'scratch_total: all final targets below 0.25 nRMSE': true outcome: not_met duration_seconds: 80.407 - started_at: '2026-09-15T18:04:05.802381+00:00' mode: full seed: 7101 steps_override: null accelerator: gpu threads: 1 provenance: revision: null source_sha256: 2db77891e89376c6926a1f11a18aa46fb4f8bebed0df14438ba8eff89c4db6c1 script_sha256: 624c767c1b4024e120b601bc1efaadbe9a478f8c3ab6d8b0e375600c396a001e code_sha256: 5f613c009ea56a099f0f766dcac921baef4eabf53cf14a0a335ebfb2b4ca271d python: 3.12.6 platform: Linux-6.8.0-138-generic-x86_64-with-glibc2.35 host: lab relflow_path: /home/grantham/relflow-mutations-v3_tgit_/src/relflow packages: relflow: 0.2.2 torch: 2.12.0 lightning: 2.6.4 numpy: 2.4.6 pyarrow: 24.0.0 polars: 1.44.1 torchmetrics: 1.9.0 tensordict: 0.12.4 pydantic: 2.13.4 cuda: '13.0' gpu: NVIDIA GeForce RTX 3090 metrics: source: u: rmse: 0.009796106068409562 baseline_rmse: 1.2908974940031006 nrmse: 0.007588601042234289 v: rmse: 0.019397250921355053 baseline_rmse: 1.2969480983896875 nrmse: 0.01495607337366777 source_fit: steps: 512 optimizer_covers_current_parameters: true validation_curve: - step: 0 scores: u: rmse: 1.5771062482408669 baseline_rmse: 1.214202473192484 nrmse: 1.2988824212276602 v: rmse: 1.3072694637116182 baseline_rmse: 1.3285155523136438 nrmse: 0.9840076478103512 - step: 32 scores: u: rmse: 0.10717741666818707 baseline_rmse: 1.214202473192484 nrmse: 0.08826980592980273 v: rmse: 0.10111543375960332 baseline_rmse: 1.3285155523136438 nrmse: 0.07611159205739684 - step: 128 scores: u: rmse: 0.017833300961804754 baseline_rmse: 1.214202473192484 nrmse: 0.014687254684069229 v: rmse: 0.022416444336445907 baseline_rmse: 1.3285155523136438 nrmse: 0.016873302158493437 - step: 512 scores: u: rmse: 0.009539652227979903 baseline_rmse: 1.214202473192484 nrmse: 0.007856722777789638 v: rmse: 0.01895260950991225 baseline_rmse: 1.3285155523136438 nrmse: 0.014266004998515672 source_prerequisite_met: true downstream_interpretable: true arms: extended: before: u: rmse: 0.017057897003102993 baseline_rmse: 1.2908974940031006 nrmse: 0.013213982583703134 v: rmse: 0.01556351569255927 baseline_rmse: 1.2969480983896875 nrmse: 0.012000106798323844 w: rmse: 1.8416579454327944 baseline_rmse: 1.8516566868110593 nrmse: 0.9946001105661304 after: u: rmse: 0.011567030862529346 baseline_rmse: 1.2908974940031006 nrmse: 0.008960456516698112 v: rmse: 0.0103254468559876 baseline_rmse: 1.2969480983896875 nrmse: 0.007961341605579937 w: rmse: 0.031046987648793805 baseline_rmse: 1.8516566868110593 nrmse: 0.01676714040455481 initial_old_prediction_drift: u: 0.03059530258178711 v: 0.021714448928833008 steps: 256 optimizer_covers_current_parameters: true validation_curve: - step: 0 scores: u: rmse: 0.016539311274679638 baseline_rmse: 1.2520463239225392 nrmse: 0.013209823757050447 v: rmse: 0.014701699835999216 baseline_rmse: 1.3128148986756512 nrmse: 0.011198608311674465 w: rmse: 1.7614574054999887 baseline_rmse: 1.7725715949102796 nrmse: 0.9937299066270701 - step: 32 scores: u: rmse: 0.0820270232479449 baseline_rmse: 1.2520463239225392 nrmse: 0.06551436770403368 v: rmse: 0.0743724926121628 baseline_rmse: 1.3128148986756512 nrmse: 0.056651164369926565 w: rmse: 0.12562222626077427 baseline_rmse: 1.7725715949102796 nrmse: 0.07087004362559063 - step: 128 scores: u: rmse: 0.02058476444392487 baseline_rmse: 1.2520463239225392 nrmse: 0.016440896834739156 v: rmse: 0.043512702291031054 baseline_rmse: 1.3128148986756512 nrmse: 0.03314458293772111 w: rmse: 0.05793298161340654 baseline_rmse: 1.7725715949102796 nrmse: 0.032683013639479466 - step: 256 scores: u: rmse: 0.010786105146310983 baseline_rmse: 1.2520463239225392 nrmse: 0.008614781210745595 v: rmse: 0.009854160995647818 baseline_rmse: 1.3128148986756512 nrmse: 0.007506131295119026 w: rmse: 0.031470882484469324 baseline_rmse: 1.7725715949102796 nrmse: 0.01775436466139595 changed_existing_state_entries: [] pooling_scale_control: source_capacity: 4 extended_capacity: 5 old_prediction_drift_with_source_capacity: u: 0.0 v: 0.0 new_head_updated_parameters: - embedder.embeddings.weight - embedder.linear.weight - embedder.scalar.weight - decoder.context_projection.0.weight - decoder.context_projection.0.bias - decoder.context_projection.1.weight - decoder.context_projection.1.bias - decoder.context_projection.3.weight - decoder.context_projection.3.bias - decoder.pool.queries - decoder.pool.blocks.0.attention_norm.weight - decoder.pool.blocks.0.attention_norm.bias - decoder.pool.blocks.0.ffn_norm.weight - decoder.pool.blocks.0.ffn_norm.bias - decoder.pool.blocks.0.attention.q_proj.weight - decoder.pool.blocks.0.attention.q_proj.bias - decoder.pool.blocks.0.attention.k_proj.weight - decoder.pool.blocks.0.attention.k_proj.bias - decoder.pool.blocks.0.attention.v_proj.weight - decoder.pool.blocks.0.attention.v_proj.bias - decoder.pool.blocks.0.attention.out_proj.weight - decoder.pool.blocks.0.attention.out_proj.bias - decoder.pool.blocks.0.ffn.0.weight - decoder.pool.blocks.0.ffn.0.bias - decoder.pool.blocks.0.ffn.3.weight - decoder.pool.blocks.0.ffn.3.bias - decoder.pool.norm.weight - decoder.pool.norm.bias - decoder.classification.weight - decoder.classification.bias - decoder.regression.weight - decoder.regression.bias continuation: before: u: rmse: 0.009796106068409562 baseline_rmse: 1.2908974940031006 nrmse: 0.007588601042234289 v: rmse: 0.019397250921355053 baseline_rmse: 1.2969480983896875 nrmse: 0.01495607337366777 after: u: rmse: 0.014894116016718029 baseline_rmse: 1.2908974940031006 nrmse: 0.011537799155943093 v: rmse: 0.03418484117031133 baseline_rmse: 1.2969480983896875 nrmse: 0.02635790993699424 initial_old_prediction_drift: u: 0.0 v: 0.0 steps: 256 optimizer_covers_current_parameters: true validation_curve: - step: 0 scores: u: rmse: 0.009405761385961659 baseline_rmse: 1.2520463239225392 nrmse: 0.007512311011380413 v: rmse: 0.01851867432477028 baseline_rmse: 1.3128148986756512 nrmse: 0.014106081781560868 - step: 32 scores: u: rmse: 0.08427936048393028 baseline_rmse: 1.2520463239225392 nrmse: 0.06731329254647005 v: rmse: 0.061933053385872454 baseline_rmse: 1.3128148986756512 nrmse: 0.04717576975120379 - step: 128 scores: u: rmse: 0.028769483507834138 baseline_rmse: 1.2520463239225392 nrmse: 0.02297797050967104 v: rmse: 0.052134352761181266 baseline_rmse: 1.3128148986756512 nrmse: 0.03971188384118252 - step: 256 scores: u: rmse: 0.01376603956676995 baseline_rmse: 1.2520463239225392 nrmse: 0.010994832462462162 v: rmse: 0.035074513265269745 baseline_rmse: 1.3128148986756512 nrmse: 0.026717028653965164 scratch_adaptation: before: u: rmse: 1.5767538624741606 baseline_rmse: 1.2908974940031006 nrmse: 1.2214400212247785 v: rmse: 1.3513676741288465 baseline_rmse: 1.2969480983896875 nrmse: 1.0419597174372106 w: rmse: 2.0629725509924435 baseline_rmse: 1.8516566868110593 nrmse: 1.114122593937926 after: u: rmse: 0.029171949457763285 baseline_rmse: 1.2908974940031006 nrmse: 0.022598192027858423 v: rmse: 0.02927820228914024 baseline_rmse: 1.2969480983896875 nrmse: 0.022574690787929407 w: rmse: 0.06450798959616429 baseline_rmse: 1.8516566868110593 nrmse: 0.03483798592667875 initial_old_prediction_drift: u: 3.5362414121627808 v: 3.128715068101883 steps: 256 optimizer_covers_current_parameters: true validation_curve: - step: 0 scores: u: rmse: 1.5552778950568296 baseline_rmse: 1.2520463239225392 nrmse: 1.2421887795527369 v: rmse: 1.3780781627025007 baseline_rmse: 1.3128148986756512 nrmse: 1.0497124644858054 w: rmse: 1.9900718363110987 baseline_rmse: 1.7725715949102796 nrmse: 1.1227032194498345 - step: 32 scores: u: rmse: 0.10571130232951326 baseline_rmse: 1.2520463239225392 nrmse: 0.0844308236122846 v: rmse: 0.07665508388739514 baseline_rmse: 1.3128148986756512 nrmse: 0.05838986437823313 w: rmse: 0.11916817650876253 baseline_rmse: 1.7725715949102796 nrmse: 0.06722897785959069 - step: 128 scores: u: rmse: 0.03689917851922985 baseline_rmse: 1.2520463239225392 nrmse: 0.029471096886917344 v: rmse: 0.019332327621935888 baseline_rmse: 1.3128148986756512 nrmse: 0.014725859404427893 w: rmse: 0.04996588304430775 baseline_rmse: 1.7725715949102796 nrmse: 0.028188358195391717 - step: 256 scores: u: rmse: 0.028474723085549214 baseline_rmse: 1.2520463239225392 nrmse: 0.022742547573112695 v: rmse: 0.030366150318473572 baseline_rmse: 1.3128148986756512 nrmse: 0.023130564978434135 w: rmse: 0.06598310054335318 baseline_rmse: 1.7725715949102796 nrmse: 0.03722450519505983 scratch_total: before: u: rmse: 1.611385024388229 baseline_rmse: 1.2908974940031006 nrmse: 1.2482672186397155 v: rmse: 1.1546936662757465 baseline_rmse: 1.2969480983896875 nrmse: 0.8903160178186263 w: rmse: 1.927721049616771 baseline_rmse: 1.8516566868110593 nrmse: 1.0410790852038077 after: u: rmse: 0.01769300810368644 baseline_rmse: 1.2908974940031006 nrmse: 0.013705974475804463 v: rmse: 0.013830831489853855 baseline_rmse: 1.2969480983896875 nrmse: 0.010664136450044876 w: rmse: 0.02184147382636228 baseline_rmse: 1.8516566868110593 nrmse: 0.011795638998273418 initial_old_prediction_drift: u: 3.6831963062286377 v: 2.9051989912986755 steps: 768 optimizer_covers_current_parameters: true validation_curve: - step: 0 scores: u: rmse: 1.596030310081841 baseline_rmse: 1.2520463239225392 nrmse: 1.274737427511175 v: rmse: 1.1701771178431117 baseline_rmse: 1.3128148986756512 nrmse: 0.8913496632492285 w: rmse: 1.8460251411143769 baseline_rmse: 1.7725715949102796 nrmse: 1.041438972854473 - step: 32 scores: u: rmse: 0.0783583153479978 baseline_rmse: 1.2520463239225392 nrmse: 0.06258419824476527 v: rmse: 0.0853941148445926 baseline_rmse: 1.3128148986756512 nrmse: 0.06504657658192099 w: rmse: 0.16184347906236152 baseline_rmse: 1.7725715949102796 nrmse: 0.0913043397102126 - step: 128 scores: u: rmse: 0.08590635687056956 baseline_rmse: 1.2520463239225392 nrmse: 0.06861276234687014 v: rmse: 0.04359809529439207 baseline_rmse: 1.3128148986756512 nrmse: 0.03320962866766153 w: rmse: 0.08713269901187601 baseline_rmse: 1.7725715949102796 nrmse: 0.0491560957323624 - step: 768 scores: u: rmse: 0.01772292845773104 baseline_rmse: 1.2520463239225392 nrmse: 0.014155169915923582 v: rmse: 0.014364053246255675 baseline_rmse: 1.3128148986756512 nrmse: 0.010941415473533951 w: rmse: 0.02199540328058544 baseline_rmse: 1.7725715949102796 nrmse: 0.012408753104101704 test_rows: 1024 baseline: Source-training mean for test comparisons; phase-training mean for validation curves optimizer_policy: New AdamW factory and Trainer for every fit; all available labels rehearsed mutation: extend record with hidden Number record/w checks: Source learns both original tasks below 0.25 nRMSE: true Source optimizer covers current parameters: true Extension preserves every existing state entry: true Extension immediately preserves old predictions: false Diagnostic source pooling scale restores old predictions: true 'extended: fresh optimizer covers current parameters': true 'extended: all final targets below 0.25 nRMSE': true Added head parameters actually learn: true Hidden target placeholders cannot affect predictions: true Extended checkpoint preserves schema and learned state: true Extended checkpoint preserves predictions: true 'continuation: fresh optimizer covers current parameters': true 'scratch_adaptation: schema and field order match extension': true 'scratch_adaptation: fresh optimizer covers current parameters': true 'scratch_total: schema and field order match extension': true 'scratch_total: fresh optimizer covers current parameters': true 'scratch_total: all final targets below 0.25 nRMSE': true outcome: not_met duration_seconds: 69.677 - started_at: '2026-09-15T18:05:19.254712+00:00' mode: full seed: 7102 steps_override: null accelerator: gpu threads: 1 provenance: revision: null source_sha256: 2db77891e89376c6926a1f11a18aa46fb4f8bebed0df14438ba8eff89c4db6c1 script_sha256: 624c767c1b4024e120b601bc1efaadbe9a478f8c3ab6d8b0e375600c396a001e code_sha256: 5f613c009ea56a099f0f766dcac921baef4eabf53cf14a0a335ebfb2b4ca271d python: 3.12.6 platform: Linux-6.8.0-138-generic-x86_64-with-glibc2.35 host: lab relflow_path: /home/grantham/relflow-mutations-v3_tgit_/src/relflow packages: relflow: 0.2.2 torch: 2.12.0 lightning: 2.6.4 numpy: 2.4.6 pyarrow: 24.0.0 polars: 1.44.1 torchmetrics: 1.9.0 tensordict: 0.12.4 pydantic: 2.13.4 cuda: '13.0' gpu: NVIDIA GeForce RTX 3090 metrics: source: u: rmse: 0.06565498347536482 baseline_rmse: 1.2967289998405087 nrmse: 0.050631229411419086 v: rmse: 0.041381249939003356 baseline_rmse: 1.2693145185962316 nrmse: 0.03260125787008879 source_fit: steps: 512 optimizer_covers_current_parameters: true validation_curve: - step: 0 scores: u: rmse: 1.2393028237273997 baseline_rmse: 1.2723746695686682 nrmse: 0.9740077772434124 v: rmse: 1.7021467221439637 baseline_rmse: 1.3468533739051283 nrmse: 1.2637951206289677 - step: 32 scores: u: rmse: 0.10355146261742668 baseline_rmse: 1.2723746695686682 nrmse: 0.08138441065675282 v: rmse: 0.08961803685159271 baseline_rmse: 1.3468533739051283 nrmse: 0.06653882195932738 - step: 128 scores: u: rmse: 0.03978728611733003 baseline_rmse: 1.2723746695686682 nrmse: 0.03127010232828497 v: rmse: 0.04909547654677152 baseline_rmse: 1.3468533739051283 nrmse: 0.036451983191326795 - step: 512 scores: u: rmse: 0.06498414801296525 baseline_rmse: 1.2723746695686682 nrmse: 0.05107312300943142 v: rmse: 0.03994279807800749 baseline_rmse: 1.3468533739051283 nrmse: 0.029656381943192166 source_prerequisite_met: true downstream_interpretable: true arms: extended: before: u: rmse: 0.06474350162801548 baseline_rmse: 1.2967289998405087 nrmse: 0.049928320902809 v: rmse: 0.04323893491385473 baseline_rmse: 1.2693145185962316 nrmse: 0.03406479188599671 w: rmse: 1.5212991803002514 baseline_rmse: 1.781881033784205 nrmse: 0.8537602406988123 after: u: rmse: 0.0324817988643851 baseline_rmse: 1.2967289998405087 nrmse: 0.02504902633347461 v: rmse: 0.03827953321319368 baseline_rmse: 1.2693145185962316 nrmse: 0.03015764229619624 w: rmse: 0.038706815564526605 baseline_rmse: 1.781881033784205 nrmse: 0.021722446577886523 initial_old_prediction_drift: u: 0.00907289981842041 v: 0.04569077491760254 steps: 256 optimizer_covers_current_parameters: true validation_curve: - step: 0 scores: u: rmse: 0.0639808104354256 baseline_rmse: 1.2761651462099894 nrmse: 0.05013521222189666 v: rmse: 0.04337480398792787 baseline_rmse: 1.3343993130174792 nrmse: 0.032505115646263606 w: rmse: 1.5099151836692908 baseline_rmse: 1.7848255975425062 nrmse: 0.8459735145821896 - step: 32 scores: u: rmse: 0.04423919177469243 baseline_rmse: 1.2761651462099894 nrmse: 0.03466572638037945 v: rmse: 0.07241202769045631 baseline_rmse: 1.3343993130174792 nrmse: 0.054265636218525086 w: rmse: 0.12386977698554788 baseline_rmse: 1.7848255975425062 nrmse: 0.06940161389219311 - step: 128 scores: u: rmse: 0.06044978258179214 baseline_rmse: 1.2761651462099894 nrmse: 0.04736830712021757 v: rmse: 0.07717794848657966 baseline_rmse: 1.3343993130174792 nrmse: 0.057837221387694704 w: rmse: 0.1225639564505473 baseline_rmse: 1.7848255975425062 nrmse: 0.068669990288857 - step: 256 scores: u: rmse: 0.03267472383985952 baseline_rmse: 1.2761651462099894 nrmse: 0.025603836569975547 v: rmse: 0.0398969455033619 baseline_rmse: 1.3343993130174792 nrmse: 0.029898805488098517 w: rmse: 0.03833486196638072 baseline_rmse: 1.7848255975425062 nrmse: 0.021478211663460728 changed_existing_state_entries: [] pooling_scale_control: source_capacity: 4 extended_capacity: 5 old_prediction_drift_with_source_capacity: u: 0.0 v: 0.0 new_head_updated_parameters: - embedder.embeddings.weight - embedder.linear.weight - embedder.scalar.weight - decoder.context_projection.0.weight - decoder.context_projection.0.bias - decoder.context_projection.1.weight - decoder.context_projection.1.bias - decoder.context_projection.3.weight - decoder.context_projection.3.bias - decoder.pool.queries - decoder.pool.blocks.0.attention_norm.weight - decoder.pool.blocks.0.attention_norm.bias - decoder.pool.blocks.0.ffn_norm.weight - decoder.pool.blocks.0.ffn_norm.bias - decoder.pool.blocks.0.attention.q_proj.weight - decoder.pool.blocks.0.attention.q_proj.bias - decoder.pool.blocks.0.attention.k_proj.weight - decoder.pool.blocks.0.attention.k_proj.bias - decoder.pool.blocks.0.attention.v_proj.weight - decoder.pool.blocks.0.attention.v_proj.bias - decoder.pool.blocks.0.attention.out_proj.weight - decoder.pool.blocks.0.attention.out_proj.bias - decoder.pool.blocks.0.ffn.0.weight - decoder.pool.blocks.0.ffn.0.bias - decoder.pool.blocks.0.ffn.3.weight - decoder.pool.blocks.0.ffn.3.bias - decoder.pool.norm.weight - decoder.pool.norm.bias - decoder.classification.weight - decoder.classification.bias - decoder.regression.weight - decoder.regression.bias continuation: before: u: rmse: 0.06565498347536482 baseline_rmse: 1.2967289998405087 nrmse: 0.050631229411419086 v: rmse: 0.041381249939003356 baseline_rmse: 1.2693145185962316 nrmse: 0.03260125787008879 after: u: rmse: 0.019977130988254894 baseline_rmse: 1.2967289998405087 nrmse: 0.015405787169649158 v: rmse: 0.00927443784231114 baseline_rmse: 1.2693145185962316 nrmse: 0.007306650720869391 initial_old_prediction_drift: u: 0.0 v: 0.0 steps: 256 optimizer_covers_current_parameters: true validation_curve: - step: 0 scores: u: rmse: 0.06498562323181704 baseline_rmse: 1.2761651462099894 nrmse: 0.050922581160294314 v: rmse: 0.04133891376364024 baseline_rmse: 1.3343993130174792 nrmse: 0.03097941775026884 - step: 32 scores: u: rmse: 0.05295915753943094 baseline_rmse: 1.2761651462099894 nrmse: 0.04149867099623535 v: rmse: 0.10447499427366826 baseline_rmse: 1.3343993130174792 nrmse: 0.07829365112412925 - step: 128 scores: u: rmse: 0.00932612991919312 baseline_rmse: 1.2761651462099894 nrmse: 0.007307933418249405 v: rmse: 0.01754182235084703 baseline_rmse: 1.3343993130174792 nrmse: 0.013145856850884973 - step: 256 scores: u: rmse: 0.019303910259913987 baseline_rmse: 1.2761651462099894 nrmse: 0.015126498570536561 v: rmse: 0.010437616341014606 baseline_rmse: 1.3343993130174792 nrmse: 0.007821958718947485 scratch_adaptation: before: u: rmse: 1.652515252729601 baseline_rmse: 1.2967289998405087 nrmse: 1.2743720954284605 v: rmse: 1.140062918660766 baseline_rmse: 1.2693145185962316 nrmse: 0.8981721251574367 w: rmse: 1.8464977545612473 baseline_rmse: 1.781881033784205 nrmse: 1.0362632069997484 after: u: rmse: 0.024923579195574745 baseline_rmse: 1.2967289998405087 nrmse: 0.019220345344817785 v: rmse: 0.011155455963321062 baseline_rmse: 1.2693145185962316 nrmse: 0.008788567214734276 w: rmse: 0.042661741833751904 baseline_rmse: 1.781881033784205 nrmse: 0.02394196976391324 initial_old_prediction_drift: u: 3.7187434434890747 v: 2.7047845982015133 steps: 256 optimizer_covers_current_parameters: true validation_curve: - step: 0 scores: u: rmse: 1.651418760910742 baseline_rmse: 1.2761651462099894 nrmse: 1.2940478478159327 v: rmse: 1.206956370920089 baseline_rmse: 1.3343993130174792 nrmse: 0.9044941488996998 w: rmse: 1.85347494086642 baseline_rmse: 1.7848255975425062 nrmse: 1.0384627738522103 - step: 32 scores: u: rmse: 0.14078241903040933 baseline_rmse: 1.2761651462099894 nrmse: 0.11031677165648275 v: rmse: 0.11582958129957194 baseline_rmse: 1.3343993130174792 nrmse: 0.08680278846790346 w: rmse: 0.14517225019432958 baseline_rmse: 1.7848255975425062 nrmse: 0.08133693868701491 - step: 128 scores: u: rmse: 0.0333081031072592 baseline_rmse: 1.2761651462099894 nrmse: 0.026100151070712948 v: rmse: 0.052612817313052604 baseline_rmse: 1.3343993130174792 nrmse: 0.039428090826934825 w: rmse: 0.08863975331154425 baseline_rmse: 1.7848255975425062 nrmse: 0.04966297739879499 - step: 256 scores: u: rmse: 0.025304039807153837 baseline_rmse: 1.2761651462099894 nrmse: 0.019828185938397448 v: rmse: 0.011635730652798045 baseline_rmse: 1.3343993130174792 nrmse: 0.008719826621078026 w: rmse: 0.04262402630844258 baseline_rmse: 1.7848255975425062 nrmse: 0.023881339648608144 scratch_total: before: u: rmse: 1.330800912463188 baseline_rmse: 1.2967289998405087 nrmse: 1.0262752761963911 v: rmse: 1.161068047429632 baseline_rmse: 1.2693145185962316 nrmse: 0.914720528615467 w: rmse: 1.7520921620157683 baseline_rmse: 1.781881033784205 nrmse: 0.9832823453397594 after: u: rmse: 0.029193433240425424 baseline_rmse: 1.2967289998405087 nrmse: 0.02251313361852482 v: rmse: 0.013451315286365112 baseline_rmse: 1.2693145185962316 nrmse: 0.010597306726816044 w: rmse: 0.020846806768631748 baseline_rmse: 1.781881033784205 nrmse: 0.011699325809848875 initial_old_prediction_drift: u: 3.064458005130291 v: 2.9282822608947754 steps: 768 optimizer_covers_current_parameters: true validation_curve: - step: 0 scores: u: rmse: 1.3057670518995306 baseline_rmse: 1.2761651462099894 nrmse: 1.023195983511581 v: rmse: 1.1998756548951877 baseline_rmse: 1.3343993130174792 nrmse: 0.8991878541827986 w: rmse: 1.7733134774154757 baseline_rmse: 1.7848255975425062 nrmse: 0.9935500027885742 - step: 32 scores: u: rmse: 0.07665584843421662 baseline_rmse: 1.2761651462099894 nrmse: 0.0600673421162398 v: rmse: 0.0806347671003684 baseline_rmse: 1.3343993130174792 nrmse: 0.06042776424849087 w: rmse: 0.10556096299850835 baseline_rmse: 1.7848255975425062 nrmse: 0.05914357298766519 - step: 128 scores: u: rmse: 0.034880790811436704 baseline_rmse: 1.2761651462099894 nrmse: 0.02733250544808185 v: rmse: 0.08712851028959709 baseline_rmse: 1.3343993130174792 nrmse: 0.06529418101435713 w: rmse: 0.10003553361552128 baseline_rmse: 1.7848255975425062 nrmse: 0.05604779187011794 - step: 768 scores: u: rmse: 0.02899928418370321 baseline_rmse: 1.2761651462099894 nrmse: 0.0227237707202916 v: rmse: 0.013598748076073829 baseline_rmse: 1.3343993130174792 nrmse: 0.010190913576928452 w: rmse: 0.020866178728345897 baseline_rmse: 1.7848255975425062 nrmse: 0.01169087823318769 test_rows: 1024 baseline: Source-training mean for test comparisons; phase-training mean for validation curves optimizer_policy: New AdamW factory and Trainer for every fit; all available labels rehearsed mutation: extend record with hidden Number record/w checks: Source learns both original tasks below 0.25 nRMSE: true Source optimizer covers current parameters: true Extension preserves every existing state entry: true Extension immediately preserves old predictions: false Diagnostic source pooling scale restores old predictions: true 'extended: fresh optimizer covers current parameters': true 'extended: all final targets below 0.25 nRMSE': true Added head parameters actually learn: true Hidden target placeholders cannot affect predictions: true Extended checkpoint preserves schema and learned state: true Extended checkpoint preserves predictions: true 'continuation: fresh optimizer covers current parameters': true 'scratch_adaptation: schema and field order match extension': true 'scratch_adaptation: fresh optimizer covers current parameters': true 'scratch_total: schema and field order match extension': true 'scratch_total: fresh optimizer covers current parameters': true 'scratch_total: all final targets below 0.25 nRMSE': true outcome: not_met duration_seconds: 64.355 P048: script: proofs/mutations/reset_prediction_target.py page: docs/proofs/mutations/reset-prediction-target.py historical: status: Provisional pass source_revision: null sources: [] summary: New experiment; no historical run predates registration. insights: 'Initial hypothesis: A selected hidden head loses skill locally and learns again.' evidence: No historical measurements. See recorded runs for actual outcomes. runs: - started_at: '2026-09-15T17:53:09.364423+00:00' mode: smoke seed: 4801 steps_override: 2 accelerator: cpu threads: 1 provenance: revision: 5d985953fd0416bc5982a3518d8f43cadb52110c source_sha256: 2252cee7efe14048db1252663050b74fce7ce20d11f143c1b24b93df622d6f7c script_sha256: 84de49318b9298b52f63fe1743ebcf79674875911b46ffd1222e86ea2521027e code_sha256: f7ad2a7bee509149b1c4bad0f6392d223f7a873be15ca244b94bae386803a02e python: 3.12.10 platform: macOS-26.5.2-arm64-arm-64bit host: Granthams-MacBook-Pro.local relflow_path: /Users/grantham/Desktop/relflow/src/relflow packages: relflow: 0.2.2 torch: 2.12.0 lightning: 2.6.4 numpy: 2.4.6 pyarrow: 24.0.0 polars: 1.44.1 torchmetrics: 1.9.0 tensordict: 0.12.4 pydantic: 2.13.4 cuda: null gpu: null metrics: source: u: rmse: 0.6927066134466747 baseline_rmse: 1.3095368006346164 nrmse: 0.5289707117134709 v: rmse: 0.8132343410861116 baseline_rmse: 1.3299997452869836 nrmse: 0.6114545088958901 source_fit: steps: 2 optimizer_covers_current_parameters: true validation_curve: - step: 0 scores: u: rmse: 1.3068514901822017 baseline_rmse: 1.3363771131216209 nrmse: 0.9779062192479107 v: rmse: 1.239518438388088 baseline_rmse: 1.2518777806823944 nrmse: 0.9901273570910657 - step: 2 scores: u: rmse: 0.706964916375878 baseline_rmse: 1.3363771131216209 nrmse: 0.5290160310546553 v: rmse: 0.7580390717333731 baseline_rmse: 1.2518777806823944 nrmse: 0.6055216279341331 source_prerequisite_met: false downstream_interpretable: false arms: selective_reset: before: u: rmse: 1.3737681353196793 baseline_rmse: 1.3095368006346164 nrmse: 1.049048896261591 v: rmse: 0.8132343410861116 baseline_rmse: 1.3299997452869836 nrmse: 0.6114545088958901 after: u: rmse: 0.8479980597975598 baseline_rmse: 1.3095368006346164 nrmse: 0.6475557306878358 v: rmse: 0.39194342676012783 baseline_rmse: 1.3299997452869836 nrmse: 0.29469436227264495 changed_state_entries: - nodes.record/u.embedder.embeddings.weight - nodes.record/u.embedder.counter.counts - nodes.record/u.embedder.linear.weight - nodes.record/u.embedder.linear.bias - nodes.record/u.embedder.scalar.weight - nodes.record/u.embedder.normalizer.mean - nodes.record/u.embedder.normalizer.var - nodes.record/u.embedder.normalizer.count - nodes.record/u.decoder.context_projection.0.weight - nodes.record/u.decoder.context_projection.0.bias - nodes.record/u.decoder.context_projection.1.weight - nodes.record/u.decoder.context_projection.1.bias - nodes.record/u.decoder.context_projection.3.weight - nodes.record/u.decoder.context_projection.3.bias - nodes.record/u.decoder.pool.queries - nodes.record/u.decoder.pool.blocks.0.ffn_norm.weight - nodes.record/u.decoder.pool.blocks.0.ffn_norm.bias - nodes.record/u.decoder.pool.blocks.0.attention.q_proj.weight - nodes.record/u.decoder.pool.blocks.0.attention.q_proj.bias - nodes.record/u.decoder.pool.blocks.0.attention.k_proj.weight - nodes.record/u.decoder.pool.blocks.0.attention.k_proj.bias - nodes.record/u.decoder.pool.blocks.0.attention.v_proj.weight - nodes.record/u.decoder.pool.blocks.0.attention.v_proj.bias - nodes.record/u.decoder.pool.blocks.0.attention.out_proj.weight - nodes.record/u.decoder.pool.blocks.0.attention.out_proj.bias - nodes.record/u.decoder.pool.blocks.0.ffn.0.weight - nodes.record/u.decoder.pool.blocks.0.ffn.0.bias - nodes.record/u.decoder.pool.blocks.0.ffn.3.weight - nodes.record/u.decoder.pool.blocks.0.ffn.3.bias - nodes.record/u.decoder.pool.norm.weight - nodes.record/u.decoder.pool.norm.bias - nodes.record/u.decoder.classification.weight - nodes.record/u.decoder.classification.bias - nodes.record/u.decoder.regression.weight - nodes.record/u.decoder.regression.bias initial_prediction_drift: u: 1.4475162327289581 v: 0.0 steps: 2 optimizer_covers_current_parameters: true validation_curve: - step: 0 scores: u: rmse: 1.38716854729249 baseline_rmse: 1.3218992662289384 nrmse: 1.0493753818698677 v: rmse: 0.7820919004651208 baseline_rmse: 1.3013515850804187 nrmse: 0.6009843223242322 - step: 2 scores: u: rmse: 0.869197105878184 baseline_rmse: 1.3218992662289384 nrmse: 0.6575365673345103 v: rmse: 0.37362651997995877 baseline_rmse: 1.3013515850804187 nrmse: 0.2871065162277956 reset_head_updated_parameters: - embedder.embeddings.weight - embedder.linear.weight - embedder.scalar.weight - decoder.context_projection.0.weight - decoder.context_projection.0.bias - decoder.context_projection.1.weight - decoder.context_projection.1.bias - decoder.context_projection.3.weight - decoder.context_projection.3.bias - decoder.pool.queries - decoder.pool.blocks.0.ffn_norm.weight - decoder.pool.blocks.0.ffn_norm.bias - decoder.pool.blocks.0.attention.q_proj.weight - decoder.pool.blocks.0.attention.k_proj.weight - decoder.pool.blocks.0.attention.v_proj.weight - decoder.pool.blocks.0.attention.v_proj.bias - decoder.pool.blocks.0.attention.out_proj.weight - decoder.pool.blocks.0.attention.out_proj.bias - decoder.pool.blocks.0.ffn.0.weight - decoder.pool.blocks.0.ffn.0.bias - decoder.pool.blocks.0.ffn.3.weight - decoder.pool.blocks.0.ffn.3.bias - decoder.pool.norm.weight - decoder.pool.norm.bias - decoder.classification.weight - decoder.classification.bias - decoder.regression.weight - decoder.regression.bias continuation: before: u: rmse: 0.6927066134466747 baseline_rmse: 1.3095368006346164 nrmse: 0.5289707117134709 v: rmse: 0.8132343410861116 baseline_rmse: 1.3299997452869836 nrmse: 0.6114545088958901 after: u: rmse: 0.35633074373644424 baseline_rmse: 1.3095368006346164 nrmse: 0.27210441399108626 v: rmse: 0.4470884888794418 baseline_rmse: 1.3299997452869836 nrmse: 0.3361568229345565 changed_state_entries: [] initial_prediction_drift: u: 0.0 v: 0.0 steps: 2 optimizer_covers_current_parameters: true validation_curve: - step: 0 scores: u: rmse: 0.7069684164108889 baseline_rmse: 1.3218992662289384 nrmse: 0.5348126248891115 v: rmse: 0.7820919004651208 baseline_rmse: 1.3013515850804187 nrmse: 0.6009843223242322 - step: 2 scores: u: rmse: 0.3762503097698815 baseline_rmse: 1.3218992662289384 nrmse: 0.2846285790317695 v: rmse: 0.4382457039406172 baseline_rmse: 1.3013515850804187 nrmse: 0.3367619550050613 complete_reset: before: u: rmse: 1.2697078532128654 baseline_rmse: 1.3095368006346164 nrmse: 0.9695854691502755 v: rmse: 1.4672726027853868 baseline_rmse: 1.3299997452869836 nrmse: 1.1032126945774587 after: u: rmse: 0.8630547312414026 baseline_rmse: 1.3095368006346164 nrmse: 0.6590534384548464 v: rmse: 0.9049750511344486 baseline_rmse: 1.3299997452869836 nrmse: 0.6804324995860624 changed_state_entries: - nodes.record/a.embedder.embeddings.weight - nodes.record/a.embedder.counter.counts - nodes.record/a.embedder.linear.weight - nodes.record/a.embedder.linear.bias - nodes.record/a.embedder.scalar.weight - nodes.record/a.embedder.normalizer.mean - nodes.record/a.embedder.normalizer.var - nodes.record/a.embedder.normalizer.count - nodes.record/b.embedder.embeddings.weight - nodes.record/b.embedder.counter.counts - nodes.record/b.embedder.linear.weight - nodes.record/b.embedder.linear.bias - nodes.record/b.embedder.scalar.weight - nodes.record/b.embedder.normalizer.mean - nodes.record/b.embedder.normalizer.var - nodes.record/b.embedder.normalizer.count - nodes.record/u.embedder.embeddings.weight - nodes.record/u.embedder.counter.counts - nodes.record/u.embedder.linear.weight - nodes.record/u.embedder.linear.bias - nodes.record/u.embedder.scalar.weight - nodes.record/u.embedder.normalizer.mean - nodes.record/u.embedder.normalizer.var - nodes.record/u.embedder.normalizer.count - nodes.record/u.decoder.context_projection.0.weight - nodes.record/u.decoder.context_projection.0.bias - nodes.record/u.decoder.context_projection.1.weight - nodes.record/u.decoder.context_projection.1.bias - nodes.record/u.decoder.context_projection.3.weight - nodes.record/u.decoder.context_projection.3.bias - nodes.record/u.decoder.pool.queries - nodes.record/u.decoder.pool.blocks.0.ffn_norm.weight - nodes.record/u.decoder.pool.blocks.0.ffn_norm.bias - nodes.record/u.decoder.pool.blocks.0.attention.q_proj.weight - nodes.record/u.decoder.pool.blocks.0.attention.q_proj.bias - nodes.record/u.decoder.pool.blocks.0.attention.k_proj.weight - nodes.record/u.decoder.pool.blocks.0.attention.k_proj.bias - nodes.record/u.decoder.pool.blocks.0.attention.v_proj.weight - nodes.record/u.decoder.pool.blocks.0.attention.v_proj.bias - nodes.record/u.decoder.pool.blocks.0.attention.out_proj.weight - nodes.record/u.decoder.pool.blocks.0.attention.out_proj.bias - nodes.record/u.decoder.pool.blocks.0.ffn.0.weight - nodes.record/u.decoder.pool.blocks.0.ffn.0.bias - nodes.record/u.decoder.pool.blocks.0.ffn.3.weight - nodes.record/u.decoder.pool.blocks.0.ffn.3.bias - nodes.record/u.decoder.pool.norm.weight - nodes.record/u.decoder.pool.norm.bias - nodes.record/u.decoder.classification.weight - nodes.record/u.decoder.classification.bias - nodes.record/u.decoder.regression.weight - nodes.record/u.decoder.regression.bias - nodes.record/v.embedder.embeddings.weight - nodes.record/v.embedder.counter.counts - nodes.record/v.embedder.linear.weight - nodes.record/v.embedder.linear.bias - nodes.record/v.embedder.scalar.weight - nodes.record/v.embedder.normalizer.mean - nodes.record/v.embedder.normalizer.var - nodes.record/v.embedder.normalizer.count - nodes.record/v.decoder.context_projection.0.weight - nodes.record/v.decoder.context_projection.0.bias - nodes.record/v.decoder.context_projection.1.weight - nodes.record/v.decoder.context_projection.1.bias - nodes.record/v.decoder.context_projection.3.weight - nodes.record/v.decoder.context_projection.3.bias - nodes.record/v.decoder.pool.queries - nodes.record/v.decoder.pool.blocks.0.ffn_norm.weight - nodes.record/v.decoder.pool.blocks.0.ffn_norm.bias - nodes.record/v.decoder.pool.blocks.0.attention.q_proj.weight - nodes.record/v.decoder.pool.blocks.0.attention.q_proj.bias - nodes.record/v.decoder.pool.blocks.0.attention.k_proj.weight - nodes.record/v.decoder.pool.blocks.0.attention.k_proj.bias - nodes.record/v.decoder.pool.blocks.0.attention.v_proj.weight - nodes.record/v.decoder.pool.blocks.0.attention.v_proj.bias - nodes.record/v.decoder.pool.blocks.0.attention.out_proj.weight - nodes.record/v.decoder.pool.blocks.0.attention.out_proj.bias - nodes.record/v.decoder.pool.blocks.0.ffn.0.weight - nodes.record/v.decoder.pool.blocks.0.ffn.0.bias - nodes.record/v.decoder.pool.blocks.0.ffn.3.weight - nodes.record/v.decoder.pool.blocks.0.ffn.3.bias - nodes.record/v.decoder.pool.norm.weight - nodes.record/v.decoder.pool.norm.bias - nodes.record/v.decoder.classification.weight - nodes.record/v.decoder.classification.bias - nodes.record/v.decoder.regression.weight - nodes.record/v.decoder.regression.bias - nodes.record.encoder.encoder.0.attention_norm.weight - nodes.record.encoder.encoder.0.attention_norm.bias - nodes.record.encoder.encoder.0.ffn_norm.weight - nodes.record.encoder.encoder.0.ffn_norm.bias - nodes.record.encoder.encoder.0.attention.q_proj.weight - nodes.record.encoder.encoder.0.attention.q_proj.bias - nodes.record.encoder.encoder.0.attention.k_proj.weight - nodes.record.encoder.encoder.0.attention.k_proj.bias - nodes.record.encoder.encoder.0.attention.v_proj.weight - nodes.record.encoder.encoder.0.attention.v_proj.bias - nodes.record.encoder.encoder.0.attention.out_proj.weight - nodes.record.encoder.encoder.0.attention.out_proj.bias - nodes.record.encoder.encoder.0.ffn.0.weight - nodes.record.encoder.encoder.0.ffn.0.bias - nodes.record.encoder.encoder.0.ffn.3.weight - nodes.record.encoder.encoder.0.ffn.3.bias - nodes.record.encoder.coordinate_encoder.attention_norm.weight - nodes.record.encoder.coordinate_encoder.attention_norm.bias - nodes.record.encoder.coordinate_encoder.ffn_norm.weight - nodes.record.encoder.coordinate_encoder.ffn_norm.bias - nodes.record.encoder.coordinate_encoder.attention.q_proj.weight - nodes.record.encoder.coordinate_encoder.attention.q_proj.bias - nodes.record.encoder.coordinate_encoder.attention.k_proj.weight - nodes.record.encoder.coordinate_encoder.attention.k_proj.bias - nodes.record.encoder.coordinate_encoder.attention.v_proj.weight - nodes.record.encoder.coordinate_encoder.attention.v_proj.bias - nodes.record.encoder.coordinate_encoder.attention.out_proj.weight - nodes.record.encoder.coordinate_encoder.attention.out_proj.bias - nodes.record.encoder.coordinate_encoder.ffn.0.weight - nodes.record.encoder.coordinate_encoder.ffn.0.bias - nodes.record.encoder.coordinate_encoder.ffn.3.weight - nodes.record.encoder.coordinate_encoder.ffn.3.bias - nodes.record.encoder.pool.queries - nodes.record.encoder.pool.mass_direction - nodes.record.encoder.pool.blocks.0.attention_norm.weight - nodes.record.encoder.pool.blocks.0.attention_norm.bias - nodes.record.encoder.pool.blocks.0.ffn_norm.weight - nodes.record.encoder.pool.blocks.0.ffn_norm.bias - nodes.record.encoder.pool.blocks.0.attention.q_proj.weight - nodes.record.encoder.pool.blocks.0.attention.q_proj.bias - nodes.record.encoder.pool.blocks.0.attention.k_proj.weight - nodes.record.encoder.pool.blocks.0.attention.k_proj.bias - nodes.record.encoder.pool.blocks.0.attention.v_proj.weight - nodes.record.encoder.pool.blocks.0.attention.v_proj.bias - nodes.record.encoder.pool.blocks.0.attention.out_proj.weight - nodes.record.encoder.pool.blocks.0.attention.out_proj.bias - nodes.record.encoder.pool.blocks.0.ffn.0.weight - nodes.record.encoder.pool.blocks.0.ffn.0.bias - nodes.record.encoder.pool.blocks.0.ffn.3.weight - nodes.record.encoder.pool.blocks.0.ffn.3.bias - nodes.record.encoder.pool.norm.weight - nodes.record.encoder.pool.norm.bias - nodes.record.encoder.pool.mass_projection.weight initial_prediction_drift: u: 1.2862106263637543 v: 1.6154753267765045 steps: 2 optimizer_covers_current_parameters: true validation_curve: - step: 0 scores: u: rmse: 1.2645541266222733 baseline_rmse: 1.3218992662289384 nrmse: 0.9566191304650187 v: rmse: 1.4453547400996387 baseline_rmse: 1.3013515850804187 nrmse: 1.1106566101506852 - step: 2 scores: u: rmse: 0.8800129261808026 baseline_rmse: 1.3218992662289384 nrmse: 0.6657185979770368 v: rmse: 0.894858753647061 baseline_rmse: 1.3013515850804187 nrmse: 0.6876379634115265 test_rows: 1024 baseline: Source-training mean for test comparisons; phase-training mean for validation curves optimizer_policy: New AdamW factory and Trainer for every fit; both hidden labels rehearsed mutation: reset record/u; complete-reset control resets record with descendants=True checks: Source learns both targets below 0.25 nRMSE: false Reset changes selected learned state: true Reset preserves every unselected state entry: true Reset preserves the schema: true Selected target immediately loses its skill: false Unselected target immediately preserves predictions: true 'selective_reset: fresh optimizer covers current parameters': true 'selective_reset: both final targets below 0.25 nRMSE': false Reset head parameters actually relearn: true Hidden target placeholders cannot affect predictions: true Relearned checkpoint preserves schema and state: true Relearned checkpoint preserves predictions: true 'continuation: fresh optimizer covers current parameters': true 'continuation: both final targets below 0.25 nRMSE': false Complete reset loses both learned tasks: false 'complete_reset: fresh optimizer covers current parameters': true 'complete_reset: both final targets below 0.25 nRMSE': false outcome: not_met duration_seconds: 0.678 - started_at: '2026-09-15T17:54:33.039955+00:00' mode: full seed: 4801 steps_override: null accelerator: cpu threads: 1 provenance: revision: 5d985953fd0416bc5982a3518d8f43cadb52110c source_sha256: 2252cee7efe14048db1252663050b74fce7ce20d11f143c1b24b93df622d6f7c script_sha256: 84de49318b9298b52f63fe1743ebcf79674875911b46ffd1222e86ea2521027e code_sha256: f7ad2a7bee509149b1c4bad0f6392d223f7a873be15ca244b94bae386803a02e python: 3.12.10 platform: macOS-26.5.2-arm64-arm-64bit host: Granthams-MacBook-Pro.local relflow_path: /Users/grantham/Desktop/relflow/src/relflow packages: relflow: 0.2.2 torch: 2.12.0 lightning: 2.6.4 numpy: 2.4.6 pyarrow: 24.0.0 polars: 1.44.1 torchmetrics: 1.9.0 tensordict: 0.12.4 pydantic: 2.13.4 cuda: null gpu: null metrics: source: u: rmse: 0.03358791864760073 baseline_rmse: 1.3095368006346164 nrmse: 0.02564870161061808 v: rmse: 0.03156370575360709 baseline_rmse: 1.3299997452869836 nrmse: 0.023732114134199598 source_fit: steps: 512 optimizer_covers_current_parameters: true validation_curve: - step: 0 scores: u: rmse: 1.3068514901822017 baseline_rmse: 1.3363771131216209 nrmse: 0.9779062192479107 v: rmse: 1.239518438388088 baseline_rmse: 1.2518777806823944 nrmse: 0.9901273570910657 - step: 32 scores: u: rmse: 0.12416073898651163 baseline_rmse: 1.3363771131216209 nrmse: 0.0929084595713306 v: rmse: 0.0746600123136656 baseline_rmse: 1.2518777806823944 nrmse: 0.05963841955320006 - step: 128 scores: u: rmse: 0.0258042552892143 baseline_rmse: 1.3363771131216209 nrmse: 0.019309111953390592 v: rmse: 0.053975986675749904 baseline_rmse: 1.2518777806823944 nrmse: 0.04311601939793817 - step: 512 scores: u: rmse: 0.0343734284251833 baseline_rmse: 1.3363771131216209 nrmse: 0.02572135371646031 v: rmse: 0.0322785317593936 baseline_rmse: 1.2518777806823944 nrmse: 0.02578409191175091 source_prerequisite_met: true downstream_interpretable: true arms: selective_reset: before: u: rmse: 1.4491657976216348 baseline_rmse: 1.3095368006346164 nrmse: 1.1066247217484477 v: rmse: 0.03156370575360709 baseline_rmse: 1.3299997452869836 nrmse: 0.023732114134199598 after: u: rmse: 0.019335197482159963 baseline_rmse: 1.3095368006346164 nrmse: 0.014764913420371163 v: rmse: 0.03096444665173069 baseline_rmse: 1.3299997452869836 nrmse: 0.023281543294618654 changed_state_entries: - nodes.record/u.embedder.embeddings.weight - nodes.record/u.embedder.counter.counts - nodes.record/u.embedder.linear.weight - nodes.record/u.embedder.linear.bias - nodes.record/u.embedder.scalar.weight - nodes.record/u.embedder.normalizer.mean - nodes.record/u.embedder.normalizer.var - nodes.record/u.embedder.normalizer.count - nodes.record/u.decoder.context_projection.0.weight - nodes.record/u.decoder.context_projection.0.bias - nodes.record/u.decoder.context_projection.1.weight - nodes.record/u.decoder.context_projection.1.bias - nodes.record/u.decoder.context_projection.3.weight - nodes.record/u.decoder.context_projection.3.bias - nodes.record/u.decoder.pool.queries - nodes.record/u.decoder.pool.blocks.0.ffn_norm.weight - nodes.record/u.decoder.pool.blocks.0.ffn_norm.bias - nodes.record/u.decoder.pool.blocks.0.attention.q_proj.weight - nodes.record/u.decoder.pool.blocks.0.attention.q_proj.bias - nodes.record/u.decoder.pool.blocks.0.attention.k_proj.weight - nodes.record/u.decoder.pool.blocks.0.attention.k_proj.bias - nodes.record/u.decoder.pool.blocks.0.attention.v_proj.weight - nodes.record/u.decoder.pool.blocks.0.attention.v_proj.bias - nodes.record/u.decoder.pool.blocks.0.attention.out_proj.weight - nodes.record/u.decoder.pool.blocks.0.attention.out_proj.bias - nodes.record/u.decoder.pool.blocks.0.ffn.0.weight - nodes.record/u.decoder.pool.blocks.0.ffn.0.bias - nodes.record/u.decoder.pool.blocks.0.ffn.3.weight - nodes.record/u.decoder.pool.blocks.0.ffn.3.bias - nodes.record/u.decoder.pool.norm.weight - nodes.record/u.decoder.pool.norm.bias - nodes.record/u.decoder.classification.weight - nodes.record/u.decoder.classification.bias - nodes.record/u.decoder.regression.weight - nodes.record/u.decoder.regression.bias initial_prediction_drift: u: 3.087439715862274 v: 0.0 steps: 256 optimizer_covers_current_parameters: true validation_curve: - step: 0 scores: u: rmse: 1.4673986990650236 baseline_rmse: 1.3218992662289384 nrmse: 1.110068472351271 v: rmse: 0.031117111830632205 baseline_rmse: 1.3013515850804187 nrmse: 0.023911379666632735 - step: 32 scores: u: rmse: 0.07177933367028597 baseline_rmse: 1.3218992662289384 nrmse: 0.05430015395579664 v: rmse: 0.09672465815910404 baseline_rmse: 1.3013515850804187 nrmse: 0.0743263075620927 - step: 128 scores: u: rmse: 0.031076536166253863 baseline_rmse: 1.3218992662289384 nrmse: 0.023509004778335167 v: rmse: 0.02592063403195276 baseline_rmse: 1.3013515850804187 nrmse: 0.01991824064236335 - step: 256 scores: u: rmse: 0.01888875941418896 baseline_rmse: 1.3218992662289384 nrmse: 0.014289106512687657 v: rmse: 0.029985567132639054 baseline_rmse: 1.3013515850804187 nrmse: 0.023041864686233933 reset_head_updated_parameters: - embedder.embeddings.weight - embedder.linear.weight - embedder.scalar.weight - decoder.context_projection.0.weight - decoder.context_projection.0.bias - decoder.context_projection.1.weight - decoder.context_projection.1.bias - decoder.context_projection.3.weight - decoder.context_projection.3.bias - decoder.pool.queries - decoder.pool.blocks.0.ffn_norm.weight - decoder.pool.blocks.0.ffn_norm.bias - decoder.pool.blocks.0.attention.q_proj.weight - decoder.pool.blocks.0.attention.k_proj.weight - decoder.pool.blocks.0.attention.v_proj.weight - decoder.pool.blocks.0.attention.v_proj.bias - decoder.pool.blocks.0.attention.out_proj.weight - decoder.pool.blocks.0.attention.out_proj.bias - decoder.pool.blocks.0.ffn.0.weight - decoder.pool.blocks.0.ffn.0.bias - decoder.pool.blocks.0.ffn.3.weight - decoder.pool.blocks.0.ffn.3.bias - decoder.pool.norm.weight - decoder.pool.norm.bias - decoder.classification.weight - decoder.classification.bias - decoder.regression.weight - decoder.regression.bias continuation: before: u: rmse: 0.03358791864760073 baseline_rmse: 1.3095368006346164 nrmse: 0.02564870161061808 v: rmse: 0.03156370575360709 baseline_rmse: 1.3299997452869836 nrmse: 0.023732114134199598 after: u: rmse: 0.023723095213619178 baseline_rmse: 1.3095368006346164 nrmse: 0.018115638447214845 v: rmse: 0.01890324128381892 baseline_rmse: 1.3299997452869836 nrmse: 0.01421296609326796 changed_state_entries: [] initial_prediction_drift: u: 0.0 v: 0.0 steps: 256 optimizer_covers_current_parameters: true validation_curve: - step: 0 scores: u: rmse: 0.03508149973739064 baseline_rmse: 1.3218992662289384 nrmse: 0.026538708836316812 v: rmse: 0.031117111830632205 baseline_rmse: 1.3013515850804187 nrmse: 0.023911379666632735 - step: 32 scores: u: rmse: 0.08731390598207757 baseline_rmse: 1.3218992662289384 nrmse: 0.0660518605409043 v: rmse: 0.07474587581370884 baseline_rmse: 1.3013515850804187 nrmse: 0.0574371112854101 - step: 128 scores: u: rmse: 0.03262503869977538 baseline_rmse: 1.3218992662289384 nrmse: 0.024680427271018004 v: rmse: 0.02215005567457824 baseline_rmse: 1.3013515850804187 nrmse: 0.017020808157089576 - step: 256 scores: u: rmse: 0.023497773499379464 baseline_rmse: 1.3218992662289384 nrmse: 0.017775767110010568 v: rmse: 0.019109970737607465 baseline_rmse: 1.3013515850804187 nrmse: 0.014684710078888127 complete_reset: before: u: rmse: 1.2697078532128654 baseline_rmse: 1.3095368006346164 nrmse: 0.9695854691502755 v: rmse: 1.4672726027853868 baseline_rmse: 1.3299997452869836 nrmse: 1.1032126945774587 after: u: rmse: 0.02836517719142516 baseline_rmse: 1.3095368006346164 nrmse: 0.021660465882042467 v: rmse: 0.07017519663442393 baseline_rmse: 1.3299997452869836 nrmse: 0.05276331584505809 changed_state_entries: - nodes.record/a.embedder.embeddings.weight - nodes.record/a.embedder.counter.counts - nodes.record/a.embedder.linear.weight - nodes.record/a.embedder.linear.bias - nodes.record/a.embedder.scalar.weight - nodes.record/a.embedder.normalizer.mean - nodes.record/a.embedder.normalizer.var - nodes.record/a.embedder.normalizer.count - nodes.record/b.embedder.embeddings.weight - nodes.record/b.embedder.counter.counts - nodes.record/b.embedder.linear.weight - nodes.record/b.embedder.linear.bias - nodes.record/b.embedder.scalar.weight - nodes.record/b.embedder.normalizer.mean - nodes.record/b.embedder.normalizer.var - nodes.record/b.embedder.normalizer.count - nodes.record/u.embedder.embeddings.weight - nodes.record/u.embedder.counter.counts - nodes.record/u.embedder.linear.weight - nodes.record/u.embedder.linear.bias - nodes.record/u.embedder.scalar.weight - nodes.record/u.embedder.normalizer.mean - nodes.record/u.embedder.normalizer.var - nodes.record/u.embedder.normalizer.count - nodes.record/u.decoder.context_projection.0.weight - nodes.record/u.decoder.context_projection.0.bias - nodes.record/u.decoder.context_projection.1.weight - nodes.record/u.decoder.context_projection.1.bias - nodes.record/u.decoder.context_projection.3.weight - nodes.record/u.decoder.context_projection.3.bias - nodes.record/u.decoder.pool.queries - nodes.record/u.decoder.pool.blocks.0.ffn_norm.weight - nodes.record/u.decoder.pool.blocks.0.ffn_norm.bias - nodes.record/u.decoder.pool.blocks.0.attention.q_proj.weight - nodes.record/u.decoder.pool.blocks.0.attention.q_proj.bias - nodes.record/u.decoder.pool.blocks.0.attention.k_proj.weight - nodes.record/u.decoder.pool.blocks.0.attention.k_proj.bias - nodes.record/u.decoder.pool.blocks.0.attention.v_proj.weight - nodes.record/u.decoder.pool.blocks.0.attention.v_proj.bias - nodes.record/u.decoder.pool.blocks.0.attention.out_proj.weight - nodes.record/u.decoder.pool.blocks.0.attention.out_proj.bias - nodes.record/u.decoder.pool.blocks.0.ffn.0.weight - nodes.record/u.decoder.pool.blocks.0.ffn.0.bias - nodes.record/u.decoder.pool.blocks.0.ffn.3.weight - nodes.record/u.decoder.pool.blocks.0.ffn.3.bias - nodes.record/u.decoder.pool.norm.weight - nodes.record/u.decoder.pool.norm.bias - nodes.record/u.decoder.classification.weight - nodes.record/u.decoder.classification.bias - nodes.record/u.decoder.regression.weight - nodes.record/u.decoder.regression.bias - nodes.record/v.embedder.embeddings.weight - nodes.record/v.embedder.counter.counts - nodes.record/v.embedder.linear.weight - nodes.record/v.embedder.linear.bias - nodes.record/v.embedder.scalar.weight - nodes.record/v.embedder.normalizer.mean - nodes.record/v.embedder.normalizer.var - nodes.record/v.embedder.normalizer.count - nodes.record/v.decoder.context_projection.0.weight - nodes.record/v.decoder.context_projection.0.bias - nodes.record/v.decoder.context_projection.1.weight - nodes.record/v.decoder.context_projection.1.bias - nodes.record/v.decoder.context_projection.3.weight - nodes.record/v.decoder.context_projection.3.bias - nodes.record/v.decoder.pool.queries - nodes.record/v.decoder.pool.blocks.0.ffn_norm.weight - nodes.record/v.decoder.pool.blocks.0.ffn_norm.bias - nodes.record/v.decoder.pool.blocks.0.attention.q_proj.weight - nodes.record/v.decoder.pool.blocks.0.attention.q_proj.bias - nodes.record/v.decoder.pool.blocks.0.attention.k_proj.weight - nodes.record/v.decoder.pool.blocks.0.attention.k_proj.bias - nodes.record/v.decoder.pool.blocks.0.attention.v_proj.weight - nodes.record/v.decoder.pool.blocks.0.attention.v_proj.bias - nodes.record/v.decoder.pool.blocks.0.attention.out_proj.weight - nodes.record/v.decoder.pool.blocks.0.attention.out_proj.bias - nodes.record/v.decoder.pool.blocks.0.ffn.0.weight - nodes.record/v.decoder.pool.blocks.0.ffn.0.bias - nodes.record/v.decoder.pool.blocks.0.ffn.3.weight - nodes.record/v.decoder.pool.blocks.0.ffn.3.bias - nodes.record/v.decoder.pool.norm.weight - nodes.record/v.decoder.pool.norm.bias - nodes.record/v.decoder.classification.weight - nodes.record/v.decoder.classification.bias - nodes.record/v.decoder.regression.weight - nodes.record/v.decoder.regression.bias - nodes.record.encoder.encoder.0.attention_norm.weight - nodes.record.encoder.encoder.0.attention_norm.bias - nodes.record.encoder.encoder.0.ffn_norm.weight - nodes.record.encoder.encoder.0.ffn_norm.bias - nodes.record.encoder.encoder.0.attention.q_proj.weight - nodes.record.encoder.encoder.0.attention.q_proj.bias - nodes.record.encoder.encoder.0.attention.k_proj.weight - nodes.record.encoder.encoder.0.attention.k_proj.bias - nodes.record.encoder.encoder.0.attention.v_proj.weight - nodes.record.encoder.encoder.0.attention.v_proj.bias - nodes.record.encoder.encoder.0.attention.out_proj.weight - nodes.record.encoder.encoder.0.attention.out_proj.bias - nodes.record.encoder.encoder.0.ffn.0.weight - nodes.record.encoder.encoder.0.ffn.0.bias - nodes.record.encoder.encoder.0.ffn.3.weight - nodes.record.encoder.encoder.0.ffn.3.bias - nodes.record.encoder.coordinate_encoder.attention_norm.weight - nodes.record.encoder.coordinate_encoder.attention_norm.bias - nodes.record.encoder.coordinate_encoder.ffn_norm.weight - nodes.record.encoder.coordinate_encoder.ffn_norm.bias - nodes.record.encoder.coordinate_encoder.attention.q_proj.weight - nodes.record.encoder.coordinate_encoder.attention.q_proj.bias - nodes.record.encoder.coordinate_encoder.attention.k_proj.weight - nodes.record.encoder.coordinate_encoder.attention.k_proj.bias - nodes.record.encoder.coordinate_encoder.attention.v_proj.weight - nodes.record.encoder.coordinate_encoder.attention.v_proj.bias - nodes.record.encoder.coordinate_encoder.attention.out_proj.weight - nodes.record.encoder.coordinate_encoder.attention.out_proj.bias - nodes.record.encoder.coordinate_encoder.ffn.0.weight - nodes.record.encoder.coordinate_encoder.ffn.0.bias - nodes.record.encoder.coordinate_encoder.ffn.3.weight - nodes.record.encoder.coordinate_encoder.ffn.3.bias - nodes.record.encoder.pool.queries - nodes.record.encoder.pool.mass_direction - nodes.record.encoder.pool.blocks.0.attention_norm.weight - nodes.record.encoder.pool.blocks.0.attention_norm.bias - nodes.record.encoder.pool.blocks.0.ffn_norm.weight - nodes.record.encoder.pool.blocks.0.ffn_norm.bias - nodes.record.encoder.pool.blocks.0.attention.q_proj.weight - nodes.record.encoder.pool.blocks.0.attention.q_proj.bias - nodes.record.encoder.pool.blocks.0.attention.k_proj.weight - nodes.record.encoder.pool.blocks.0.attention.k_proj.bias - nodes.record.encoder.pool.blocks.0.attention.v_proj.weight - nodes.record.encoder.pool.blocks.0.attention.v_proj.bias - nodes.record.encoder.pool.blocks.0.attention.out_proj.weight - nodes.record.encoder.pool.blocks.0.attention.out_proj.bias - nodes.record.encoder.pool.blocks.0.ffn.0.weight - nodes.record.encoder.pool.blocks.0.ffn.0.bias - nodes.record.encoder.pool.blocks.0.ffn.3.weight - nodes.record.encoder.pool.blocks.0.ffn.3.bias - nodes.record.encoder.pool.norm.weight - nodes.record.encoder.pool.norm.bias - nodes.record.encoder.pool.mass_projection.weight initial_prediction_drift: u: 2.959281414747238 v: 3.2186422646045685 steps: 256 optimizer_covers_current_parameters: true validation_curve: - step: 0 scores: u: rmse: 1.2645541266222733 baseline_rmse: 1.3218992662289384 nrmse: 0.9566191304650187 v: rmse: 1.4453547400996387 baseline_rmse: 1.3013515850804187 nrmse: 1.1106566101506852 - step: 32 scores: u: rmse: 0.09326566428708902 baseline_rmse: 1.3218992662289384 nrmse: 0.07055429008078172 v: rmse: 0.07710959390132345 baseline_rmse: 1.3013515850804187 nrmse: 0.05925346753741293 - step: 128 scores: u: rmse: 0.03477009442148113 baseline_rmse: 1.3218992662289384 nrmse: 0.02630313467127633 v: rmse: 0.05007222845263086 baseline_rmse: 1.3013515850804187 nrmse: 0.03847709491170027 - step: 256 scores: u: rmse: 0.029895567385493513 baseline_rmse: 1.3218992662289384 nrmse: 0.022615616900051996 v: rmse: 0.06995185957194495 baseline_rmse: 1.3013515850804187 nrmse: 0.053753236538012276 test_rows: 1024 baseline: Source-training mean for test comparisons; phase-training mean for validation curves optimizer_policy: New AdamW factory and Trainer for every fit; both hidden labels rehearsed mutation: reset record/u; complete-reset control resets record with descendants=True checks: Source learns both targets below 0.25 nRMSE: true Reset changes selected learned state: true Reset preserves every unselected state entry: true Reset preserves the schema: true Selected target immediately loses its skill: true Unselected target immediately preserves predictions: true 'selective_reset: fresh optimizer covers current parameters': true 'selective_reset: both final targets below 0.25 nRMSE': true Reset head parameters actually relearn: true Hidden target placeholders cannot affect predictions: true Relearned checkpoint preserves schema and state: true Relearned checkpoint preserves predictions: true 'continuation: fresh optimizer covers current parameters': true 'continuation: both final targets below 0.25 nRMSE': true Complete reset loses both learned tasks: true 'complete_reset: fresh optimizer covers current parameters': true 'complete_reset: both final targets below 0.25 nRMSE': true outcome: met duration_seconds: 14.438 - started_at: '2026-09-15T17:56:07.986706+00:00' mode: full seed: 4801 steps_override: null accelerator: cpu threads: 1 provenance: revision: 5d985953fd0416bc5982a3518d8f43cadb52110c source_sha256: 2252cee7efe14048db1252663050b74fce7ce20d11f143c1b24b93df622d6f7c script_sha256: 84de49318b9298b52f63fe1743ebcf79674875911b46ffd1222e86ea2521027e code_sha256: f7ad2a7bee509149b1c4bad0f6392d223f7a873be15ca244b94bae386803a02e python: 3.12.10 platform: macOS-26.5.2-arm64-arm-64bit host: Granthams-MacBook-Pro.local relflow_path: /Users/grantham/Desktop/relflow/src/relflow packages: relflow: 0.2.2 torch: 2.12.0 lightning: 2.6.4 numpy: 2.4.6 pyarrow: 24.0.0 polars: 1.44.1 torchmetrics: 1.9.0 tensordict: 0.12.4 pydantic: 2.13.4 cuda: null gpu: null metrics: source: u: rmse: 0.03358791864760073 baseline_rmse: 1.3095368006346164 nrmse: 0.02564870161061808 v: rmse: 0.03156370575360709 baseline_rmse: 1.3299997452869836 nrmse: 0.023732114134199598 source_fit: steps: 512 optimizer_covers_current_parameters: true validation_curve: - step: 0 scores: u: rmse: 1.3068514901822017 baseline_rmse: 1.3363771131216209 nrmse: 0.9779062192479107 v: rmse: 1.239518438388088 baseline_rmse: 1.2518777806823944 nrmse: 0.9901273570910657 - step: 32 scores: u: rmse: 0.12416073898651163 baseline_rmse: 1.3363771131216209 nrmse: 0.0929084595713306 v: rmse: 0.0746600123136656 baseline_rmse: 1.2518777806823944 nrmse: 0.05963841955320006 - step: 128 scores: u: rmse: 0.0258042552892143 baseline_rmse: 1.3363771131216209 nrmse: 0.019309111953390592 v: rmse: 0.053975986675749904 baseline_rmse: 1.2518777806823944 nrmse: 0.04311601939793817 - step: 512 scores: u: rmse: 0.0343734284251833 baseline_rmse: 1.3363771131216209 nrmse: 0.02572135371646031 v: rmse: 0.0322785317593936 baseline_rmse: 1.2518777806823944 nrmse: 0.02578409191175091 source_prerequisite_met: true downstream_interpretable: true arms: selective_reset: before: u: rmse: 1.4491657976216348 baseline_rmse: 1.3095368006346164 nrmse: 1.1066247217484477 v: rmse: 0.03156370575360709 baseline_rmse: 1.3299997452869836 nrmse: 0.023732114134199598 after: u: rmse: 0.019335197482159963 baseline_rmse: 1.3095368006346164 nrmse: 0.014764913420371163 v: rmse: 0.03096444665173069 baseline_rmse: 1.3299997452869836 nrmse: 0.023281543294618654 changed_state_entries: - nodes.record/u.embedder.embeddings.weight - nodes.record/u.embedder.counter.counts - nodes.record/u.embedder.linear.weight - nodes.record/u.embedder.linear.bias - nodes.record/u.embedder.scalar.weight - nodes.record/u.embedder.normalizer.mean - nodes.record/u.embedder.normalizer.var - nodes.record/u.embedder.normalizer.count - nodes.record/u.decoder.context_projection.0.weight - nodes.record/u.decoder.context_projection.0.bias - nodes.record/u.decoder.context_projection.1.weight - nodes.record/u.decoder.context_projection.1.bias - nodes.record/u.decoder.context_projection.3.weight - nodes.record/u.decoder.context_projection.3.bias - nodes.record/u.decoder.pool.queries - nodes.record/u.decoder.pool.blocks.0.ffn_norm.weight - nodes.record/u.decoder.pool.blocks.0.ffn_norm.bias - nodes.record/u.decoder.pool.blocks.0.attention.q_proj.weight - nodes.record/u.decoder.pool.blocks.0.attention.q_proj.bias - nodes.record/u.decoder.pool.blocks.0.attention.k_proj.weight - nodes.record/u.decoder.pool.blocks.0.attention.k_proj.bias - nodes.record/u.decoder.pool.blocks.0.attention.v_proj.weight - nodes.record/u.decoder.pool.blocks.0.attention.v_proj.bias - nodes.record/u.decoder.pool.blocks.0.attention.out_proj.weight - nodes.record/u.decoder.pool.blocks.0.attention.out_proj.bias - nodes.record/u.decoder.pool.blocks.0.ffn.0.weight - nodes.record/u.decoder.pool.blocks.0.ffn.0.bias - nodes.record/u.decoder.pool.blocks.0.ffn.3.weight - nodes.record/u.decoder.pool.blocks.0.ffn.3.bias - nodes.record/u.decoder.pool.norm.weight - nodes.record/u.decoder.pool.norm.bias - nodes.record/u.decoder.classification.weight - nodes.record/u.decoder.classification.bias - nodes.record/u.decoder.regression.weight - nodes.record/u.decoder.regression.bias initial_prediction_drift: u: 3.087439715862274 v: 0.0 steps: 256 optimizer_covers_current_parameters: true validation_curve: - step: 0 scores: u: rmse: 1.4673986990650236 baseline_rmse: 1.3218992662289384 nrmse: 1.110068472351271 v: rmse: 0.031117111830632205 baseline_rmse: 1.3013515850804187 nrmse: 0.023911379666632735 - step: 32 scores: u: rmse: 0.07177933367028597 baseline_rmse: 1.3218992662289384 nrmse: 0.05430015395579664 v: rmse: 0.09672465815910404 baseline_rmse: 1.3013515850804187 nrmse: 0.0743263075620927 - step: 128 scores: u: rmse: 0.031076536166253863 baseline_rmse: 1.3218992662289384 nrmse: 0.023509004778335167 v: rmse: 0.02592063403195276 baseline_rmse: 1.3013515850804187 nrmse: 0.01991824064236335 - step: 256 scores: u: rmse: 0.01888875941418896 baseline_rmse: 1.3218992662289384 nrmse: 0.014289106512687657 v: rmse: 0.029985567132639054 baseline_rmse: 1.3013515850804187 nrmse: 0.023041864686233933 reset_head_updated_parameters: - embedder.embeddings.weight - embedder.linear.weight - embedder.scalar.weight - decoder.context_projection.0.weight - decoder.context_projection.0.bias - decoder.context_projection.1.weight - decoder.context_projection.1.bias - decoder.context_projection.3.weight - decoder.context_projection.3.bias - decoder.pool.queries - decoder.pool.blocks.0.ffn_norm.weight - decoder.pool.blocks.0.ffn_norm.bias - decoder.pool.blocks.0.attention.q_proj.weight - decoder.pool.blocks.0.attention.k_proj.weight - decoder.pool.blocks.0.attention.v_proj.weight - decoder.pool.blocks.0.attention.v_proj.bias - decoder.pool.blocks.0.attention.out_proj.weight - decoder.pool.blocks.0.attention.out_proj.bias - decoder.pool.blocks.0.ffn.0.weight - decoder.pool.blocks.0.ffn.0.bias - decoder.pool.blocks.0.ffn.3.weight - decoder.pool.blocks.0.ffn.3.bias - decoder.pool.norm.weight - decoder.pool.norm.bias - decoder.classification.weight - decoder.classification.bias - decoder.regression.weight - decoder.regression.bias continuation: before: u: rmse: 0.03358791864760073 baseline_rmse: 1.3095368006346164 nrmse: 0.02564870161061808 v: rmse: 0.03156370575360709 baseline_rmse: 1.3299997452869836 nrmse: 0.023732114134199598 after: u: rmse: 0.023723095213619178 baseline_rmse: 1.3095368006346164 nrmse: 0.018115638447214845 v: rmse: 0.01890324128381892 baseline_rmse: 1.3299997452869836 nrmse: 0.01421296609326796 changed_state_entries: [] initial_prediction_drift: u: 0.0 v: 0.0 steps: 256 optimizer_covers_current_parameters: true validation_curve: - step: 0 scores: u: rmse: 0.03508149973739064 baseline_rmse: 1.3218992662289384 nrmse: 0.026538708836316812 v: rmse: 0.031117111830632205 baseline_rmse: 1.3013515850804187 nrmse: 0.023911379666632735 - step: 32 scores: u: rmse: 0.08731390598207757 baseline_rmse: 1.3218992662289384 nrmse: 0.0660518605409043 v: rmse: 0.07474587581370884 baseline_rmse: 1.3013515850804187 nrmse: 0.0574371112854101 - step: 128 scores: u: rmse: 0.03262503869977538 baseline_rmse: 1.3218992662289384 nrmse: 0.024680427271018004 v: rmse: 0.02215005567457824 baseline_rmse: 1.3013515850804187 nrmse: 0.017020808157089576 - step: 256 scores: u: rmse: 0.023497773499379464 baseline_rmse: 1.3218992662289384 nrmse: 0.017775767110010568 v: rmse: 0.019109970737607465 baseline_rmse: 1.3013515850804187 nrmse: 0.014684710078888127 complete_reset: before: u: rmse: 1.2697078532128654 baseline_rmse: 1.3095368006346164 nrmse: 0.9695854691502755 v: rmse: 1.4672726027853868 baseline_rmse: 1.3299997452869836 nrmse: 1.1032126945774587 after: u: rmse: 0.02836517719142516 baseline_rmse: 1.3095368006346164 nrmse: 0.021660465882042467 v: rmse: 0.07017519663442393 baseline_rmse: 1.3299997452869836 nrmse: 0.05276331584505809 changed_state_entries: - nodes.record/a.embedder.embeddings.weight - nodes.record/a.embedder.counter.counts - nodes.record/a.embedder.linear.weight - nodes.record/a.embedder.linear.bias - nodes.record/a.embedder.scalar.weight - nodes.record/a.embedder.normalizer.mean - nodes.record/a.embedder.normalizer.var - nodes.record/a.embedder.normalizer.count - nodes.record/b.embedder.embeddings.weight - nodes.record/b.embedder.counter.counts - nodes.record/b.embedder.linear.weight - nodes.record/b.embedder.linear.bias - nodes.record/b.embedder.scalar.weight - nodes.record/b.embedder.normalizer.mean - nodes.record/b.embedder.normalizer.var - nodes.record/b.embedder.normalizer.count - nodes.record/u.embedder.embeddings.weight - nodes.record/u.embedder.counter.counts - nodes.record/u.embedder.linear.weight - nodes.record/u.embedder.linear.bias - nodes.record/u.embedder.scalar.weight - nodes.record/u.embedder.normalizer.mean - nodes.record/u.embedder.normalizer.var - nodes.record/u.embedder.normalizer.count - nodes.record/u.decoder.context_projection.0.weight - nodes.record/u.decoder.context_projection.0.bias - nodes.record/u.decoder.context_projection.1.weight - nodes.record/u.decoder.context_projection.1.bias - nodes.record/u.decoder.context_projection.3.weight - nodes.record/u.decoder.context_projection.3.bias - nodes.record/u.decoder.pool.queries - nodes.record/u.decoder.pool.blocks.0.ffn_norm.weight - nodes.record/u.decoder.pool.blocks.0.ffn_norm.bias - nodes.record/u.decoder.pool.blocks.0.attention.q_proj.weight - nodes.record/u.decoder.pool.blocks.0.attention.q_proj.bias - nodes.record/u.decoder.pool.blocks.0.attention.k_proj.weight - nodes.record/u.decoder.pool.blocks.0.attention.k_proj.bias - nodes.record/u.decoder.pool.blocks.0.attention.v_proj.weight - nodes.record/u.decoder.pool.blocks.0.attention.v_proj.bias - nodes.record/u.decoder.pool.blocks.0.attention.out_proj.weight - nodes.record/u.decoder.pool.blocks.0.attention.out_proj.bias - nodes.record/u.decoder.pool.blocks.0.ffn.0.weight - nodes.record/u.decoder.pool.blocks.0.ffn.0.bias - nodes.record/u.decoder.pool.blocks.0.ffn.3.weight - nodes.record/u.decoder.pool.blocks.0.ffn.3.bias - nodes.record/u.decoder.pool.norm.weight - nodes.record/u.decoder.pool.norm.bias - nodes.record/u.decoder.classification.weight - nodes.record/u.decoder.classification.bias - nodes.record/u.decoder.regression.weight - nodes.record/u.decoder.regression.bias - nodes.record/v.embedder.embeddings.weight - nodes.record/v.embedder.counter.counts - nodes.record/v.embedder.linear.weight - nodes.record/v.embedder.linear.bias - nodes.record/v.embedder.scalar.weight - nodes.record/v.embedder.normalizer.mean - nodes.record/v.embedder.normalizer.var - nodes.record/v.embedder.normalizer.count - nodes.record/v.decoder.context_projection.0.weight - nodes.record/v.decoder.context_projection.0.bias - nodes.record/v.decoder.context_projection.1.weight - nodes.record/v.decoder.context_projection.1.bias - nodes.record/v.decoder.context_projection.3.weight - nodes.record/v.decoder.context_projection.3.bias - nodes.record/v.decoder.pool.queries - nodes.record/v.decoder.pool.blocks.0.ffn_norm.weight - nodes.record/v.decoder.pool.blocks.0.ffn_norm.bias - nodes.record/v.decoder.pool.blocks.0.attention.q_proj.weight - nodes.record/v.decoder.pool.blocks.0.attention.q_proj.bias - nodes.record/v.decoder.pool.blocks.0.attention.k_proj.weight - nodes.record/v.decoder.pool.blocks.0.attention.k_proj.bias - nodes.record/v.decoder.pool.blocks.0.attention.v_proj.weight - nodes.record/v.decoder.pool.blocks.0.attention.v_proj.bias - nodes.record/v.decoder.pool.blocks.0.attention.out_proj.weight - nodes.record/v.decoder.pool.blocks.0.attention.out_proj.bias - nodes.record/v.decoder.pool.blocks.0.ffn.0.weight - nodes.record/v.decoder.pool.blocks.0.ffn.0.bias - nodes.record/v.decoder.pool.blocks.0.ffn.3.weight - nodes.record/v.decoder.pool.blocks.0.ffn.3.bias - nodes.record/v.decoder.pool.norm.weight - nodes.record/v.decoder.pool.norm.bias - nodes.record/v.decoder.classification.weight - nodes.record/v.decoder.classification.bias - nodes.record/v.decoder.regression.weight - nodes.record/v.decoder.regression.bias - nodes.record.encoder.encoder.0.attention_norm.weight - nodes.record.encoder.encoder.0.attention_norm.bias - nodes.record.encoder.encoder.0.ffn_norm.weight - nodes.record.encoder.encoder.0.ffn_norm.bias - nodes.record.encoder.encoder.0.attention.q_proj.weight - nodes.record.encoder.encoder.0.attention.q_proj.bias - nodes.record.encoder.encoder.0.attention.k_proj.weight - nodes.record.encoder.encoder.0.attention.k_proj.bias - nodes.record.encoder.encoder.0.attention.v_proj.weight - nodes.record.encoder.encoder.0.attention.v_proj.bias - nodes.record.encoder.encoder.0.attention.out_proj.weight - nodes.record.encoder.encoder.0.attention.out_proj.bias - nodes.record.encoder.encoder.0.ffn.0.weight - nodes.record.encoder.encoder.0.ffn.0.bias - nodes.record.encoder.encoder.0.ffn.3.weight - nodes.record.encoder.encoder.0.ffn.3.bias - nodes.record.encoder.coordinate_encoder.attention_norm.weight - nodes.record.encoder.coordinate_encoder.attention_norm.bias - nodes.record.encoder.coordinate_encoder.ffn_norm.weight - nodes.record.encoder.coordinate_encoder.ffn_norm.bias - nodes.record.encoder.coordinate_encoder.attention.q_proj.weight - nodes.record.encoder.coordinate_encoder.attention.q_proj.bias - nodes.record.encoder.coordinate_encoder.attention.k_proj.weight - nodes.record.encoder.coordinate_encoder.attention.k_proj.bias - nodes.record.encoder.coordinate_encoder.attention.v_proj.weight - nodes.record.encoder.coordinate_encoder.attention.v_proj.bias - nodes.record.encoder.coordinate_encoder.attention.out_proj.weight - nodes.record.encoder.coordinate_encoder.attention.out_proj.bias - nodes.record.encoder.coordinate_encoder.ffn.0.weight - nodes.record.encoder.coordinate_encoder.ffn.0.bias - nodes.record.encoder.coordinate_encoder.ffn.3.weight - nodes.record.encoder.coordinate_encoder.ffn.3.bias - nodes.record.encoder.pool.queries - nodes.record.encoder.pool.mass_direction - nodes.record.encoder.pool.blocks.0.attention_norm.weight - nodes.record.encoder.pool.blocks.0.attention_norm.bias - nodes.record.encoder.pool.blocks.0.ffn_norm.weight - nodes.record.encoder.pool.blocks.0.ffn_norm.bias - nodes.record.encoder.pool.blocks.0.attention.q_proj.weight - nodes.record.encoder.pool.blocks.0.attention.q_proj.bias - nodes.record.encoder.pool.blocks.0.attention.k_proj.weight - nodes.record.encoder.pool.blocks.0.attention.k_proj.bias - nodes.record.encoder.pool.blocks.0.attention.v_proj.weight - nodes.record.encoder.pool.blocks.0.attention.v_proj.bias - nodes.record.encoder.pool.blocks.0.attention.out_proj.weight - nodes.record.encoder.pool.blocks.0.attention.out_proj.bias - nodes.record.encoder.pool.blocks.0.ffn.0.weight - nodes.record.encoder.pool.blocks.0.ffn.0.bias - nodes.record.encoder.pool.blocks.0.ffn.3.weight - nodes.record.encoder.pool.blocks.0.ffn.3.bias - nodes.record.encoder.pool.norm.weight - nodes.record.encoder.pool.norm.bias - nodes.record.encoder.pool.mass_projection.weight initial_prediction_drift: u: 2.959281414747238 v: 3.2186422646045685 steps: 256 optimizer_covers_current_parameters: true validation_curve: - step: 0 scores: u: rmse: 1.2645541266222733 baseline_rmse: 1.3218992662289384 nrmse: 0.9566191304650187 v: rmse: 1.4453547400996387 baseline_rmse: 1.3013515850804187 nrmse: 1.1106566101506852 - step: 32 scores: u: rmse: 0.09326566428708902 baseline_rmse: 1.3218992662289384 nrmse: 0.07055429008078172 v: rmse: 0.07710959390132345 baseline_rmse: 1.3013515850804187 nrmse: 0.05925346753741293 - step: 128 scores: u: rmse: 0.03477009442148113 baseline_rmse: 1.3218992662289384 nrmse: 0.02630313467127633 v: rmse: 0.05007222845263086 baseline_rmse: 1.3013515850804187 nrmse: 0.03847709491170027 - step: 256 scores: u: rmse: 0.029895567385493513 baseline_rmse: 1.3218992662289384 nrmse: 0.022615616900051996 v: rmse: 0.06995185957194495 baseline_rmse: 1.3013515850804187 nrmse: 0.053753236538012276 test_rows: 1024 baseline: Source-training mean for test comparisons; phase-training mean for validation curves optimizer_policy: New AdamW factory and Trainer for every fit; both hidden labels rehearsed mutation: reset record/u; complete-reset control resets record with descendants=True checks: Source learns both targets below 0.25 nRMSE: true Reset changes selected learned state: true Reset preserves every unselected state entry: true Reset preserves the schema: true Selected target immediately loses its skill: true Unselected target immediately preserves predictions: true 'selective_reset: fresh optimizer covers current parameters': true 'selective_reset: both final targets below 0.25 nRMSE': true Reset head parameters actually relearn: true Hidden target placeholders cannot affect predictions: true Relearned checkpoint preserves schema and state: true Relearned checkpoint preserves predictions: true 'continuation: fresh optimizer covers current parameters': true 'continuation: both final targets below 0.25 nRMSE': true Complete reset loses both learned tasks: true 'complete_reset: fresh optimizer covers current parameters': true 'complete_reset: both final targets below 0.25 nRMSE': true outcome: met duration_seconds: 15.008 - started_at: '2026-09-15T17:59:28.750941+00:00' mode: full seed: 4801 steps_override: null accelerator: gpu threads: 1 provenance: revision: null source_sha256: 2252cee7efe14048db1252663050b74fce7ce20d11f143c1b24b93df622d6f7c script_sha256: 84de49318b9298b52f63fe1743ebcf79674875911b46ffd1222e86ea2521027e code_sha256: f7ad2a7bee509149b1c4bad0f6392d223f7a873be15ca244b94bae386803a02e python: 3.12.6 platform: Linux-6.8.0-138-generic-x86_64-with-glibc2.35 host: lab relflow_path: /home/grantham/relflow-mutations-v3_tgit_/src/relflow packages: relflow: 0.2.2 torch: 2.12.0 lightning: 2.6.4 numpy: 2.4.6 pyarrow: 24.0.0 polars: 1.44.1 torchmetrics: 1.9.0 tensordict: 0.12.4 pydantic: 2.13.4 cuda: '13.0' gpu: NVIDIA GeForce RTX 3090 metrics: source: u: rmse: 0.03486898976137349 baseline_rmse: 1.3095368006346164 nrmse: 0.026626964392658214 v: rmse: 0.023913129355889483 baseline_rmse: 1.3299997452869836 nrmse: 0.01797979995156283 source_fit: steps: 512 optimizer_covers_current_parameters: true validation_curve: - step: 0 scores: u: rmse: 1.306851460766046 baseline_rmse: 1.3363771131216209 nrmse: 0.9779061972360433 v: rmse: 1.239518428898769 baseline_rmse: 1.2518777806823944 nrmse: 0.9901273495109976 - step: 32 scores: u: rmse: 0.12416076009991522 baseline_rmse: 1.3363771131216209 nrmse: 0.09290847537031684 v: rmse: 0.0746600408669145 baseline_rmse: 1.2518777806823944 nrmse: 0.059638442361535926 - step: 128 scores: u: rmse: 0.025857928944642575 baseline_rmse: 1.3363771131216209 nrmse: 0.019349275508199534 v: rmse: 0.05415747091400457 baseline_rmse: 1.2518777806823944 nrmse: 0.04326098901162981 - step: 512 scores: u: rmse: 0.035328201845256105 baseline_rmse: 1.3363771131216209 nrmse: 0.026435802812226822 v: rmse: 0.023122613995360922 baseline_rmse: 1.2518777806823944 nrmse: 0.01847034459127221 source_prerequisite_met: true downstream_interpretable: true arms: selective_reset: before: u: rmse: 1.453095410541797 baseline_rmse: 1.3095368006346164 nrmse: 1.1096254873002505 v: rmse: 0.023913129355889483 baseline_rmse: 1.3299997452869836 nrmse: 0.01797979995156283 after: u: rmse: 0.030104610980376922 baseline_rmse: 1.3095368006346164 nrmse: 0.022988747598225483 v: rmse: 0.015094024155396957 baseline_rmse: 1.3299997452869836 nrmse: 0.011348892515870377 changed_state_entries: - nodes.record/u.embedder.embeddings.weight - nodes.record/u.embedder.counter.counts - nodes.record/u.embedder.linear.weight - nodes.record/u.embedder.linear.bias - nodes.record/u.embedder.scalar.weight - nodes.record/u.embedder.normalizer.mean - nodes.record/u.embedder.normalizer.var - nodes.record/u.embedder.normalizer.count - nodes.record/u.decoder.context_projection.0.weight - nodes.record/u.decoder.context_projection.0.bias - nodes.record/u.decoder.context_projection.1.weight - nodes.record/u.decoder.context_projection.1.bias - nodes.record/u.decoder.context_projection.3.weight - nodes.record/u.decoder.context_projection.3.bias - nodes.record/u.decoder.pool.queries - nodes.record/u.decoder.pool.blocks.0.attention_norm.weight - nodes.record/u.decoder.pool.blocks.0.attention_norm.bias - nodes.record/u.decoder.pool.blocks.0.ffn_norm.weight - nodes.record/u.decoder.pool.blocks.0.ffn_norm.bias - nodes.record/u.decoder.pool.blocks.0.attention.q_proj.weight - nodes.record/u.decoder.pool.blocks.0.attention.q_proj.bias - nodes.record/u.decoder.pool.blocks.0.attention.k_proj.weight - nodes.record/u.decoder.pool.blocks.0.attention.k_proj.bias - nodes.record/u.decoder.pool.blocks.0.attention.v_proj.weight - nodes.record/u.decoder.pool.blocks.0.attention.v_proj.bias - nodes.record/u.decoder.pool.blocks.0.attention.out_proj.weight - nodes.record/u.decoder.pool.blocks.0.attention.out_proj.bias - nodes.record/u.decoder.pool.blocks.0.ffn.0.weight - nodes.record/u.decoder.pool.blocks.0.ffn.0.bias - nodes.record/u.decoder.pool.blocks.0.ffn.3.weight - nodes.record/u.decoder.pool.blocks.0.ffn.3.bias - nodes.record/u.decoder.pool.norm.weight - nodes.record/u.decoder.pool.norm.bias - nodes.record/u.decoder.classification.weight - nodes.record/u.decoder.classification.bias - nodes.record/u.decoder.regression.weight - nodes.record/u.decoder.regression.bias initial_prediction_drift: u: 3.157834827899933 v: 0.0 steps: 256 optimizer_covers_current_parameters: true validation_curve: - step: 0 scores: u: rmse: 1.4709752782012806 baseline_rmse: 1.3218992662289384 nrmse: 1.112774109026946 v: rmse: 0.024370825129755812 baseline_rmse: 1.3013515850804187 nrmse: 0.018727318127675533 - step: 32 scores: u: rmse: 0.07010150875836921 baseline_rmse: 1.3218992662289384 nrmse: 0.05303090072691545 v: rmse: 0.11446097666958786 baseline_rmse: 1.3013515850804187 nrmse: 0.08795545952519403 - step: 128 scores: u: rmse: 0.033955746317564664 baseline_rmse: 1.3218992662289384 nrmse: 0.025687090677062152 v: rmse: 0.030254737304352235 baseline_rmse: 1.3013515850804187 nrmse: 0.02324870361800236 - step: 256 scores: u: rmse: 0.029946861070816684 baseline_rmse: 1.3218992662289384 nrmse: 0.02265441992130603 v: rmse: 0.014465925603878202 baseline_rmse: 1.3013515850804187 nrmse: 0.011116077906789703 reset_head_updated_parameters: - embedder.embeddings.weight - embedder.linear.weight - embedder.scalar.weight - decoder.context_projection.0.weight - decoder.context_projection.0.bias - decoder.context_projection.1.weight - decoder.context_projection.1.bias - decoder.context_projection.3.weight - decoder.context_projection.3.bias - decoder.pool.queries - decoder.pool.blocks.0.attention_norm.weight - decoder.pool.blocks.0.attention_norm.bias - decoder.pool.blocks.0.ffn_norm.weight - decoder.pool.blocks.0.ffn_norm.bias - decoder.pool.blocks.0.attention.q_proj.weight - decoder.pool.blocks.0.attention.q_proj.bias - decoder.pool.blocks.0.attention.k_proj.weight - decoder.pool.blocks.0.attention.k_proj.bias - decoder.pool.blocks.0.attention.v_proj.weight - decoder.pool.blocks.0.attention.v_proj.bias - decoder.pool.blocks.0.attention.out_proj.weight - decoder.pool.blocks.0.attention.out_proj.bias - decoder.pool.blocks.0.ffn.0.weight - decoder.pool.blocks.0.ffn.0.bias - decoder.pool.blocks.0.ffn.3.weight - decoder.pool.blocks.0.ffn.3.bias - decoder.pool.norm.weight - decoder.pool.norm.bias - decoder.classification.weight - decoder.classification.bias - decoder.regression.weight - decoder.regression.bias continuation: before: u: rmse: 0.03486898976137349 baseline_rmse: 1.3095368006346164 nrmse: 0.026626964392658214 v: rmse: 0.023913129355889483 baseline_rmse: 1.3299997452869836 nrmse: 0.01797979995156283 after: u: rmse: 0.011157568747554308 baseline_rmse: 1.3095368006346164 nrmse: 0.008520240700488313 v: rmse: 0.03783177597342966 baseline_rmse: 1.3299997452869836 nrmse: 0.028444949788517762 changed_state_entries: [] initial_prediction_drift: u: 0.0 v: 0.0 steps: 256 optimizer_covers_current_parameters: true validation_curve: - step: 0 scores: u: rmse: 0.03575045788708082 baseline_rmse: 1.3218992662289384 nrmse: 0.02704476717735709 v: rmse: 0.024370825129755812 baseline_rmse: 1.3013515850804187 nrmse: 0.018727318127675533 - step: 32 scores: u: rmse: 0.08799193586778015 baseline_rmse: 1.3218992662289384 nrmse: 0.06656478153498038 v: rmse: 0.12242212700592879 baseline_rmse: 1.3013515850804187 nrmse: 0.09407306096942554 - step: 128 scores: u: rmse: 0.010680149996980285 baseline_rmse: 1.3218992662289384 nrmse: 0.008079397780020101 v: rmse: 0.010780049453845484 baseline_rmse: 1.3013515850804187 nrmse: 0.008283733295010599 - step: 256 scores: u: rmse: 0.01178134751345195 baseline_rmse: 1.3218992662289384 nrmse: 0.008912439710373173 v: rmse: 0.037828475979184796 baseline_rmse: 1.3013515850804187 nrmse: 0.029068605604262693 complete_reset: before: u: rmse: 1.2697078081789366 baseline_rmse: 1.3095368006346164 nrmse: 0.9695854347610711 v: rmse: 1.4672726032015133 baseline_rmse: 1.3299997452869836 nrmse: 1.1032126948903358 after: u: rmse: 0.02626351852449911 baseline_rmse: 1.3095368006346164 nrmse: 0.02005557882128361 v: rmse: 0.04900859171262993 baseline_rmse: 1.3299997452869836 nrmse: 0.03684857225446686 changed_state_entries: - nodes.record/a.embedder.embeddings.weight - nodes.record/a.embedder.counter.counts - nodes.record/a.embedder.linear.weight - nodes.record/a.embedder.linear.bias - nodes.record/a.embedder.scalar.weight - nodes.record/a.embedder.normalizer.mean - nodes.record/a.embedder.normalizer.var - nodes.record/a.embedder.normalizer.count - nodes.record/b.embedder.embeddings.weight - nodes.record/b.embedder.counter.counts - nodes.record/b.embedder.linear.weight - nodes.record/b.embedder.linear.bias - nodes.record/b.embedder.scalar.weight - nodes.record/b.embedder.normalizer.mean - nodes.record/b.embedder.normalizer.var - nodes.record/b.embedder.normalizer.count - nodes.record/u.embedder.embeddings.weight - nodes.record/u.embedder.counter.counts - nodes.record/u.embedder.linear.weight - nodes.record/u.embedder.linear.bias - nodes.record/u.embedder.scalar.weight - nodes.record/u.embedder.normalizer.mean - nodes.record/u.embedder.normalizer.var - nodes.record/u.embedder.normalizer.count - nodes.record/u.decoder.context_projection.0.weight - nodes.record/u.decoder.context_projection.0.bias - nodes.record/u.decoder.context_projection.1.weight - nodes.record/u.decoder.context_projection.1.bias - nodes.record/u.decoder.context_projection.3.weight - nodes.record/u.decoder.context_projection.3.bias - nodes.record/u.decoder.pool.queries - nodes.record/u.decoder.pool.blocks.0.attention_norm.weight - nodes.record/u.decoder.pool.blocks.0.attention_norm.bias - nodes.record/u.decoder.pool.blocks.0.ffn_norm.weight - nodes.record/u.decoder.pool.blocks.0.ffn_norm.bias - nodes.record/u.decoder.pool.blocks.0.attention.q_proj.weight - nodes.record/u.decoder.pool.blocks.0.attention.q_proj.bias - nodes.record/u.decoder.pool.blocks.0.attention.k_proj.weight - nodes.record/u.decoder.pool.blocks.0.attention.k_proj.bias - nodes.record/u.decoder.pool.blocks.0.attention.v_proj.weight - nodes.record/u.decoder.pool.blocks.0.attention.v_proj.bias - nodes.record/u.decoder.pool.blocks.0.attention.out_proj.weight - nodes.record/u.decoder.pool.blocks.0.attention.out_proj.bias - nodes.record/u.decoder.pool.blocks.0.ffn.0.weight - nodes.record/u.decoder.pool.blocks.0.ffn.0.bias - nodes.record/u.decoder.pool.blocks.0.ffn.3.weight - nodes.record/u.decoder.pool.blocks.0.ffn.3.bias - nodes.record/u.decoder.pool.norm.weight - nodes.record/u.decoder.pool.norm.bias - nodes.record/u.decoder.classification.weight - nodes.record/u.decoder.classification.bias - nodes.record/u.decoder.regression.weight - nodes.record/u.decoder.regression.bias - nodes.record/v.embedder.embeddings.weight - nodes.record/v.embedder.counter.counts - nodes.record/v.embedder.linear.weight - nodes.record/v.embedder.linear.bias - nodes.record/v.embedder.scalar.weight - nodes.record/v.embedder.normalizer.mean - nodes.record/v.embedder.normalizer.var - nodes.record/v.embedder.normalizer.count - nodes.record/v.decoder.context_projection.0.weight - nodes.record/v.decoder.context_projection.0.bias - nodes.record/v.decoder.context_projection.1.weight - nodes.record/v.decoder.context_projection.1.bias - nodes.record/v.decoder.context_projection.3.weight - nodes.record/v.decoder.context_projection.3.bias - nodes.record/v.decoder.pool.queries - nodes.record/v.decoder.pool.blocks.0.attention_norm.weight - nodes.record/v.decoder.pool.blocks.0.attention_norm.bias - nodes.record/v.decoder.pool.blocks.0.ffn_norm.weight - nodes.record/v.decoder.pool.blocks.0.ffn_norm.bias - nodes.record/v.decoder.pool.blocks.0.attention.q_proj.weight - nodes.record/v.decoder.pool.blocks.0.attention.q_proj.bias - nodes.record/v.decoder.pool.blocks.0.attention.k_proj.weight - nodes.record/v.decoder.pool.blocks.0.attention.k_proj.bias - nodes.record/v.decoder.pool.blocks.0.attention.v_proj.weight - nodes.record/v.decoder.pool.blocks.0.attention.v_proj.bias - nodes.record/v.decoder.pool.blocks.0.attention.out_proj.weight - nodes.record/v.decoder.pool.blocks.0.attention.out_proj.bias - nodes.record/v.decoder.pool.blocks.0.ffn.0.weight - nodes.record/v.decoder.pool.blocks.0.ffn.0.bias - nodes.record/v.decoder.pool.blocks.0.ffn.3.weight - nodes.record/v.decoder.pool.blocks.0.ffn.3.bias - nodes.record/v.decoder.pool.norm.weight - nodes.record/v.decoder.pool.norm.bias - nodes.record/v.decoder.classification.weight - nodes.record/v.decoder.classification.bias - nodes.record/v.decoder.regression.weight - nodes.record/v.decoder.regression.bias - nodes.record.encoder.encoder.0.attention_norm.weight - nodes.record.encoder.encoder.0.attention_norm.bias - nodes.record.encoder.encoder.0.ffn_norm.weight - nodes.record.encoder.encoder.0.ffn_norm.bias - nodes.record.encoder.encoder.0.attention.q_proj.weight - nodes.record.encoder.encoder.0.attention.q_proj.bias - nodes.record.encoder.encoder.0.attention.k_proj.weight - nodes.record.encoder.encoder.0.attention.k_proj.bias - nodes.record.encoder.encoder.0.attention.v_proj.weight - nodes.record.encoder.encoder.0.attention.v_proj.bias - nodes.record.encoder.encoder.0.attention.out_proj.weight - nodes.record.encoder.encoder.0.attention.out_proj.bias - nodes.record.encoder.encoder.0.ffn.0.weight - nodes.record.encoder.encoder.0.ffn.0.bias - nodes.record.encoder.encoder.0.ffn.3.weight - nodes.record.encoder.encoder.0.ffn.3.bias - nodes.record.encoder.coordinate_encoder.attention_norm.weight - nodes.record.encoder.coordinate_encoder.attention_norm.bias - nodes.record.encoder.coordinate_encoder.ffn_norm.weight - nodes.record.encoder.coordinate_encoder.ffn_norm.bias - nodes.record.encoder.coordinate_encoder.attention.q_proj.weight - nodes.record.encoder.coordinate_encoder.attention.q_proj.bias - nodes.record.encoder.coordinate_encoder.attention.k_proj.weight - nodes.record.encoder.coordinate_encoder.attention.k_proj.bias - nodes.record.encoder.coordinate_encoder.attention.v_proj.weight - nodes.record.encoder.coordinate_encoder.attention.v_proj.bias - nodes.record.encoder.coordinate_encoder.attention.out_proj.weight - nodes.record.encoder.coordinate_encoder.attention.out_proj.bias - nodes.record.encoder.coordinate_encoder.ffn.0.weight - nodes.record.encoder.coordinate_encoder.ffn.0.bias - nodes.record.encoder.coordinate_encoder.ffn.3.weight - nodes.record.encoder.coordinate_encoder.ffn.3.bias - nodes.record.encoder.pool.queries - nodes.record.encoder.pool.mass_direction - nodes.record.encoder.pool.blocks.0.attention_norm.weight - nodes.record.encoder.pool.blocks.0.attention_norm.bias - nodes.record.encoder.pool.blocks.0.ffn_norm.weight - nodes.record.encoder.pool.blocks.0.ffn_norm.bias - nodes.record.encoder.pool.blocks.0.attention.q_proj.weight - nodes.record.encoder.pool.blocks.0.attention.q_proj.bias - nodes.record.encoder.pool.blocks.0.attention.k_proj.weight - nodes.record.encoder.pool.blocks.0.attention.k_proj.bias - nodes.record.encoder.pool.blocks.0.attention.v_proj.weight - nodes.record.encoder.pool.blocks.0.attention.v_proj.bias - nodes.record.encoder.pool.blocks.0.attention.out_proj.weight - nodes.record.encoder.pool.blocks.0.attention.out_proj.bias - nodes.record.encoder.pool.blocks.0.ffn.0.weight - nodes.record.encoder.pool.blocks.0.ffn.0.bias - nodes.record.encoder.pool.blocks.0.ffn.3.weight - nodes.record.encoder.pool.blocks.0.ffn.3.bias - nodes.record.encoder.pool.norm.weight - nodes.record.encoder.pool.norm.bias - nodes.record.encoder.pool.mass_projection.weight initial_prediction_drift: u: 3.0307161808013916 v: 3.276441514492035 steps: 256 optimizer_covers_current_parameters: true validation_curve: - step: 0 scores: u: rmse: 1.2645540796908759 baseline_rmse: 1.3218992662289384 nrmse: 0.9566190949620128 v: rmse: 1.445354731245329 baseline_rmse: 1.3013515850804187 nrmse: 1.1106566033467515 - step: 32 scores: u: rmse: 0.09326570235963688 baseline_rmse: 1.3218992662289384 nrmse: 0.07055431888218046 v: rmse: 0.07710958198337366 baseline_rmse: 1.3013515850804187 nrmse: 0.059253458379280775 - step: 128 scores: u: rmse: 0.0348029081713234 baseline_rmse: 1.3218992662289384 nrmse: 0.026327957856129045 v: rmse: 0.050016179758700076 baseline_rmse: 1.3013515850804187 nrmse: 0.038434025310392396 - step: 256 scores: u: rmse: 0.025879687118018732 baseline_rmse: 1.3218992662289384 nrmse: 0.019577654499988696 v: rmse: 0.04897061230943835 baseline_rmse: 1.3013515850804187 nrmse: 0.03763057798589621 test_rows: 1024 baseline: Source-training mean for test comparisons; phase-training mean for validation curves optimizer_policy: New AdamW factory and Trainer for every fit; both hidden labels rehearsed mutation: reset record/u; complete-reset control resets record with descendants=True checks: Source learns both targets below 0.25 nRMSE: true Reset changes selected learned state: true Reset preserves every unselected state entry: true Reset preserves the schema: true Selected target immediately loses its skill: true Unselected target immediately preserves predictions: true 'selective_reset: fresh optimizer covers current parameters': true 'selective_reset: both final targets below 0.25 nRMSE': true Reset head parameters actually relearn: true Hidden target placeholders cannot affect predictions: true Relearned checkpoint preserves schema and state: true Relearned checkpoint preserves predictions: true 'continuation: fresh optimizer covers current parameters': true 'continuation: both final targets below 0.25 nRMSE': true Complete reset loses both learned tasks: true 'complete_reset: fresh optimizer covers current parameters': true 'complete_reset: both final targets below 0.25 nRMSE': true outcome: met duration_seconds: 35.627 - started_at: '2026-09-15T18:01:31.237558+00:00' mode: full seed: 7101 steps_override: null accelerator: gpu threads: 1 provenance: revision: null source_sha256: 2252cee7efe14048db1252663050b74fce7ce20d11f143c1b24b93df622d6f7c script_sha256: 84de49318b9298b52f63fe1743ebcf79674875911b46ffd1222e86ea2521027e code_sha256: f7ad2a7bee509149b1c4bad0f6392d223f7a873be15ca244b94bae386803a02e python: 3.12.6 platform: Linux-6.8.0-138-generic-x86_64-with-glibc2.35 host: lab relflow_path: /home/grantham/relflow-mutations-v3_tgit_/src/relflow packages: relflow: 0.2.2 torch: 2.12.0 lightning: 2.6.4 numpy: 2.4.6 pyarrow: 24.0.0 polars: 1.44.1 torchmetrics: 1.9.0 tensordict: 0.12.4 pydantic: 2.13.4 cuda: '13.0' gpu: NVIDIA GeForce RTX 3090 metrics: source: u: rmse: 0.009796106068409562 baseline_rmse: 1.2908974940031006 nrmse: 0.007588601042234289 v: rmse: 0.019397250921355053 baseline_rmse: 1.2969480983896875 nrmse: 0.01495607337366777 source_fit: steps: 512 optimizer_covers_current_parameters: true validation_curve: - step: 0 scores: u: rmse: 1.5771062482408669 baseline_rmse: 1.214202473192484 nrmse: 1.2988824212276602 v: rmse: 1.3072694637116182 baseline_rmse: 1.3285155523136438 nrmse: 0.9840076478103512 - step: 32 scores: u: rmse: 0.10717741666818707 baseline_rmse: 1.214202473192484 nrmse: 0.08826980592980273 v: rmse: 0.10111543375960332 baseline_rmse: 1.3285155523136438 nrmse: 0.07611159205739684 - step: 128 scores: u: rmse: 0.017833300961804754 baseline_rmse: 1.214202473192484 nrmse: 0.014687254684069229 v: rmse: 0.022416444336445907 baseline_rmse: 1.3285155523136438 nrmse: 0.016873302158493437 - step: 512 scores: u: rmse: 0.009539652227979903 baseline_rmse: 1.214202473192484 nrmse: 0.007856722777789638 v: rmse: 0.01895260950991225 baseline_rmse: 1.3285155523136438 nrmse: 0.014266004998515672 source_prerequisite_met: true downstream_interpretable: true arms: selective_reset: before: u: rmse: 1.609977754081612 baseline_rmse: 1.2908974940031006 nrmse: 1.2471770698764288 v: rmse: 0.019397250921355053 baseline_rmse: 1.2969480983896875 nrmse: 0.01495607337366777 after: u: rmse: 0.04632457690965369 baseline_rmse: 1.2908974940031006 nrmse: 0.03588555801282114 v: rmse: 0.015373638433522189 baseline_rmse: 1.2969480983896875 nrmse: 0.011853703669877272 changed_state_entries: - nodes.record/u.embedder.embeddings.weight - nodes.record/u.embedder.counter.counts - nodes.record/u.embedder.linear.weight - nodes.record/u.embedder.linear.bias - nodes.record/u.embedder.scalar.weight - nodes.record/u.embedder.normalizer.mean - nodes.record/u.embedder.normalizer.var - nodes.record/u.embedder.normalizer.count - nodes.record/u.decoder.context_projection.0.weight - nodes.record/u.decoder.context_projection.0.bias - nodes.record/u.decoder.context_projection.1.weight - nodes.record/u.decoder.context_projection.1.bias - nodes.record/u.decoder.context_projection.3.weight - nodes.record/u.decoder.context_projection.3.bias - nodes.record/u.decoder.pool.queries - nodes.record/u.decoder.pool.blocks.0.attention_norm.weight - nodes.record/u.decoder.pool.blocks.0.attention_norm.bias - nodes.record/u.decoder.pool.blocks.0.ffn_norm.weight - nodes.record/u.decoder.pool.blocks.0.ffn_norm.bias - nodes.record/u.decoder.pool.blocks.0.attention.q_proj.weight - nodes.record/u.decoder.pool.blocks.0.attention.q_proj.bias - nodes.record/u.decoder.pool.blocks.0.attention.k_proj.weight - nodes.record/u.decoder.pool.blocks.0.attention.k_proj.bias - nodes.record/u.decoder.pool.blocks.0.attention.v_proj.weight - nodes.record/u.decoder.pool.blocks.0.attention.v_proj.bias - nodes.record/u.decoder.pool.blocks.0.attention.out_proj.weight - nodes.record/u.decoder.pool.blocks.0.attention.out_proj.bias - nodes.record/u.decoder.pool.blocks.0.ffn.0.weight - nodes.record/u.decoder.pool.blocks.0.ffn.0.bias - nodes.record/u.decoder.pool.blocks.0.ffn.3.weight - nodes.record/u.decoder.pool.blocks.0.ffn.3.bias - nodes.record/u.decoder.pool.norm.weight - nodes.record/u.decoder.pool.norm.bias - nodes.record/u.decoder.classification.weight - nodes.record/u.decoder.classification.bias - nodes.record/u.decoder.regression.weight - nodes.record/u.decoder.regression.bias initial_prediction_drift: u: 3.6637556552886963 v: 0.0 steps: 256 optimizer_covers_current_parameters: true validation_curve: - step: 0 scores: u: rmse: 1.5776053052517884 baseline_rmse: 1.2520463239225392 nrmse: 1.2600215144670563 v: rmse: 0.01851867432477028 baseline_rmse: 1.3128148986756512 nrmse: 0.014106081781560868 - step: 32 scores: u: rmse: 0.09245453226491651 baseline_rmse: 1.2520463239225392 nrmse: 0.07384274087820127 v: rmse: 0.029584106461261212 baseline_rmse: 1.3128148986756512 nrmse: 0.022534864961621957 - step: 128 scores: u: rmse: 0.05615155865109741 baseline_rmse: 1.2520463239225392 nrmse: 0.04484782837361803 v: rmse: 0.012019017248036383 baseline_rmse: 1.3128148986756512 nrmse: 0.009155149945480506 - step: 256 scores: u: rmse: 0.04690964108847677 baseline_rmse: 1.2520463239225392 nrmse: 0.037466378193990006 v: rmse: 0.01498911598433781 baseline_rmse: 1.3128148986756512 nrmse: 0.011417539517154029 reset_head_updated_parameters: - embedder.embeddings.weight - embedder.linear.weight - embedder.scalar.weight - decoder.context_projection.0.weight - decoder.context_projection.0.bias - decoder.context_projection.1.weight - decoder.context_projection.1.bias - decoder.context_projection.3.weight - decoder.context_projection.3.bias - decoder.pool.queries - decoder.pool.blocks.0.attention_norm.weight - decoder.pool.blocks.0.attention_norm.bias - decoder.pool.blocks.0.ffn_norm.weight - decoder.pool.blocks.0.ffn_norm.bias - decoder.pool.blocks.0.attention.q_proj.weight - decoder.pool.blocks.0.attention.q_proj.bias - decoder.pool.blocks.0.attention.k_proj.weight - decoder.pool.blocks.0.attention.k_proj.bias - decoder.pool.blocks.0.attention.v_proj.weight - decoder.pool.blocks.0.attention.v_proj.bias - decoder.pool.blocks.0.attention.out_proj.weight - decoder.pool.blocks.0.attention.out_proj.bias - decoder.pool.blocks.0.ffn.0.weight - decoder.pool.blocks.0.ffn.0.bias - decoder.pool.blocks.0.ffn.3.weight - decoder.pool.blocks.0.ffn.3.bias - decoder.pool.norm.weight - decoder.pool.norm.bias - decoder.classification.weight - decoder.classification.bias - decoder.regression.weight - decoder.regression.bias continuation: before: u: rmse: 0.009796106068409562 baseline_rmse: 1.2908974940031006 nrmse: 0.007588601042234289 v: rmse: 0.019397250921355053 baseline_rmse: 1.2969480983896875 nrmse: 0.01495607337366777 after: u: rmse: 0.014894116016718029 baseline_rmse: 1.2908974940031006 nrmse: 0.011537799155943093 v: rmse: 0.03418484117031133 baseline_rmse: 1.2969480983896875 nrmse: 0.02635790993699424 changed_state_entries: [] initial_prediction_drift: u: 0.0 v: 0.0 steps: 256 optimizer_covers_current_parameters: true validation_curve: - step: 0 scores: u: rmse: 0.009405761385961659 baseline_rmse: 1.2520463239225392 nrmse: 0.007512311011380413 v: rmse: 0.01851867432477028 baseline_rmse: 1.3128148986756512 nrmse: 0.014106081781560868 - step: 32 scores: u: rmse: 0.08427936048393028 baseline_rmse: 1.2520463239225392 nrmse: 0.06731329254647005 v: rmse: 0.061933053385872454 baseline_rmse: 1.3128148986756512 nrmse: 0.04717576975120379 - step: 128 scores: u: rmse: 0.028769483507834138 baseline_rmse: 1.2520463239225392 nrmse: 0.02297797050967104 v: rmse: 0.052134352761181266 baseline_rmse: 1.3128148986756512 nrmse: 0.03971188384118252 - step: 256 scores: u: rmse: 0.01376603956676995 baseline_rmse: 1.2520463239225392 nrmse: 0.010994832462462162 v: rmse: 0.035074513265269745 baseline_rmse: 1.3128148986756512 nrmse: 0.026717028653965164 complete_reset: before: u: rmse: 1.4410291398560169 baseline_rmse: 1.2908974940031006 nrmse: 1.116300207065516 v: rmse: 1.4272504185751518 baseline_rmse: 1.2969480983896875 nrmse: 1.1004684153107205 after: u: rmse: 0.019270368654935185 baseline_rmse: 1.2908974940031006 nrmse: 0.01492788447142876 v: rmse: 0.03453584099338142 baseline_rmse: 1.2969480983896875 nrmse: 0.026628545148615967 changed_state_entries: - nodes.record/a.embedder.embeddings.weight - nodes.record/a.embedder.counter.counts - nodes.record/a.embedder.linear.weight - nodes.record/a.embedder.linear.bias - nodes.record/a.embedder.scalar.weight - nodes.record/a.embedder.normalizer.mean - nodes.record/a.embedder.normalizer.var - nodes.record/a.embedder.normalizer.count - nodes.record/b.embedder.embeddings.weight - nodes.record/b.embedder.counter.counts - nodes.record/b.embedder.linear.weight - nodes.record/b.embedder.linear.bias - nodes.record/b.embedder.scalar.weight - nodes.record/b.embedder.normalizer.mean - nodes.record/b.embedder.normalizer.var - nodes.record/b.embedder.normalizer.count - nodes.record/u.embedder.embeddings.weight - nodes.record/u.embedder.counter.counts - nodes.record/u.embedder.linear.weight - nodes.record/u.embedder.linear.bias - nodes.record/u.embedder.scalar.weight - nodes.record/u.embedder.normalizer.mean - nodes.record/u.embedder.normalizer.var - nodes.record/u.embedder.normalizer.count - nodes.record/u.decoder.context_projection.0.weight - nodes.record/u.decoder.context_projection.0.bias - nodes.record/u.decoder.context_projection.1.weight - nodes.record/u.decoder.context_projection.1.bias - nodes.record/u.decoder.context_projection.3.weight - nodes.record/u.decoder.context_projection.3.bias - nodes.record/u.decoder.pool.queries - nodes.record/u.decoder.pool.blocks.0.attention_norm.weight - nodes.record/u.decoder.pool.blocks.0.attention_norm.bias - nodes.record/u.decoder.pool.blocks.0.ffn_norm.weight - nodes.record/u.decoder.pool.blocks.0.ffn_norm.bias - nodes.record/u.decoder.pool.blocks.0.attention.q_proj.weight - nodes.record/u.decoder.pool.blocks.0.attention.q_proj.bias - nodes.record/u.decoder.pool.blocks.0.attention.k_proj.weight - nodes.record/u.decoder.pool.blocks.0.attention.k_proj.bias - nodes.record/u.decoder.pool.blocks.0.attention.v_proj.weight - nodes.record/u.decoder.pool.blocks.0.attention.v_proj.bias - nodes.record/u.decoder.pool.blocks.0.attention.out_proj.weight - nodes.record/u.decoder.pool.blocks.0.attention.out_proj.bias - nodes.record/u.decoder.pool.blocks.0.ffn.0.weight - nodes.record/u.decoder.pool.blocks.0.ffn.0.bias - nodes.record/u.decoder.pool.blocks.0.ffn.3.weight - nodes.record/u.decoder.pool.blocks.0.ffn.3.bias - nodes.record/u.decoder.pool.norm.weight - nodes.record/u.decoder.pool.norm.bias - nodes.record/u.decoder.classification.weight - nodes.record/u.decoder.classification.bias - nodes.record/u.decoder.regression.weight - nodes.record/u.decoder.regression.bias - nodes.record/v.embedder.embeddings.weight - nodes.record/v.embedder.counter.counts - nodes.record/v.embedder.linear.weight - nodes.record/v.embedder.linear.bias - nodes.record/v.embedder.scalar.weight - nodes.record/v.embedder.normalizer.mean - nodes.record/v.embedder.normalizer.var - nodes.record/v.embedder.normalizer.count - nodes.record/v.decoder.context_projection.0.weight - nodes.record/v.decoder.context_projection.0.bias - nodes.record/v.decoder.context_projection.1.weight - nodes.record/v.decoder.context_projection.1.bias - nodes.record/v.decoder.context_projection.3.weight - nodes.record/v.decoder.context_projection.3.bias - nodes.record/v.decoder.pool.queries - nodes.record/v.decoder.pool.blocks.0.attention_norm.weight - nodes.record/v.decoder.pool.blocks.0.attention_norm.bias - nodes.record/v.decoder.pool.blocks.0.ffn_norm.weight - nodes.record/v.decoder.pool.blocks.0.ffn_norm.bias - nodes.record/v.decoder.pool.blocks.0.attention.q_proj.weight - nodes.record/v.decoder.pool.blocks.0.attention.q_proj.bias - nodes.record/v.decoder.pool.blocks.0.attention.k_proj.weight - nodes.record/v.decoder.pool.blocks.0.attention.k_proj.bias - nodes.record/v.decoder.pool.blocks.0.attention.v_proj.weight - nodes.record/v.decoder.pool.blocks.0.attention.v_proj.bias - nodes.record/v.decoder.pool.blocks.0.attention.out_proj.weight - nodes.record/v.decoder.pool.blocks.0.attention.out_proj.bias - nodes.record/v.decoder.pool.blocks.0.ffn.0.weight - nodes.record/v.decoder.pool.blocks.0.ffn.0.bias - nodes.record/v.decoder.pool.blocks.0.ffn.3.weight - nodes.record/v.decoder.pool.blocks.0.ffn.3.bias - nodes.record/v.decoder.pool.norm.weight - nodes.record/v.decoder.pool.norm.bias - nodes.record/v.decoder.classification.weight - nodes.record/v.decoder.classification.bias - nodes.record/v.decoder.regression.weight - nodes.record/v.decoder.regression.bias - nodes.record.encoder.encoder.0.attention_norm.weight - nodes.record.encoder.encoder.0.attention_norm.bias - nodes.record.encoder.encoder.0.ffn_norm.weight - nodes.record.encoder.encoder.0.ffn_norm.bias - nodes.record.encoder.encoder.0.attention.q_proj.weight - nodes.record.encoder.encoder.0.attention.q_proj.bias - nodes.record.encoder.encoder.0.attention.k_proj.weight - nodes.record.encoder.encoder.0.attention.k_proj.bias - nodes.record.encoder.encoder.0.attention.v_proj.weight - nodes.record.encoder.encoder.0.attention.v_proj.bias - nodes.record.encoder.encoder.0.attention.out_proj.weight - nodes.record.encoder.encoder.0.attention.out_proj.bias - nodes.record.encoder.encoder.0.ffn.0.weight - nodes.record.encoder.encoder.0.ffn.0.bias - nodes.record.encoder.encoder.0.ffn.3.weight - nodes.record.encoder.encoder.0.ffn.3.bias - nodes.record.encoder.coordinate_encoder.attention_norm.weight - nodes.record.encoder.coordinate_encoder.attention_norm.bias - nodes.record.encoder.coordinate_encoder.ffn_norm.weight - nodes.record.encoder.coordinate_encoder.ffn_norm.bias - nodes.record.encoder.coordinate_encoder.attention.q_proj.weight - nodes.record.encoder.coordinate_encoder.attention.q_proj.bias - nodes.record.encoder.coordinate_encoder.attention.k_proj.weight - nodes.record.encoder.coordinate_encoder.attention.k_proj.bias - nodes.record.encoder.coordinate_encoder.attention.v_proj.weight - nodes.record.encoder.coordinate_encoder.attention.v_proj.bias - nodes.record.encoder.coordinate_encoder.attention.out_proj.weight - nodes.record.encoder.coordinate_encoder.attention.out_proj.bias - nodes.record.encoder.coordinate_encoder.ffn.0.weight - nodes.record.encoder.coordinate_encoder.ffn.0.bias - nodes.record.encoder.coordinate_encoder.ffn.3.weight - nodes.record.encoder.coordinate_encoder.ffn.3.bias - nodes.record.encoder.pool.queries - nodes.record.encoder.pool.mass_direction - nodes.record.encoder.pool.blocks.0.attention_norm.weight - nodes.record.encoder.pool.blocks.0.attention_norm.bias - nodes.record.encoder.pool.blocks.0.ffn_norm.weight - nodes.record.encoder.pool.blocks.0.ffn_norm.bias - nodes.record.encoder.pool.blocks.0.attention.q_proj.weight - nodes.record.encoder.pool.blocks.0.attention.q_proj.bias - nodes.record.encoder.pool.blocks.0.attention.k_proj.weight - nodes.record.encoder.pool.blocks.0.attention.k_proj.bias - nodes.record.encoder.pool.blocks.0.attention.v_proj.weight - nodes.record.encoder.pool.blocks.0.attention.v_proj.bias - nodes.record.encoder.pool.blocks.0.attention.out_proj.weight - nodes.record.encoder.pool.blocks.0.attention.out_proj.bias - nodes.record.encoder.pool.blocks.0.ffn.0.weight - nodes.record.encoder.pool.blocks.0.ffn.0.bias - nodes.record.encoder.pool.blocks.0.ffn.3.weight - nodes.record.encoder.pool.blocks.0.ffn.3.bias - nodes.record.encoder.pool.norm.weight - nodes.record.encoder.pool.norm.bias - nodes.record.encoder.pool.mass_projection.weight initial_prediction_drift: u: 3.4898728728294373 v: 3.0272326171398163 steps: 256 optimizer_covers_current_parameters: true validation_curve: - step: 0 scores: u: rmse: 1.3874749972975753 baseline_rmse: 1.2520463239225392 nrmse: 1.1081658647826633 v: rmse: 1.4373958297140221 baseline_rmse: 1.3128148986756512 nrmse: 1.0948960368777398 - step: 32 scores: u: rmse: 0.0686681868131216 baseline_rmse: 1.2520463239225392 nrmse: 0.05484476532624676 v: rmse: 0.09046761199796963 baseline_rmse: 1.3128148986756512 nrmse: 0.06891117101826925 - step: 128 scores: u: rmse: 0.02379988831312726 baseline_rmse: 1.2520463239225392 nrmse: 0.019008792133636498 v: rmse: 0.02647999021107485 baseline_rmse: 1.3128148986756512 nrmse: 0.020170391300241553 - step: 256 scores: u: rmse: 0.01781185651776526 baseline_rmse: 1.2520463239225392 nrmse: 0.014226196089903803 v: rmse: 0.034212078750043576 baseline_rmse: 1.3128148986756512 nrmse: 0.02606009330375229 test_rows: 1024 baseline: Source-training mean for test comparisons; phase-training mean for validation curves optimizer_policy: New AdamW factory and Trainer for every fit; both hidden labels rehearsed mutation: reset record/u; complete-reset control resets record with descendants=True checks: Source learns both targets below 0.25 nRMSE: true Reset changes selected learned state: true Reset preserves every unselected state entry: true Reset preserves the schema: true Selected target immediately loses its skill: true Unselected target immediately preserves predictions: true 'selective_reset: fresh optimizer covers current parameters': true 'selective_reset: both final targets below 0.25 nRMSE': true Reset head parameters actually relearn: true Hidden target placeholders cannot affect predictions: true Relearned checkpoint preserves schema and state: true Relearned checkpoint preserves predictions: true 'continuation: fresh optimizer covers current parameters': true 'continuation: both final targets below 0.25 nRMSE': true Complete reset loses both learned tasks: true 'complete_reset: fresh optimizer covers current parameters': true 'complete_reset: both final targets below 0.25 nRMSE': true outcome: met duration_seconds: 35.489 - started_at: '2026-09-15T18:03:47.205739+00:00' mode: full seed: 7102 steps_override: null accelerator: gpu threads: 1 provenance: revision: null source_sha256: 2252cee7efe14048db1252663050b74fce7ce20d11f143c1b24b93df622d6f7c script_sha256: 84de49318b9298b52f63fe1743ebcf79674875911b46ffd1222e86ea2521027e code_sha256: f7ad2a7bee509149b1c4bad0f6392d223f7a873be15ca244b94bae386803a02e python: 3.12.6 platform: Linux-6.8.0-138-generic-x86_64-with-glibc2.35 host: lab relflow_path: /home/grantham/relflow-mutations-v3_tgit_/src/relflow packages: relflow: 0.2.2 torch: 2.12.0 lightning: 2.6.4 numpy: 2.4.6 pyarrow: 24.0.0 polars: 1.44.1 torchmetrics: 1.9.0 tensordict: 0.12.4 pydantic: 2.13.4 cuda: '13.0' gpu: NVIDIA GeForce RTX 3090 metrics: source: u: rmse: 0.06565498347536482 baseline_rmse: 1.2967289998405087 nrmse: 0.050631229411419086 v: rmse: 0.041381249939003356 baseline_rmse: 1.2693145185962316 nrmse: 0.03260125787008879 source_fit: steps: 512 optimizer_covers_current_parameters: true validation_curve: - step: 0 scores: u: rmse: 1.2393028237273997 baseline_rmse: 1.2723746695686682 nrmse: 0.9740077772434124 v: rmse: 1.7021467221439637 baseline_rmse: 1.3468533739051283 nrmse: 1.2637951206289677 - step: 32 scores: u: rmse: 0.10355146261742668 baseline_rmse: 1.2723746695686682 nrmse: 0.08138441065675282 v: rmse: 0.08961803685159271 baseline_rmse: 1.3468533739051283 nrmse: 0.06653882195932738 - step: 128 scores: u: rmse: 0.03978728611733003 baseline_rmse: 1.2723746695686682 nrmse: 0.03127010232828497 v: rmse: 0.04909547654677152 baseline_rmse: 1.3468533739051283 nrmse: 0.036451983191326795 - step: 512 scores: u: rmse: 0.06498414801296525 baseline_rmse: 1.2723746695686682 nrmse: 0.05107312300943142 v: rmse: 0.03994279807800749 baseline_rmse: 1.3468533739051283 nrmse: 0.029656381943192166 source_prerequisite_met: true downstream_interpretable: true arms: selective_reset: before: u: rmse: 1.2489358540227307 baseline_rmse: 1.2967289998405087 nrmse: 0.9631433045581179 v: rmse: 0.041381249939003356 baseline_rmse: 1.2693145185962316 nrmse: 0.03260125787008879 after: u: rmse: 0.021251006458372784 baseline_rmse: 1.2967289998405087 nrmse: 0.016388163186746463 v: rmse: 0.012912155587262835 baseline_rmse: 1.2693145185962316 nrmse: 0.010172542264420586 changed_state_entries: - nodes.record/u.embedder.embeddings.weight - nodes.record/u.embedder.counter.counts - nodes.record/u.embedder.linear.weight - nodes.record/u.embedder.linear.bias - nodes.record/u.embedder.scalar.weight - nodes.record/u.embedder.normalizer.mean - nodes.record/u.embedder.normalizer.var - nodes.record/u.embedder.normalizer.count - nodes.record/u.decoder.context_projection.0.weight - nodes.record/u.decoder.context_projection.0.bias - nodes.record/u.decoder.context_projection.1.weight - nodes.record/u.decoder.context_projection.1.bias - nodes.record/u.decoder.context_projection.3.weight - nodes.record/u.decoder.context_projection.3.bias - nodes.record/u.decoder.pool.queries - nodes.record/u.decoder.pool.blocks.0.attention_norm.weight - nodes.record/u.decoder.pool.blocks.0.attention_norm.bias - nodes.record/u.decoder.pool.blocks.0.ffn_norm.weight - nodes.record/u.decoder.pool.blocks.0.ffn_norm.bias - nodes.record/u.decoder.pool.blocks.0.attention.q_proj.weight - nodes.record/u.decoder.pool.blocks.0.attention.q_proj.bias - nodes.record/u.decoder.pool.blocks.0.attention.k_proj.weight - nodes.record/u.decoder.pool.blocks.0.attention.k_proj.bias - nodes.record/u.decoder.pool.blocks.0.attention.v_proj.weight - nodes.record/u.decoder.pool.blocks.0.attention.v_proj.bias - nodes.record/u.decoder.pool.blocks.0.attention.out_proj.weight - nodes.record/u.decoder.pool.blocks.0.attention.out_proj.bias - nodes.record/u.decoder.pool.blocks.0.ffn.0.weight - nodes.record/u.decoder.pool.blocks.0.ffn.0.bias - nodes.record/u.decoder.pool.blocks.0.ffn.3.weight - nodes.record/u.decoder.pool.blocks.0.ffn.3.bias - nodes.record/u.decoder.pool.norm.weight - nodes.record/u.decoder.pool.norm.bias - nodes.record/u.decoder.classification.weight - nodes.record/u.decoder.classification.bias - nodes.record/u.decoder.regression.weight - nodes.record/u.decoder.regression.bias initial_prediction_drift: u: 2.9901226460933685 v: 0.0 steps: 256 optimizer_covers_current_parameters: true validation_curve: - step: 0 scores: u: rmse: 1.2203848159275321 baseline_rmse: 1.2761651462099894 nrmse: 0.9562906646932678 v: rmse: 0.04133891376364024 baseline_rmse: 1.3343993130174792 nrmse: 0.03097941775026884 - step: 32 scores: u: rmse: 0.04211010364844126 baseline_rmse: 1.2761651462099894 nrmse: 0.03299737794399234 v: rmse: 0.06197801814497376 baseline_rmse: 1.3343993130174792 nrmse: 0.04644638043527074 - step: 128 scores: u: rmse: 0.018226766827659342 baseline_rmse: 1.2761651462099894 nrmse: 0.014282451516396592 v: rmse: 0.013220150875833902 baseline_rmse: 1.3343993130174792 nrmse: 0.009907192507420552 - step: 256 scores: u: rmse: 0.02077646370410201 baseline_rmse: 1.2761651462099894 nrmse: 0.016280387977845073 v: rmse: 0.01272842878576654 baseline_rmse: 1.3343993130174792 nrmse: 0.00953869554757468 reset_head_updated_parameters: - embedder.embeddings.weight - embedder.linear.weight - embedder.scalar.weight - decoder.context_projection.0.weight - decoder.context_projection.0.bias - decoder.context_projection.1.weight - decoder.context_projection.1.bias - decoder.context_projection.3.weight - decoder.context_projection.3.bias - decoder.pool.queries - decoder.pool.blocks.0.attention_norm.weight - decoder.pool.blocks.0.attention_norm.bias - decoder.pool.blocks.0.ffn_norm.weight - decoder.pool.blocks.0.ffn_norm.bias - decoder.pool.blocks.0.attention.q_proj.weight - decoder.pool.blocks.0.attention.q_proj.bias - decoder.pool.blocks.0.attention.k_proj.weight - decoder.pool.blocks.0.attention.k_proj.bias - decoder.pool.blocks.0.attention.v_proj.weight - decoder.pool.blocks.0.attention.v_proj.bias - decoder.pool.blocks.0.attention.out_proj.weight - decoder.pool.blocks.0.attention.out_proj.bias - decoder.pool.blocks.0.ffn.0.weight - decoder.pool.blocks.0.ffn.0.bias - decoder.pool.blocks.0.ffn.3.weight - decoder.pool.blocks.0.ffn.3.bias - decoder.pool.norm.weight - decoder.pool.norm.bias - decoder.classification.weight - decoder.classification.bias - decoder.regression.weight - decoder.regression.bias continuation: before: u: rmse: 0.06565498347536482 baseline_rmse: 1.2967289998405087 nrmse: 0.050631229411419086 v: rmse: 0.041381249939003356 baseline_rmse: 1.2693145185962316 nrmse: 0.03260125787008879 after: u: rmse: 0.019977130988254894 baseline_rmse: 1.2967289998405087 nrmse: 0.015405787169649158 v: rmse: 0.00927443784231114 baseline_rmse: 1.2693145185962316 nrmse: 0.007306650720869391 changed_state_entries: [] initial_prediction_drift: u: 0.0 v: 0.0 steps: 256 optimizer_covers_current_parameters: true validation_curve: - step: 0 scores: u: rmse: 0.06498562323181704 baseline_rmse: 1.2761651462099894 nrmse: 0.050922581160294314 v: rmse: 0.04133891376364024 baseline_rmse: 1.3343993130174792 nrmse: 0.03097941775026884 - step: 32 scores: u: rmse: 0.05295915753943094 baseline_rmse: 1.2761651462099894 nrmse: 0.04149867099623535 v: rmse: 0.10447499427366826 baseline_rmse: 1.3343993130174792 nrmse: 0.07829365112412925 - step: 128 scores: u: rmse: 0.00932612991919312 baseline_rmse: 1.2761651462099894 nrmse: 0.007307933418249405 v: rmse: 0.01754182235084703 baseline_rmse: 1.3343993130174792 nrmse: 0.013145856850884973 - step: 256 scores: u: rmse: 0.019303910259913987 baseline_rmse: 1.2761651462099894 nrmse: 0.015126498570536561 v: rmse: 0.010437616341014606 baseline_rmse: 1.3343993130174792 nrmse: 0.007821958718947485 complete_reset: before: u: rmse: 1.348654753140097 baseline_rmse: 1.2967289998405087 nrmse: 1.0400436431251054 v: rmse: 1.2085798710962348 baseline_rmse: 1.2693145185962316 nrmse: 0.9521516167898522 after: u: rmse: 0.021755791973197713 baseline_rmse: 1.2967289998405087 nrmse: 0.016777439215035345 v: rmse: 0.012716984718581223 baseline_rmse: 1.2693145185962316 nrmse: 0.010018781422783434 changed_state_entries: - nodes.record/a.embedder.embeddings.weight - nodes.record/a.embedder.counter.counts - nodes.record/a.embedder.linear.weight - nodes.record/a.embedder.linear.bias - nodes.record/a.embedder.scalar.weight - nodes.record/a.embedder.normalizer.mean - nodes.record/a.embedder.normalizer.var - nodes.record/a.embedder.normalizer.count - nodes.record/b.embedder.embeddings.weight - nodes.record/b.embedder.counter.counts - nodes.record/b.embedder.linear.weight - nodes.record/b.embedder.linear.bias - nodes.record/b.embedder.scalar.weight - nodes.record/b.embedder.normalizer.mean - nodes.record/b.embedder.normalizer.var - nodes.record/b.embedder.normalizer.count - nodes.record/u.embedder.embeddings.weight - nodes.record/u.embedder.counter.counts - nodes.record/u.embedder.linear.weight - nodes.record/u.embedder.linear.bias - nodes.record/u.embedder.scalar.weight - nodes.record/u.embedder.normalizer.mean - nodes.record/u.embedder.normalizer.var - nodes.record/u.embedder.normalizer.count - nodes.record/u.decoder.context_projection.0.weight - nodes.record/u.decoder.context_projection.0.bias - nodes.record/u.decoder.context_projection.1.weight - nodes.record/u.decoder.context_projection.1.bias - nodes.record/u.decoder.context_projection.3.weight - nodes.record/u.decoder.context_projection.3.bias - nodes.record/u.decoder.pool.queries - nodes.record/u.decoder.pool.blocks.0.attention_norm.weight - nodes.record/u.decoder.pool.blocks.0.attention_norm.bias - nodes.record/u.decoder.pool.blocks.0.ffn_norm.weight - nodes.record/u.decoder.pool.blocks.0.ffn_norm.bias - nodes.record/u.decoder.pool.blocks.0.attention.q_proj.weight - nodes.record/u.decoder.pool.blocks.0.attention.q_proj.bias - nodes.record/u.decoder.pool.blocks.0.attention.k_proj.weight - nodes.record/u.decoder.pool.blocks.0.attention.k_proj.bias - nodes.record/u.decoder.pool.blocks.0.attention.v_proj.weight - nodes.record/u.decoder.pool.blocks.0.attention.v_proj.bias - nodes.record/u.decoder.pool.blocks.0.attention.out_proj.weight - nodes.record/u.decoder.pool.blocks.0.attention.out_proj.bias - nodes.record/u.decoder.pool.blocks.0.ffn.0.weight - nodes.record/u.decoder.pool.blocks.0.ffn.0.bias - nodes.record/u.decoder.pool.blocks.0.ffn.3.weight - nodes.record/u.decoder.pool.blocks.0.ffn.3.bias - nodes.record/u.decoder.pool.norm.weight - nodes.record/u.decoder.pool.norm.bias - nodes.record/u.decoder.classification.weight - nodes.record/u.decoder.classification.bias - nodes.record/u.decoder.regression.weight - nodes.record/u.decoder.regression.bias - nodes.record/v.embedder.embeddings.weight - nodes.record/v.embedder.counter.counts - nodes.record/v.embedder.linear.weight - nodes.record/v.embedder.linear.bias - nodes.record/v.embedder.scalar.weight - nodes.record/v.embedder.normalizer.mean - nodes.record/v.embedder.normalizer.var - nodes.record/v.embedder.normalizer.count - nodes.record/v.decoder.context_projection.0.weight - nodes.record/v.decoder.context_projection.0.bias - nodes.record/v.decoder.context_projection.1.weight - nodes.record/v.decoder.context_projection.1.bias - nodes.record/v.decoder.context_projection.3.weight - nodes.record/v.decoder.context_projection.3.bias - nodes.record/v.decoder.pool.queries - nodes.record/v.decoder.pool.blocks.0.attention_norm.weight - nodes.record/v.decoder.pool.blocks.0.attention_norm.bias - nodes.record/v.decoder.pool.blocks.0.ffn_norm.weight - nodes.record/v.decoder.pool.blocks.0.ffn_norm.bias - nodes.record/v.decoder.pool.blocks.0.attention.q_proj.weight - nodes.record/v.decoder.pool.blocks.0.attention.q_proj.bias - nodes.record/v.decoder.pool.blocks.0.attention.k_proj.weight - nodes.record/v.decoder.pool.blocks.0.attention.k_proj.bias - nodes.record/v.decoder.pool.blocks.0.attention.v_proj.weight - nodes.record/v.decoder.pool.blocks.0.attention.v_proj.bias - nodes.record/v.decoder.pool.blocks.0.attention.out_proj.weight - nodes.record/v.decoder.pool.blocks.0.attention.out_proj.bias - nodes.record/v.decoder.pool.blocks.0.ffn.0.weight - nodes.record/v.decoder.pool.blocks.0.ffn.0.bias - nodes.record/v.decoder.pool.blocks.0.ffn.3.weight - nodes.record/v.decoder.pool.blocks.0.ffn.3.bias - nodes.record/v.decoder.pool.norm.weight - nodes.record/v.decoder.pool.norm.bias - nodes.record/v.decoder.classification.weight - nodes.record/v.decoder.classification.bias - nodes.record/v.decoder.regression.weight - nodes.record/v.decoder.regression.bias - nodes.record.encoder.encoder.0.attention_norm.weight - nodes.record.encoder.encoder.0.attention_norm.bias - nodes.record.encoder.encoder.0.ffn_norm.weight - nodes.record.encoder.encoder.0.ffn_norm.bias - nodes.record.encoder.encoder.0.attention.q_proj.weight - nodes.record.encoder.encoder.0.attention.q_proj.bias - nodes.record.encoder.encoder.0.attention.k_proj.weight - nodes.record.encoder.encoder.0.attention.k_proj.bias - nodes.record.encoder.encoder.0.attention.v_proj.weight - nodes.record.encoder.encoder.0.attention.v_proj.bias - nodes.record.encoder.encoder.0.attention.out_proj.weight - nodes.record.encoder.encoder.0.attention.out_proj.bias - nodes.record.encoder.encoder.0.ffn.0.weight - nodes.record.encoder.encoder.0.ffn.0.bias - nodes.record.encoder.encoder.0.ffn.3.weight - nodes.record.encoder.encoder.0.ffn.3.bias - nodes.record.encoder.coordinate_encoder.attention_norm.weight - nodes.record.encoder.coordinate_encoder.attention_norm.bias - nodes.record.encoder.coordinate_encoder.ffn_norm.weight - nodes.record.encoder.coordinate_encoder.ffn_norm.bias - nodes.record.encoder.coordinate_encoder.attention.q_proj.weight - nodes.record.encoder.coordinate_encoder.attention.q_proj.bias - nodes.record.encoder.coordinate_encoder.attention.k_proj.weight - nodes.record.encoder.coordinate_encoder.attention.k_proj.bias - nodes.record.encoder.coordinate_encoder.attention.v_proj.weight - nodes.record.encoder.coordinate_encoder.attention.v_proj.bias - nodes.record.encoder.coordinate_encoder.attention.out_proj.weight - nodes.record.encoder.coordinate_encoder.attention.out_proj.bias - nodes.record.encoder.coordinate_encoder.ffn.0.weight - nodes.record.encoder.coordinate_encoder.ffn.0.bias - nodes.record.encoder.coordinate_encoder.ffn.3.weight - nodes.record.encoder.coordinate_encoder.ffn.3.bias - nodes.record.encoder.pool.queries - nodes.record.encoder.pool.mass_direction - nodes.record.encoder.pool.blocks.0.attention_norm.weight - nodes.record.encoder.pool.blocks.0.attention_norm.bias - nodes.record.encoder.pool.blocks.0.ffn_norm.weight - nodes.record.encoder.pool.blocks.0.ffn_norm.bias - nodes.record.encoder.pool.blocks.0.attention.q_proj.weight - nodes.record.encoder.pool.blocks.0.attention.q_proj.bias - nodes.record.encoder.pool.blocks.0.attention.k_proj.weight - nodes.record.encoder.pool.blocks.0.attention.k_proj.bias - nodes.record.encoder.pool.blocks.0.attention.v_proj.weight - nodes.record.encoder.pool.blocks.0.attention.v_proj.bias - nodes.record.encoder.pool.blocks.0.attention.out_proj.weight - nodes.record.encoder.pool.blocks.0.attention.out_proj.bias - nodes.record.encoder.pool.blocks.0.ffn.0.weight - nodes.record.encoder.pool.blocks.0.ffn.0.bias - nodes.record.encoder.pool.blocks.0.ffn.3.weight - nodes.record.encoder.pool.blocks.0.ffn.3.bias - nodes.record.encoder.pool.norm.weight - nodes.record.encoder.pool.norm.bias - nodes.record.encoder.pool.mass_projection.weight initial_prediction_drift: u: 3.201029494404793 v: 2.767811194062233 steps: 256 optimizer_covers_current_parameters: true validation_curve: - step: 0 scores: u: rmse: 1.3161852264157115 baseline_rmse: 1.2761651462099894 nrmse: 1.0313596405015255 v: rmse: 1.274914898643669 baseline_rmse: 1.3343993130174792 nrmse: 0.9554223283888703 - step: 32 scores: u: rmse: 0.07635042643943989 baseline_rmse: 1.2761651462099894 nrmse: 0.05982801416116769 v: rmse: 0.08731643516861808 baseline_rmse: 1.3343993130174792 nrmse: 0.06543501208133065 - step: 128 scores: u: rmse: 0.04924193097719149 baseline_rmse: 1.2761651462099894 nrmse: 0.03858586102545765 v: rmse: 0.03647201738984849 baseline_rmse: 1.3343993130174792 nrmse: 0.02733216139580757 - step: 256 scores: u: rmse: 0.022333670874404462 baseline_rmse: 1.2761651462099894 nrmse: 0.01750061184536497 v: rmse: 0.012428205365517732 baseline_rmse: 1.3343993130174792 nrmse: 0.009313707856618881 test_rows: 1024 baseline: Source-training mean for test comparisons; phase-training mean for validation curves optimizer_policy: New AdamW factory and Trainer for every fit; both hidden labels rehearsed mutation: reset record/u; complete-reset control resets record with descendants=True checks: Source learns both targets below 0.25 nRMSE: true Reset changes selected learned state: true Reset preserves every unselected state entry: true Reset preserves the schema: true Selected target immediately loses its skill: true Unselected target immediately preserves predictions: true 'selective_reset: fresh optimizer covers current parameters': true 'selective_reset: both final targets below 0.25 nRMSE': true Reset head parameters actually relearn: true Hidden target placeholders cannot affect predictions: true Relearned checkpoint preserves schema and state: true Relearned checkpoint preserves predictions: true 'continuation: fresh optimizer covers current parameters': true 'continuation: both final targets below 0.25 nRMSE': true Complete reset loses both learned tasks: true 'complete_reset: fresh optimizer covers current parameters': true 'complete_reset: both final targets below 0.25 nRMSE': true outcome: met duration_seconds: 44.535 P049: script: proofs/mutations/deactivate_and_restore_input.py page: docs/proofs/mutations/deactivate-and-restore-input.py historical: status: Provisional pass source_revision: null sources: [] summary: New experiment; no historical run predates registration. insights: 'Initial hypothesis: An informative pure input can be deactivated and reactivated without erasing its trained state.' evidence: No historical measurements. See recorded runs for actual outcomes. runs: - started_at: '2026-09-15T18:20:51.362453+00:00' mode: smoke seed: 4901 steps_override: 2 accelerator: cpu threads: 1 provenance: revision: 5d985953fd0416bc5982a3518d8f43cadb52110c source_sha256: b6c88a66cf37b1ae7164ef5c44787d1d01d26b6f935db20e0b5332dd3966e061 script_sha256: 5b099da2356648d54a137d62b71f1065c198990ab0bce3b79f4e690666aaac9e code_sha256: da353a706be8810fe4bc2f522e690989a46d8d095e3be7d988daa3e86127809c python: 3.12.10 platform: macOS-26.5.2-arm64-arm-64bit host: Granthams-MacBook-Pro.local relflow_path: /Users/grantham/Desktop/relflow/src/relflow packages: relflow: 0.2.2 torch: 2.12.0 lightning: 2.6.4 numpy: 2.4.6 pyarrow: 24.0.0 polars: 1.44.1 torchmetrics: 1.9.0 tensordict: 0.12.4 pydantic: 2.13.4 cuda: null gpu: null metrics: source: rmse: 0.45946386324889843 baseline_rmse: 0.8228176545152625 nrmse: 0.5584030200708046 source_steps: 2 source_prerequisite_met: false downstream_interpretable: false controls: shuffled_b: rmse: 0.7360061023172202 baseline_rmse: 0.8228176545152625 nrmse: 0.8944947866375271 zero_filled_b: rmse: 0.656315068211645 baseline_rmse: 0.8228176545152625 nrmse: 0.7976433959700253 only_a_oracle: rmse: 0.5857724467092484 baseline_rmse: 0.8228176545152625 nrmse: 0.7119104014052033 edits: - edit: cycle 1 deactivate inactive: true scores: rmse: 0.8676151829725777 baseline_rmse: 0.8228176545152625 nrmse: 1.0544440535658004 max_source_prediction_drift: 0.8917339276522398 learned_state_preserved: true - edit: cycle 1 reactivate inactive: false scores: rmse: 0.45946386324889843 baseline_rmse: 0.8228176545152625 nrmse: 0.5584030200708046 max_source_prediction_drift: 0.0 learned_state_preserved: true - edit: cycle 2 deactivate inactive: true scores: rmse: 0.8676151829725777 baseline_rmse: 0.8228176545152625 nrmse: 1.0544440535658004 max_source_prediction_drift: 0.8917339276522398 learned_state_preserved: true - edit: cycle 2 reactivate inactive: false scores: rmse: 0.45946386324889843 baseline_rmse: 0.8228176545152625 nrmse: 0.5584030200708046 max_source_prediction_drift: 0.0 learned_state_preserved: true - edit: cycle 3 deactivate inactive: true scores: rmse: 0.8676151829725777 baseline_rmse: 0.8228176545152625 nrmse: 1.0544440535658004 max_source_prediction_drift: 0.8917339276522398 learned_state_preserved: true - edit: cycle 3 reactivate inactive: false scores: rmse: 0.45946386324889843 baseline_rmse: 0.8228176545152625 nrmse: 0.5584030200708046 max_source_prediction_drift: 0.0 learned_state_preserved: true - edit: temporary override inactive: true scores: rmse: 0.8676151829725777 baseline_rmse: 0.8228176545152625 nrmse: 1.0544440535658004 max_source_prediction_drift: 0.8917339276522398 learned_state_preserved: true - edit: normal override exit inactive: false scores: rmse: 0.45946386324889843 baseline_rmse: 0.8228176545152625 nrmse: 0.5584030200708046 max_source_prediction_drift: 0.0 learned_state_preserved: true - edit: override before exception inactive: true scores: rmse: 0.8676151829725777 baseline_rmse: 0.8228176545152625 nrmse: 1.0544440535658004 max_source_prediction_drift: 0.8917339276522398 learned_state_preserved: true - edit: exceptional override exit inactive: false scores: rmse: 0.45946386324889843 baseline_rmse: 0.8228176545152625 nrmse: 0.5584030200708046 max_source_prediction_drift: 0.0 learned_state_preserved: true test_rows: 2048 optimizer_policy: Fresh AdamW for source; no fitting inside updates or overrides mutation: update record/b active=False/True three times; override with normal and exceptional exits checks: Source learns both-input relationship below 0.25 nRMSE: false Shuffling b destroys useful signal: false Source hidden target values cannot affect predictions: true Unchanged checkpoint preserves predictions: true 'cycle 1 deactivate: all learned state survives': true 'cycle 1 deactivate: input is inactive': true 'cycle 1 deactivate: b values cannot affect predictions': true 'cycle 1 deactivate: removing b loses information': true Inactive checkpoint preserves schema and state: true Inactive checkpoint preserves predictions: true Loaded inactive input can restore its trained function: true 'cycle 1 reactivate: all learned state survives': true 'cycle 1 reactivate: original schema is restored': true 'cycle 1 reactivate: trained predictions are restored': true 'cycle 2 deactivate: all learned state survives': true 'cycle 2 deactivate: input is inactive': true 'cycle 2 deactivate: b values cannot affect predictions': true 'cycle 2 deactivate: removing b loses information': true 'cycle 2 reactivate: all learned state survives': true 'cycle 2 reactivate: original schema is restored': true 'cycle 2 reactivate: trained predictions are restored': true 'cycle 3 deactivate: all learned state survives': true 'cycle 3 deactivate: input is inactive': true 'cycle 3 deactivate: b values cannot affect predictions': true 'cycle 3 deactivate: removing b loses information': true 'cycle 3 reactivate: all learned state survives': true 'cycle 3 reactivate: original schema is restored': true 'cycle 3 reactivate: trained predictions are restored': true 'temporary override: all learned state survives': true 'temporary override: input is inactive': true 'temporary override: b values cannot affect predictions': true 'temporary override: removing b loses information': true 'normal override exit: all learned state survives': true 'normal override exit: original schema is restored': true 'normal override exit: trained predictions are restored': true 'override before exception: all learned state survives': true 'override before exception: input is inactive': true 'override before exception: b values cannot affect predictions': true 'override before exception: removing b loses information': true 'exceptional override exit: all learned state survives': true 'exceptional override exit: original schema is restored': true 'exceptional override exit: trained predictions are restored': true outcome: not_met duration_seconds: 1.079 - started_at: '2026-09-15T18:21:55.145131+00:00' mode: full seed: 4901 steps_override: null accelerator: cpu threads: 1 provenance: revision: 5d985953fd0416bc5982a3518d8f43cadb52110c source_sha256: b6c88a66cf37b1ae7164ef5c44787d1d01d26b6f935db20e0b5332dd3966e061 script_sha256: 5b099da2356648d54a137d62b71f1065c198990ab0bce3b79f4e690666aaac9e code_sha256: da353a706be8810fe4bc2f522e690989a46d8d095e3be7d988daa3e86127809c python: 3.12.10 platform: macOS-26.5.2-arm64-arm-64bit host: Granthams-MacBook-Pro.local relflow_path: /Users/grantham/Desktop/relflow/src/relflow packages: relflow: 0.2.2 torch: 2.12.0 lightning: 2.6.4 numpy: 2.4.6 pyarrow: 24.0.0 polars: 1.44.1 torchmetrics: 1.9.0 tensordict: 0.12.4 pydantic: 2.13.4 cuda: null gpu: null metrics: source: rmse: 0.02268946924387725 baseline_rmse: 0.8228176545152625 nrmse: 0.027575331082612765 source_steps: 512 source_prerequisite_met: true downstream_interpretable: true controls: shuffled_b: rmse: 0.8323123094901392 baseline_rmse: 0.8228176545152625 nrmse: 1.0115391969566698 zero_filled_b: rmse: 0.5855899480602854 baseline_rmse: 0.8228176545152625 nrmse: 0.711688604208751 only_a_oracle: rmse: 0.5857724467092484 baseline_rmse: 0.8228176545152625 nrmse: 0.7119104014052033 edits: - edit: cycle 1 deactivate inactive: true scores: rmse: 0.6912687578436512 baseline_rmse: 0.8228176545152625 nrmse: 0.8401238768398701 max_source_prediction_drift: 1.6222696229815483 learned_state_preserved: true - edit: cycle 1 reactivate inactive: false scores: rmse: 0.02268946924387725 baseline_rmse: 0.8228176545152625 nrmse: 0.027575331082612765 max_source_prediction_drift: 0.0 learned_state_preserved: true - edit: cycle 2 deactivate inactive: true scores: rmse: 0.6912687578436512 baseline_rmse: 0.8228176545152625 nrmse: 0.8401238768398701 max_source_prediction_drift: 1.6222696229815483 learned_state_preserved: true - edit: cycle 2 reactivate inactive: false scores: rmse: 0.02268946924387725 baseline_rmse: 0.8228176545152625 nrmse: 0.027575331082612765 max_source_prediction_drift: 0.0 learned_state_preserved: true - edit: cycle 3 deactivate inactive: true scores: rmse: 0.6912687578436512 baseline_rmse: 0.8228176545152625 nrmse: 0.8401238768398701 max_source_prediction_drift: 1.6222696229815483 learned_state_preserved: true - edit: cycle 3 reactivate inactive: false scores: rmse: 0.02268946924387725 baseline_rmse: 0.8228176545152625 nrmse: 0.027575331082612765 max_source_prediction_drift: 0.0 learned_state_preserved: true - edit: temporary override inactive: true scores: rmse: 0.6912687578436512 baseline_rmse: 0.8228176545152625 nrmse: 0.8401238768398701 max_source_prediction_drift: 1.6222696229815483 learned_state_preserved: true - edit: normal override exit inactive: false scores: rmse: 0.02268946924387725 baseline_rmse: 0.8228176545152625 nrmse: 0.027575331082612765 max_source_prediction_drift: 0.0 learned_state_preserved: true - edit: override before exception inactive: true scores: rmse: 0.6912687578436512 baseline_rmse: 0.8228176545152625 nrmse: 0.8401238768398701 max_source_prediction_drift: 1.6222696229815483 learned_state_preserved: true - edit: exceptional override exit inactive: false scores: rmse: 0.02268946924387725 baseline_rmse: 0.8228176545152625 nrmse: 0.027575331082612765 max_source_prediction_drift: 0.0 learned_state_preserved: true test_rows: 2048 optimizer_policy: Fresh AdamW for source; no fitting inside updates or overrides mutation: update record/b active=False/True three times; override with normal and exceptional exits checks: Source learns both-input relationship below 0.25 nRMSE: true Shuffling b destroys useful signal: true Source hidden target values cannot affect predictions: true Unchanged checkpoint preserves predictions: true 'cycle 1 deactivate: all learned state survives': true 'cycle 1 deactivate: input is inactive': true 'cycle 1 deactivate: b values cannot affect predictions': true 'cycle 1 deactivate: removing b loses information': true Inactive checkpoint preserves schema and state: true Inactive checkpoint preserves predictions: true Loaded inactive input can restore its trained function: true 'cycle 1 reactivate: all learned state survives': true 'cycle 1 reactivate: original schema is restored': true 'cycle 1 reactivate: trained predictions are restored': true 'cycle 2 deactivate: all learned state survives': true 'cycle 2 deactivate: input is inactive': true 'cycle 2 deactivate: b values cannot affect predictions': true 'cycle 2 deactivate: removing b loses information': true 'cycle 2 reactivate: all learned state survives': true 'cycle 2 reactivate: original schema is restored': true 'cycle 2 reactivate: trained predictions are restored': true 'cycle 3 deactivate: all learned state survives': true 'cycle 3 deactivate: input is inactive': true 'cycle 3 deactivate: b values cannot affect predictions': true 'cycle 3 deactivate: removing b loses information': true 'cycle 3 reactivate: all learned state survives': true 'cycle 3 reactivate: original schema is restored': true 'cycle 3 reactivate: trained predictions are restored': true 'temporary override: all learned state survives': true 'temporary override: input is inactive': true 'temporary override: b values cannot affect predictions': true 'temporary override: removing b loses information': true 'normal override exit: all learned state survives': true 'normal override exit: original schema is restored': true 'normal override exit: trained predictions are restored': true 'override before exception: all learned state survives': true 'override before exception: input is inactive': true 'override before exception: b values cannot affect predictions': true 'override before exception: removing b loses information': true 'exceptional override exit: all learned state survives': true 'exceptional override exit: original schema is restored': true 'exceptional override exit: trained predictions are restored': true outcome: met duration_seconds: 7.187 - started_at: '2026-09-15T18:23:15.633080+00:00' mode: full seed: 4901 steps_override: null accelerator: gpu threads: 1 provenance: revision: null source_sha256: b6c88a66cf37b1ae7164ef5c44787d1d01d26b6f935db20e0b5332dd3966e061 script_sha256: 5b099da2356648d54a137d62b71f1065c198990ab0bce3b79f4e690666aaac9e code_sha256: da353a706be8810fe4bc2f522e690989a46d8d095e3be7d988daa3e86127809c python: 3.12.6 platform: Linux-6.8.0-138-generic-x86_64-with-glibc2.35 host: lab relflow_path: /home/grantham/relflow-removal-6b17okdx/src/relflow packages: relflow: 0.2.2 torch: 2.12.0 lightning: 2.6.4 numpy: 2.4.6 pyarrow: 24.0.0 polars: 1.44.1 torchmetrics: 1.9.0 tensordict: 0.12.4 pydantic: 2.13.4 cuda: '13.0' gpu: NVIDIA GeForce RTX 3090 metrics: source: rmse: 0.01799741589811066 baseline_rmse: 0.8228176545152625 nrmse: 0.021872909264098475 source_steps: 512 source_prerequisite_met: true downstream_interpretable: true controls: shuffled_b: rmse: 0.8320161740170012 baseline_rmse: 0.8228176545152625 nrmse: 1.0111792928254046 zero_filled_b: rmse: 0.5857425690628448 baseline_rmse: 0.8228176545152625 nrmse: 0.7118740900229188 only_a_oracle: rmse: 0.5857724467092484 baseline_rmse: 0.8228176545152625 nrmse: 0.7119104014052033 edits: - edit: cycle 1 deactivate inactive: true scores: rmse: 0.6899821178281003 baseline_rmse: 0.8228176545152625 nrmse: 0.8385601767800933 max_source_prediction_drift: 1.674555629491806 learned_state_preserved: true - edit: cycle 1 reactivate inactive: false scores: rmse: 0.01799741589811066 baseline_rmse: 0.8228176545152625 nrmse: 0.021872909264098475 max_source_prediction_drift: 0.0 learned_state_preserved: true - edit: cycle 2 deactivate inactive: true scores: rmse: 0.6899821178281003 baseline_rmse: 0.8228176545152625 nrmse: 0.8385601767800933 max_source_prediction_drift: 1.674555629491806 learned_state_preserved: true - edit: cycle 2 reactivate inactive: false scores: rmse: 0.01799741589811066 baseline_rmse: 0.8228176545152625 nrmse: 0.021872909264098475 max_source_prediction_drift: 0.0 learned_state_preserved: true - edit: cycle 3 deactivate inactive: true scores: rmse: 0.6899821178281003 baseline_rmse: 0.8228176545152625 nrmse: 0.8385601767800933 max_source_prediction_drift: 1.674555629491806 learned_state_preserved: true - edit: cycle 3 reactivate inactive: false scores: rmse: 0.01799741589811066 baseline_rmse: 0.8228176545152625 nrmse: 0.021872909264098475 max_source_prediction_drift: 0.0 learned_state_preserved: true - edit: temporary override inactive: true scores: rmse: 0.6899821178281003 baseline_rmse: 0.8228176545152625 nrmse: 0.8385601767800933 max_source_prediction_drift: 1.674555629491806 learned_state_preserved: true - edit: normal override exit inactive: false scores: rmse: 0.01799741589811066 baseline_rmse: 0.8228176545152625 nrmse: 0.021872909264098475 max_source_prediction_drift: 0.0 learned_state_preserved: true - edit: override before exception inactive: true scores: rmse: 0.6899821178281003 baseline_rmse: 0.8228176545152625 nrmse: 0.8385601767800933 max_source_prediction_drift: 1.674555629491806 learned_state_preserved: true - edit: exceptional override exit inactive: false scores: rmse: 0.01799741589811066 baseline_rmse: 0.8228176545152625 nrmse: 0.021872909264098475 max_source_prediction_drift: 0.0 learned_state_preserved: true test_rows: 2048 optimizer_policy: Fresh AdamW for source; no fitting inside updates or overrides mutation: update record/b active=False/True three times; override with normal and exceptional exits checks: Source learns both-input relationship below 0.25 nRMSE: true Shuffling b destroys useful signal: true Source hidden target values cannot affect predictions: true Unchanged checkpoint preserves predictions: true 'cycle 1 deactivate: all learned state survives': true 'cycle 1 deactivate: input is inactive': true 'cycle 1 deactivate: b values cannot affect predictions': true 'cycle 1 deactivate: removing b loses information': true Inactive checkpoint preserves schema and state: true Inactive checkpoint preserves predictions: true Loaded inactive input can restore its trained function: true 'cycle 1 reactivate: all learned state survives': true 'cycle 1 reactivate: original schema is restored': true 'cycle 1 reactivate: trained predictions are restored': true 'cycle 2 deactivate: all learned state survives': true 'cycle 2 deactivate: input is inactive': true 'cycle 2 deactivate: b values cannot affect predictions': true 'cycle 2 deactivate: removing b loses information': true 'cycle 2 reactivate: all learned state survives': true 'cycle 2 reactivate: original schema is restored': true 'cycle 2 reactivate: trained predictions are restored': true 'cycle 3 deactivate: all learned state survives': true 'cycle 3 deactivate: input is inactive': true 'cycle 3 deactivate: b values cannot affect predictions': true 'cycle 3 deactivate: removing b loses information': true 'cycle 3 reactivate: all learned state survives': true 'cycle 3 reactivate: original schema is restored': true 'cycle 3 reactivate: trained predictions are restored': true 'temporary override: all learned state survives': true 'temporary override: input is inactive': true 'temporary override: b values cannot affect predictions': true 'temporary override: removing b loses information': true 'normal override exit: all learned state survives': true 'normal override exit: original schema is restored': true 'normal override exit: trained predictions are restored': true 'override before exception: all learned state survives': true 'override before exception: input is inactive': true 'override before exception: b values cannot affect predictions': true 'override before exception: removing b loses information': true 'exceptional override exit: all learned state survives': true 'exceptional override exit: original schema is restored': true 'exceptional override exit: trained predictions are restored': true outcome: met duration_seconds: 11.958 - started_at: '2026-09-15T18:24:06.897983+00:00' mode: full seed: 7201 steps_override: null accelerator: gpu threads: 1 provenance: revision: null source_sha256: b6c88a66cf37b1ae7164ef5c44787d1d01d26b6f935db20e0b5332dd3966e061 script_sha256: 5b099da2356648d54a137d62b71f1065c198990ab0bce3b79f4e690666aaac9e code_sha256: da353a706be8810fe4bc2f522e690989a46d8d095e3be7d988daa3e86127809c python: 3.12.6 platform: Linux-6.8.0-138-generic-x86_64-with-glibc2.35 host: lab relflow_path: /home/grantham/relflow-removal-6b17okdx/src/relflow packages: relflow: 0.2.2 torch: 2.12.0 lightning: 2.6.4 numpy: 2.4.6 pyarrow: 24.0.0 polars: 1.44.1 torchmetrics: 1.9.0 tensordict: 0.12.4 pydantic: 2.13.4 cuda: '13.0' gpu: NVIDIA GeForce RTX 3090 metrics: source: rmse: 0.012387407114893877 baseline_rmse: 0.8120016401971133 nrmse: 0.015255396666300866 source_steps: 512 source_prerequisite_met: true downstream_interpretable: true controls: shuffled_b: rmse: 0.8415880130521701 baseline_rmse: 0.8120016401971133 nrmse: 1.0364363461727426 zero_filled_b: rmse: 0.5809422472772158 baseline_rmse: 0.8120016401971133 nrmse: 0.7154446721760218 only_a_oracle: rmse: 0.5809035857842813 baseline_rmse: 0.8120016401971133 nrmse: 0.7153970595961691 edits: - edit: cycle 1 deactivate inactive: true scores: rmse: 0.6978588819705238 baseline_rmse: 0.8120016401971133 nrmse: 0.8594303846493693 max_source_prediction_drift: 1.6460874378681183 learned_state_preserved: true - edit: cycle 1 reactivate inactive: false scores: rmse: 0.012387407114893877 baseline_rmse: 0.8120016401971133 nrmse: 0.015255396666300866 max_source_prediction_drift: 0.0 learned_state_preserved: true - edit: cycle 2 deactivate inactive: true scores: rmse: 0.6978588819705238 baseline_rmse: 0.8120016401971133 nrmse: 0.8594303846493693 max_source_prediction_drift: 1.6460874378681183 learned_state_preserved: true - edit: cycle 2 reactivate inactive: false scores: rmse: 0.012387407114893877 baseline_rmse: 0.8120016401971133 nrmse: 0.015255396666300866 max_source_prediction_drift: 0.0 learned_state_preserved: true - edit: cycle 3 deactivate inactive: true scores: rmse: 0.6978588819705238 baseline_rmse: 0.8120016401971133 nrmse: 0.8594303846493693 max_source_prediction_drift: 1.6460874378681183 learned_state_preserved: true - edit: cycle 3 reactivate inactive: false scores: rmse: 0.012387407114893877 baseline_rmse: 0.8120016401971133 nrmse: 0.015255396666300866 max_source_prediction_drift: 0.0 learned_state_preserved: true - edit: temporary override inactive: true scores: rmse: 0.6978588819705238 baseline_rmse: 0.8120016401971133 nrmse: 0.8594303846493693 max_source_prediction_drift: 1.6460874378681183 learned_state_preserved: true - edit: normal override exit inactive: false scores: rmse: 0.012387407114893877 baseline_rmse: 0.8120016401971133 nrmse: 0.015255396666300866 max_source_prediction_drift: 0.0 learned_state_preserved: true - edit: override before exception inactive: true scores: rmse: 0.6978588819705238 baseline_rmse: 0.8120016401971133 nrmse: 0.8594303846493693 max_source_prediction_drift: 1.6460874378681183 learned_state_preserved: true - edit: exceptional override exit inactive: false scores: rmse: 0.012387407114893877 baseline_rmse: 0.8120016401971133 nrmse: 0.015255396666300866 max_source_prediction_drift: 0.0 learned_state_preserved: true test_rows: 2048 optimizer_policy: Fresh AdamW for source; no fitting inside updates or overrides mutation: update record/b active=False/True three times; override with normal and exceptional exits checks: Source learns both-input relationship below 0.25 nRMSE: true Shuffling b destroys useful signal: true Source hidden target values cannot affect predictions: true Unchanged checkpoint preserves predictions: true 'cycle 1 deactivate: all learned state survives': true 'cycle 1 deactivate: input is inactive': true 'cycle 1 deactivate: b values cannot affect predictions': true 'cycle 1 deactivate: removing b loses information': true Inactive checkpoint preserves schema and state: true Inactive checkpoint preserves predictions: true Loaded inactive input can restore its trained function: true 'cycle 1 reactivate: all learned state survives': true 'cycle 1 reactivate: original schema is restored': true 'cycle 1 reactivate: trained predictions are restored': true 'cycle 2 deactivate: all learned state survives': true 'cycle 2 deactivate: input is inactive': true 'cycle 2 deactivate: b values cannot affect predictions': true 'cycle 2 deactivate: removing b loses information': true 'cycle 2 reactivate: all learned state survives': true 'cycle 2 reactivate: original schema is restored': true 'cycle 2 reactivate: trained predictions are restored': true 'cycle 3 deactivate: all learned state survives': true 'cycle 3 deactivate: input is inactive': true 'cycle 3 deactivate: b values cannot affect predictions': true 'cycle 3 deactivate: removing b loses information': true 'cycle 3 reactivate: all learned state survives': true 'cycle 3 reactivate: original schema is restored': true 'cycle 3 reactivate: trained predictions are restored': true 'temporary override: all learned state survives': true 'temporary override: input is inactive': true 'temporary override: b values cannot affect predictions': true 'temporary override: removing b loses information': true 'normal override exit: all learned state survives': true 'normal override exit: original schema is restored': true 'normal override exit: trained predictions are restored': true 'override before exception: all learned state survives': true 'override before exception: input is inactive': true 'override before exception: b values cannot affect predictions': true 'override before exception: removing b loses information': true 'exceptional override exit: all learned state survives': true 'exceptional override exit: original schema is restored': true 'exceptional override exit: trained predictions are restored': true outcome: met duration_seconds: 12.022 - started_at: '2026-09-15T18:24:58.506617+00:00' mode: full seed: 7202 steps_override: null accelerator: gpu threads: 1 provenance: revision: null source_sha256: b6c88a66cf37b1ae7164ef5c44787d1d01d26b6f935db20e0b5332dd3966e061 script_sha256: 5b099da2356648d54a137d62b71f1065c198990ab0bce3b79f4e690666aaac9e code_sha256: da353a706be8810fe4bc2f522e690989a46d8d095e3be7d988daa3e86127809c python: 3.12.6 platform: Linux-6.8.0-138-generic-x86_64-with-glibc2.35 host: lab relflow_path: /home/grantham/relflow-removal-6b17okdx/src/relflow packages: relflow: 0.2.2 torch: 2.12.0 lightning: 2.6.4 numpy: 2.4.6 pyarrow: 24.0.0 polars: 1.44.1 torchmetrics: 1.9.0 tensordict: 0.12.4 pydantic: 2.13.4 cuda: '13.0' gpu: NVIDIA GeForce RTX 3090 metrics: source: rmse: 0.02008381101831335 baseline_rmse: 0.8352833232272557 nrmse: 0.02404430982856956 source_steps: 512 source_prerequisite_met: true downstream_interpretable: true controls: shuffled_b: rmse: 0.82572553566775 baseline_rmse: 0.8352833232272557 nrmse: 0.9885574304026834 zero_filled_b: rmse: 0.5776697674270659 baseline_rmse: 0.8352833232272557 nrmse: 0.6915854194180998 only_a_oracle: rmse: 0.5780872291911047 baseline_rmse: 0.8352833232272557 nrmse: 0.6920852040449805 edits: - edit: cycle 1 deactivate inactive: true scores: rmse: 0.8051448159247964 baseline_rmse: 0.8352833232272557 nrmse: 0.963918222159621 max_source_prediction_drift: 1.9680075012147427 learned_state_preserved: true - edit: cycle 1 reactivate inactive: false scores: rmse: 0.02008381101831335 baseline_rmse: 0.8352833232272557 nrmse: 0.02404430982856956 max_source_prediction_drift: 0.0 learned_state_preserved: true - edit: cycle 2 deactivate inactive: true scores: rmse: 0.8051448159247964 baseline_rmse: 0.8352833232272557 nrmse: 0.963918222159621 max_source_prediction_drift: 1.9680075012147427 learned_state_preserved: true - edit: cycle 2 reactivate inactive: false scores: rmse: 0.02008381101831335 baseline_rmse: 0.8352833232272557 nrmse: 0.02404430982856956 max_source_prediction_drift: 0.0 learned_state_preserved: true - edit: cycle 3 deactivate inactive: true scores: rmse: 0.8051448159247964 baseline_rmse: 0.8352833232272557 nrmse: 0.963918222159621 max_source_prediction_drift: 1.9680075012147427 learned_state_preserved: true - edit: cycle 3 reactivate inactive: false scores: rmse: 0.02008381101831335 baseline_rmse: 0.8352833232272557 nrmse: 0.02404430982856956 max_source_prediction_drift: 0.0 learned_state_preserved: true - edit: temporary override inactive: true scores: rmse: 0.8051448159247964 baseline_rmse: 0.8352833232272557 nrmse: 0.963918222159621 max_source_prediction_drift: 1.9680075012147427 learned_state_preserved: true - edit: normal override exit inactive: false scores: rmse: 0.02008381101831335 baseline_rmse: 0.8352833232272557 nrmse: 0.02404430982856956 max_source_prediction_drift: 0.0 learned_state_preserved: true - edit: override before exception inactive: true scores: rmse: 0.8051448159247964 baseline_rmse: 0.8352833232272557 nrmse: 0.963918222159621 max_source_prediction_drift: 1.9680075012147427 learned_state_preserved: true - edit: exceptional override exit inactive: false scores: rmse: 0.02008381101831335 baseline_rmse: 0.8352833232272557 nrmse: 0.02404430982856956 max_source_prediction_drift: 0.0 learned_state_preserved: true test_rows: 2048 optimizer_policy: Fresh AdamW for source; no fitting inside updates or overrides mutation: update record/b active=False/True three times; override with normal and exceptional exits checks: Source learns both-input relationship below 0.25 nRMSE: true Shuffling b destroys useful signal: true Source hidden target values cannot affect predictions: true Unchanged checkpoint preserves predictions: true 'cycle 1 deactivate: all learned state survives': true 'cycle 1 deactivate: input is inactive': true 'cycle 1 deactivate: b values cannot affect predictions': true 'cycle 1 deactivate: removing b loses information': true Inactive checkpoint preserves schema and state: true Inactive checkpoint preserves predictions: true Loaded inactive input can restore its trained function: true 'cycle 1 reactivate: all learned state survives': true 'cycle 1 reactivate: original schema is restored': true 'cycle 1 reactivate: trained predictions are restored': true 'cycle 2 deactivate: all learned state survives': true 'cycle 2 deactivate: input is inactive': true 'cycle 2 deactivate: b values cannot affect predictions': true 'cycle 2 deactivate: removing b loses information': true 'cycle 2 reactivate: all learned state survives': true 'cycle 2 reactivate: original schema is restored': true 'cycle 2 reactivate: trained predictions are restored': true 'cycle 3 deactivate: all learned state survives': true 'cycle 3 deactivate: input is inactive': true 'cycle 3 deactivate: b values cannot affect predictions': true 'cycle 3 deactivate: removing b loses information': true 'cycle 3 reactivate: all learned state survives': true 'cycle 3 reactivate: original schema is restored': true 'cycle 3 reactivate: trained predictions are restored': true 'temporary override: all learned state survives': true 'temporary override: input is inactive': true 'temporary override: b values cannot affect predictions': true 'temporary override: removing b loses information': true 'normal override exit: all learned state survives': true 'normal override exit: original schema is restored': true 'normal override exit: trained predictions are restored': true 'override before exception: all learned state survives': true 'override before exception: input is inactive': true 'override before exception: b values cannot affect predictions': true 'override before exception: removing b loses information': true 'exceptional override exit: all learned state survives': true 'exceptional override exit: original schema is restored': true 'exceptional override exit: trained predictions are restored': true outcome: met duration_seconds: 12.029 P050: script: proofs/mutations/delete_input_and_adapt.py page: docs/proofs/mutations/delete-input-and-adapt.py historical: status: Provisional pass source_revision: null sources: [] summary: New experiment; no historical run predates registration. insights: 'Initial hypothesis: Deleting an informative input permits adaptation toward the conditional mean from remaining inputs.' evidence: No historical measurements. See recorded runs for actual outcomes. runs: - started_at: '2026-09-15T18:20:56.160212+00:00' mode: smoke seed: 5001 steps_override: 2 accelerator: cpu threads: 1 provenance: revision: 5d985953fd0416bc5982a3518d8f43cadb52110c source_sha256: de366e85036478c26153e1ff1f82d3a7faec6fd712790249662f6d5fea48ec37 script_sha256: fe688bd87233e2bfcd5b1c5ca6916536a0049dfd012699881fa7eeb6edb1f487 code_sha256: 71ad78be90abfc30fba52b2e0182b4539ca9101f868befec33b05fdf349d023a python: 3.12.10 platform: macOS-26.5.2-arm64-arm-64bit host: Granthams-MacBook-Pro.local relflow_path: /Users/grantham/Desktop/relflow/src/relflow packages: relflow: 0.2.2 torch: 2.12.0 lightning: 2.6.4 numpy: 2.4.6 pyarrow: 24.0.0 polars: 1.44.1 torchmetrics: 1.9.0 tensordict: 0.12.4 pydantic: 2.13.4 cuda: null gpu: null metrics: source: rmse: 0.4408785755960084 baseline_rmse: 0.8103532807277047 nrmse: 0.5440572477229874 only_a_distance_nrmse: 0.8002568809338788 source_fit: steps: 2 optimizer_covers_current_parameters: true validation_curve: - step: 0 scores: rmse: 1.1152491751251399 baseline_rmse: 0.8227835642459033 nrmse: 1.3554587422359208 only_a_distance_nrmse: 0.9971723998482822 - step: 2 scores: rmse: 0.4400419785163047 baseline_rmse: 0.8227835642459033 nrmse: 0.5348210606511222 only_a_distance_nrmse: 0.7906639334091937 source_prerequisite_met: false downstream_interpretable: false only_a_oracle: rmse: 0.5791384491604891 baseline_rmse: 0.8103532807277047 nrmse: 0.7146740353052158 only_a_distance_nrmse: 0.0 arms: deleted: before: rmse: 0.7271518158927684 baseline_rmse: 0.8103532807277047 nrmse: 0.8973269229437554 only_a_distance_nrmse: 0.5529407654147872 after: rmse: 0.724493864955975 baseline_rmse: 0.8103532807277047 nrmse: 0.8940469325987954 only_a_distance_nrmse: 0.5337538704664715 steps: 2 optimizer_covers_current_parameters: true validation_curve: - step: 0 scores: rmse: 0.7479704676843564 baseline_rmse: 0.8185735883959189 nrmse: 0.9137485966901073 only_a_distance_nrmse: 0.5496018825796432 - step: 2 scores: rmse: 0.7124954308348078 baseline_rmse: 0.8185735883959189 nrmse: 0.8704109696857159 only_a_distance_nrmse: 0.5287892518881139 updated_parameter_count: 77 pooling_capacity: 2 removed_state_entries: - nodes.record/b.embedder.weights - nodes.record/b.embedder.max_fourier_input - nodes.record/b.embedder.embeddings.weight - nodes.record/b.embedder.counter.counts - nodes.record/b.embedder.linear.weight - nodes.record/b.embedder.linear.bias - nodes.record/b.embedder.scalar.weight - nodes.record/b.embedder.normalizer.mean - nodes.record/b.embedder.normalizer.var - nodes.record/b.embedder.normalizer.count changed_unselected_state_entries: [] inactive: before: rmse: 0.7271518158927684 baseline_rmse: 0.8103532807277047 nrmse: 0.8973269229437554 only_a_distance_nrmse: 0.5529407654147872 after: rmse: 0.724493864955975 baseline_rmse: 0.8103532807277047 nrmse: 0.8940469325987954 only_a_distance_nrmse: 0.5337538704664715 steps: 2 optimizer_covers_current_parameters: true validation_curve: - step: 0 scores: rmse: 0.7479704676843564 baseline_rmse: 0.8185735883959189 nrmse: 0.9137485966901073 only_a_distance_nrmse: 0.5496018825796432 - step: 2 scores: rmse: 0.7124954308348078 baseline_rmse: 0.8185735883959189 nrmse: 0.8704109696857159 only_a_distance_nrmse: 0.5287892518881139 updated_parameter_count: 77 pooling_capacity: 2 scratch_restricted: before: rmse: 1.0277799777901702 baseline_rmse: 0.8103532807277047 nrmse: 1.268310997479043 only_a_distance_nrmse: 1.0606146291576006 after: rmse: 0.7628028335183838 baseline_rmse: 0.8103532807277047 nrmse: 0.9413213368290184 only_a_distance_nrmse: 0.6038361586464066 steps: 2 optimizer_covers_current_parameters: true validation_curve: - step: 0 scores: rmse: 1.050771858874917 baseline_rmse: 0.8185735883959189 nrmse: 1.283662060162502 only_a_distance_nrmse: 1.0486131884111216 - step: 2 scores: rmse: 0.7511156090856626 baseline_rmse: 0.8185735883959189 nrmse: 0.9175908186306776 only_a_distance_nrmse: 0.6028711571904821 updated_parameter_count: 77 pooling_capacity: 2 continuation: before: rmse: 0.4408785755960084 baseline_rmse: 0.8103532807277047 nrmse: 0.5440572477229874 only_a_distance_nrmse: 0.8002568809338788 after: rmse: 0.5387691939859782 baseline_rmse: 0.8103532807277047 nrmse: 0.6648571762455981 only_a_distance_nrmse: 0.7777005031350472 steps: 2 optimizer_covers_current_parameters: true validation_curve: - step: 0 scores: rmse: 0.44655369235189984 baseline_rmse: 0.8185735883959189 nrmse: 0.5455266315481407 only_a_distance_nrmse: 0.8152021444503296 - step: 2 scores: rmse: 0.5509526345815351 baseline_rmse: 0.8185735883959189 nrmse: 0.6730642698369792 only_a_distance_nrmse: 0.7482000327420087 updated_parameter_count: 89 pooling_capacity: 3 readdition: scores: rmse: 0.4757161400717045 baseline_rmse: 0.8103532807277047 nrmse: 0.5870478362776628 only_a_distance_nrmse: 0.5407052956250064 source_normalization: mean: -0.03621933236718178 variance: 0.31412187218666077 std: 0.5604746846974097 count: 256 alpha: null fresh_normalization: mean: 0.0 variance: 1.0 std: 1.0000049999875 count: 0 alpha: null source_field_order: - record/a - record/b - record/y readded_field_order: - record/a - record/y - record/b test_rows: 2048 optimizer_policy: New AdamW factory and Trainer per phase; matched adaptation data and updates mutation: delete record/b; inactive control updates active=False; separate source fork deletes and re-adds b checks: Source learns both-input relationship below 0.25 nRMSE: false Source optimizer covers current parameters: true Deletion removes b from schema and runtime: true Deletion removes the selected state entries: true Deletion preserves all unselected state entries: true Deletion immediately loses the selected information: true 'deleted before: removed input values cannot affect predictions': true 'deleted: optimizer covers current parameters': true 'deleted: parameters actually learn': true 'deleted: hidden target values cannot affect predictions': true 'deleted: adapts near the remaining-information limit': false 'deleted after: removed input values cannot affect predictions': true Deleted checkpoint preserves schema and learned state: true Deleted checkpoint preserves adapted predictions: true Inactive control preserves all original state: true Deletion and deactivation initially give the same predictions: true 'inactive before: removed input values cannot affect predictions': true 'inactive: optimizer covers current parameters': true 'inactive: parameters actually learn': true 'inactive: hidden target values cannot affect predictions': true 'inactive: adapts near the remaining-information limit': false 'inactive after: removed input values cannot affect predictions': true Scratch schema and field order match deletion: true 'scratch_restricted before: removed input values cannot affect predictions': true 'scratch_restricted: optimizer covers current parameters': true 'scratch_restricted: parameters actually learn': true 'scratch_restricted: hidden target values cannot affect predictions': true 'scratch_restricted: adapts near the remaining-information limit': false 'scratch_restricted after: removed input values cannot affect predictions': true Unchanged checkpoint restores source predictions: true 'continuation: optimizer covers current parameters': true 'continuation: parameters actually learn': true 'continuation: hidden target values cannot affect predictions': true Continuation retains full-information accuracy below 0.25 nRMSE: false Re-adding the same name does not restore its learned state: true Re-added input has fresh normalization: true outcome: not_met duration_seconds: 1.068 - started_at: '2026-09-15T18:22:05.979454+00:00' mode: full seed: 5001 steps_override: null accelerator: cpu threads: 1 provenance: revision: 5d985953fd0416bc5982a3518d8f43cadb52110c source_sha256: de366e85036478c26153e1ff1f82d3a7faec6fd712790249662f6d5fea48ec37 script_sha256: fe688bd87233e2bfcd5b1c5ca6916536a0049dfd012699881fa7eeb6edb1f487 code_sha256: 71ad78be90abfc30fba52b2e0182b4539ca9101f868befec33b05fdf349d023a python: 3.12.10 platform: macOS-26.5.2-arm64-arm-64bit host: Granthams-MacBook-Pro.local relflow_path: /Users/grantham/Desktop/relflow/src/relflow packages: relflow: 0.2.2 torch: 2.12.0 lightning: 2.6.4 numpy: 2.4.6 pyarrow: 24.0.0 polars: 1.44.1 torchmetrics: 1.9.0 tensordict: 0.12.4 pydantic: 2.13.4 cuda: null gpu: null metrics: source: rmse: 0.030233970450265445 baseline_rmse: 0.8103532807277047 nrmse: 0.03730961689093806 only_a_distance_nrmse: 0.728618107549543 source_fit: steps: 512 optimizer_covers_current_parameters: true validation_curve: - step: 0 scores: rmse: 1.1152491751251399 baseline_rmse: 0.8227835642459033 nrmse: 1.3554587422359208 only_a_distance_nrmse: 0.9971723998482822 - step: 32 scores: rmse: 0.05782952452186843 baseline_rmse: 0.8227835642459033 nrmse: 0.07028522084646925 only_a_distance_nrmse: 0.6607188312497041 - step: 128 scores: rmse: 0.052864771156084595 baseline_rmse: 0.8227835642459033 nrmse: 0.06425112684954537 only_a_distance_nrmse: 0.7206757033372879 - step: 512 scores: rmse: 0.02961701962644633 baseline_rmse: 0.8227835642459033 nrmse: 0.03599612451373028 only_a_distance_nrmse: 0.7132140583987974 source_prerequisite_met: true downstream_interpretable: true only_a_oracle: rmse: 0.5791384491604891 baseline_rmse: 0.8103532807277047 nrmse: 0.7146740353052158 only_a_distance_nrmse: 0.0 arms: deleted: before: rmse: 0.690778493781905 baseline_rmse: 0.8103532807277047 nrmse: 0.8524411638854347 only_a_distance_nrmse: 0.46328294230539013 after: rmse: 0.5842682661881069 baseline_rmse: 0.8103532807277047 nrmse: 0.7210043817721433 only_a_distance_nrmse: 0.07714394439231267 steps: 256 optimizer_covers_current_parameters: true validation_curve: - step: 0 scores: rmse: 0.6940508255036514 baseline_rmse: 0.8185735883959189 nrmse: 0.8478783524688562 only_a_distance_nrmse: 0.4634211862312215 - step: 32 scores: rmse: 0.5962430241585547 baseline_rmse: 0.8185735883959189 nrmse: 0.7283926975056154 only_a_distance_nrmse: 0.14292109841311226 - step: 128 scores: rmse: 0.5884883191315411 baseline_rmse: 0.8185735883959189 nrmse: 0.7189192608629676 only_a_distance_nrmse: 0.14937961596717464 - step: 256 scores: rmse: 0.5845684618941069 baseline_rmse: 0.8185735883959189 nrmse: 0.7141306171869415 only_a_distance_nrmse: 0.07933424217544038 updated_parameter_count: 77 pooling_capacity: 2 removed_state_entries: - nodes.record/b.embedder.weights - nodes.record/b.embedder.max_fourier_input - nodes.record/b.embedder.embeddings.weight - nodes.record/b.embedder.counter.counts - nodes.record/b.embedder.linear.weight - nodes.record/b.embedder.linear.bias - nodes.record/b.embedder.scalar.weight - nodes.record/b.embedder.normalizer.mean - nodes.record/b.embedder.normalizer.var - nodes.record/b.embedder.normalizer.count changed_unselected_state_entries: [] inactive: before: rmse: 0.690778493781905 baseline_rmse: 0.8103532807277047 nrmse: 0.8524411638854347 only_a_distance_nrmse: 0.46328294230539013 after: rmse: 0.5842682661881069 baseline_rmse: 0.8103532807277047 nrmse: 0.7210043817721433 only_a_distance_nrmse: 0.07714394439231267 steps: 256 optimizer_covers_current_parameters: true validation_curve: - step: 0 scores: rmse: 0.6940508255036514 baseline_rmse: 0.8185735883959189 nrmse: 0.8478783524688562 only_a_distance_nrmse: 0.4634211862312215 - step: 32 scores: rmse: 0.5962430241585547 baseline_rmse: 0.8185735883959189 nrmse: 0.7283926975056154 only_a_distance_nrmse: 0.14292109841311226 - step: 128 scores: rmse: 0.5884883191315411 baseline_rmse: 0.8185735883959189 nrmse: 0.7189192608629676 only_a_distance_nrmse: 0.14937961596717464 - step: 256 scores: rmse: 0.5845684618941069 baseline_rmse: 0.8185735883959189 nrmse: 0.7141306171869415 only_a_distance_nrmse: 0.07933424217544038 updated_parameter_count: 77 pooling_capacity: 2 scratch_restricted: before: rmse: 1.0277799777901702 baseline_rmse: 0.8103532807277047 nrmse: 1.268310997479043 only_a_distance_nrmse: 1.0606146291576006 after: rmse: 0.5831567734064373 baseline_rmse: 0.8103532807277047 nrmse: 0.7196327666900505 only_a_distance_nrmse: 0.06692009098405423 steps: 256 optimizer_covers_current_parameters: true validation_curve: - step: 0 scores: rmse: 1.050771858874917 baseline_rmse: 0.8185735883959189 nrmse: 1.283662060162502 only_a_distance_nrmse: 1.0486131884111216 - step: 32 scores: rmse: 0.5823476303396594 baseline_rmse: 0.8185735883959189 nrmse: 0.7114175666000059 only_a_distance_nrmse: 0.0577569667245416 - step: 128 scores: rmse: 0.5892498555184403 baseline_rmse: 0.8185735883959189 nrmse: 0.7198495820921091 only_a_distance_nrmse: 0.15640313331972078 - step: 256 scores: rmse: 0.5832848278837531 baseline_rmse: 0.8185735883959189 nrmse: 0.7125624820448473 only_a_distance_nrmse: 0.06956897090926455 updated_parameter_count: 77 pooling_capacity: 2 continuation: before: rmse: 0.030233970450265445 baseline_rmse: 0.8103532807277047 nrmse: 0.03730961689093806 only_a_distance_nrmse: 0.728618107549543 after: rmse: 0.013617242093645654 baseline_rmse: 0.8103532807277047 nrmse: 0.016804080908288846 only_a_distance_nrmse: 0.7173870300374742 steps: 256 optimizer_covers_current_parameters: true validation_curve: - step: 0 scores: rmse: 0.029924764154731227 baseline_rmse: 0.8185735883959189 nrmse: 0.03655720704765463 only_a_distance_nrmse: 0.7180107575257996 - step: 32 scores: rmse: 0.07121546077745729 baseline_rmse: 0.8185735883959189 nrmse: 0.08699946075344488 only_a_distance_nrmse: 0.6795849223812774 - step: 128 scores: rmse: 0.008683239416743024 baseline_rmse: 0.8185735883959189 nrmse: 0.01060776885528245 only_a_distance_nrmse: 0.7073769509370645 - step: 256 scores: rmse: 0.013771687757451129 baseline_rmse: 0.8185735883959189 nrmse: 0.016824006971002083 only_a_distance_nrmse: 0.708008635122541 updated_parameter_count: 89 pooling_capacity: 3 readdition: scores: rmse: 0.6852289823134198 baseline_rmse: 0.8103532807277047 nrmse: 0.845592902021792 only_a_distance_nrmse: 0.7385194119698516 source_normalization: mean: -0.0004043062508571893 variance: 0.3330785930156708 std: 0.5771382789381335 count: 65536 alpha: null fresh_normalization: mean: 0.0 variance: 1.0 std: 1.0000049999875 count: 0 alpha: null source_field_order: - record/a - record/b - record/y readded_field_order: - record/a - record/y - record/b test_rows: 2048 optimizer_policy: New AdamW factory and Trainer per phase; matched adaptation data and updates mutation: delete record/b; inactive control updates active=False; separate source fork deletes and re-adds b checks: Source learns both-input relationship below 0.25 nRMSE: true Source optimizer covers current parameters: true Deletion removes b from schema and runtime: true Deletion removes the selected state entries: true Deletion preserves all unselected state entries: true Deletion immediately loses the selected information: true 'deleted before: removed input values cannot affect predictions': true 'deleted: optimizer covers current parameters': true 'deleted: parameters actually learn': true 'deleted: hidden target values cannot affect predictions': true 'deleted: adapts near the remaining-information limit': true 'deleted after: removed input values cannot affect predictions': true Deleted checkpoint preserves schema and learned state: true Deleted checkpoint preserves adapted predictions: true Inactive control preserves all original state: true Deletion and deactivation initially give the same predictions: true 'inactive before: removed input values cannot affect predictions': true 'inactive: optimizer covers current parameters': true 'inactive: parameters actually learn': true 'inactive: hidden target values cannot affect predictions': true 'inactive: adapts near the remaining-information limit': true 'inactive after: removed input values cannot affect predictions': true Scratch schema and field order match deletion: true 'scratch_restricted before: removed input values cannot affect predictions': true 'scratch_restricted: optimizer covers current parameters': true 'scratch_restricted: parameters actually learn': true 'scratch_restricted: hidden target values cannot affect predictions': true 'scratch_restricted: adapts near the remaining-information limit': true 'scratch_restricted after: removed input values cannot affect predictions': true Unchanged checkpoint restores source predictions: true 'continuation: optimizer covers current parameters': true 'continuation: parameters actually learn': true 'continuation: hidden target values cannot affect predictions': true Continuation retains full-information accuracy below 0.25 nRMSE: true Re-adding the same name does not restore its learned state: true Re-added input has fresh normalization: true outcome: met duration_seconds: 13.82 - started_at: '2026-09-15T18:23:31.594803+00:00' mode: full seed: 5001 steps_override: null accelerator: gpu threads: 1 provenance: revision: null source_sha256: de366e85036478c26153e1ff1f82d3a7faec6fd712790249662f6d5fea48ec37 script_sha256: fe688bd87233e2bfcd5b1c5ca6916536a0049dfd012699881fa7eeb6edb1f487 code_sha256: 71ad78be90abfc30fba52b2e0182b4539ca9101f868befec33b05fdf349d023a python: 3.12.6 platform: Linux-6.8.0-138-generic-x86_64-with-glibc2.35 host: lab relflow_path: /home/grantham/relflow-removal-6b17okdx/src/relflow packages: relflow: 0.2.2 torch: 2.12.0 lightning: 2.6.4 numpy: 2.4.6 pyarrow: 24.0.0 polars: 1.44.1 torchmetrics: 1.9.0 tensordict: 0.12.4 pydantic: 2.13.4 cuda: '13.0' gpu: NVIDIA GeForce RTX 3090 metrics: source: rmse: 0.034193175689539355 baseline_rmse: 0.8103532807277047 nrmse: 0.04219539366686289 only_a_distance_nrmse: 0.7263763636959647 source_fit: steps: 512 optimizer_covers_current_parameters: true validation_curve: - step: 0 scores: rmse: 1.1152492049615597 baseline_rmse: 0.8227835642459033 nrmse: 1.3554587784987011 only_a_distance_nrmse: 0.9971724517883713 - step: 32 scores: rmse: 0.057829628655470124 baseline_rmse: 0.8227835642459033 nrmse: 0.07028534740903833 only_a_distance_nrmse: 0.6607187908389289 - step: 128 scores: rmse: 0.05289033813159905 baseline_rmse: 0.8227835642459033 nrmse: 0.0642822006053002 only_a_distance_nrmse: 0.7206997159576051 - step: 512 scores: rmse: 0.03409068593820913 baseline_rmse: 0.8227835642459033 nrmse: 0.04143335795660173 only_a_distance_nrmse: 0.7134729135579938 source_prerequisite_met: true downstream_interpretable: true only_a_oracle: rmse: 0.5791384491604891 baseline_rmse: 0.8103532807277047 nrmse: 0.7146740353052158 only_a_distance_nrmse: 0.0 arms: deleted: before: rmse: 0.6869920652316659 baseline_rmse: 0.8103532807277047 nrmse: 0.8477685986718542 only_a_distance_nrmse: 0.4544723842827458 after: rmse: 0.5844560505348129 baseline_rmse: 0.8103532807277047 nrmse: 0.7212361132294867 only_a_distance_nrmse: 0.07872369638242804 steps: 256 optimizer_covers_current_parameters: true validation_curve: - step: 0 scores: rmse: 0.6876196483060345 baseline_rmse: 0.8185735883959189 nrmse: 0.8400217867443018 only_a_distance_nrmse: 0.4543510302166673 - step: 32 scores: rmse: 0.5889345966290531 baseline_rmse: 0.8185735883959189 nrmse: 0.7194644500846069 only_a_distance_nrmse: 0.10265846999341365 - step: 128 scores: rmse: 0.5879907604012774 baseline_rmse: 0.8185735883959189 nrmse: 0.7183114245763869 only_a_distance_nrmse: 0.14707627628724834 - step: 256 scores: rmse: 0.5845773005020589 baseline_rmse: 0.8185735883959189 nrmse: 0.7141414147597892 only_a_distance_nrmse: 0.08087778371578416 updated_parameter_count: 89 pooling_capacity: 2 removed_state_entries: - nodes.record/b.embedder.weights - nodes.record/b.embedder.max_fourier_input - nodes.record/b.embedder.embeddings.weight - nodes.record/b.embedder.counter.counts - nodes.record/b.embedder.linear.weight - nodes.record/b.embedder.linear.bias - nodes.record/b.embedder.scalar.weight - nodes.record/b.embedder.normalizer.mean - nodes.record/b.embedder.normalizer.var - nodes.record/b.embedder.normalizer.count changed_unselected_state_entries: [] inactive: before: rmse: 0.6869920652316659 baseline_rmse: 0.8103532807277047 nrmse: 0.8477685986718542 only_a_distance_nrmse: 0.4544723842827458 after: rmse: 0.5844560505348129 baseline_rmse: 0.8103532807277047 nrmse: 0.7212361132294867 only_a_distance_nrmse: 0.07872369638242804 steps: 256 optimizer_covers_current_parameters: true validation_curve: - step: 0 scores: rmse: 0.6876196483060345 baseline_rmse: 0.8185735883959189 nrmse: 0.8400217867443018 only_a_distance_nrmse: 0.4543510302166673 - step: 32 scores: rmse: 0.5889345966290531 baseline_rmse: 0.8185735883959189 nrmse: 0.7194644500846069 only_a_distance_nrmse: 0.10265846999341365 - step: 128 scores: rmse: 0.5879907604012774 baseline_rmse: 0.8185735883959189 nrmse: 0.7183114245763869 only_a_distance_nrmse: 0.14707627628724834 - step: 256 scores: rmse: 0.5845773005020589 baseline_rmse: 0.8185735883959189 nrmse: 0.7141414147597892 only_a_distance_nrmse: 0.08087778371578416 updated_parameter_count: 89 pooling_capacity: 2 scratch_restricted: before: rmse: 1.0277799965979517 baseline_rmse: 0.8103532807277047 nrmse: 1.2683110206884036 only_a_distance_nrmse: 1.06061465559322 after: rmse: 0.5831567760991642 baseline_rmse: 0.8103532807277047 nrmse: 0.7196327700129553 only_a_distance_nrmse: 0.06692010222859263 steps: 256 optimizer_covers_current_parameters: true validation_curve: - step: 0 scores: rmse: 1.0507718721225583 baseline_rmse: 0.8185735883959189 nrmse: 1.283662076346314 only_a_distance_nrmse: 1.048613208499735 - step: 32 scores: rmse: 0.5823476208361499 baseline_rmse: 0.8185735883959189 nrmse: 0.7114175549901645 only_a_distance_nrmse: 0.05775696373949843 - step: 128 scores: rmse: 0.5892498569077196 baseline_rmse: 0.8185735883959189 nrmse: 0.7198495837893044 only_a_distance_nrmse: 0.15640314885348733 - step: 256 scores: rmse: 0.5832848293252761 baseline_rmse: 0.8185735883959189 nrmse: 0.7125624838058655 only_a_distance_nrmse: 0.06956898556800413 updated_parameter_count: 89 pooling_capacity: 2 continuation: before: rmse: 0.034193175689539355 baseline_rmse: 0.8103532807277047 nrmse: 0.04219539366686289 only_a_distance_nrmse: 0.7263763636959647 after: rmse: 0.017375520799651097 baseline_rmse: 0.8103532807277047 nrmse: 0.021441908378587324 only_a_distance_nrmse: 0.7010713008943545 steps: 256 optimizer_covers_current_parameters: true validation_curve: - step: 0 scores: rmse: 0.035068273201261636 baseline_rmse: 0.8185735883959189 nrmse: 0.04284070937346221 only_a_distance_nrmse: 0.7203080813425458 - step: 32 scores: rmse: 0.08031614652766814 baseline_rmse: 0.8185735883959189 nrmse: 0.09811719760596734 only_a_distance_nrmse: 0.7179605857986787 - step: 128 scores: rmse: 0.004807879730056251 baseline_rmse: 0.8185735883959189 nrmse: 0.005873485045465243 only_a_distance_nrmse: 0.705783136099351 - step: 256 scores: rmse: 0.016764382588168907 baseline_rmse: 0.8185735883959189 nrmse: 0.020479994499969734 only_a_distance_nrmse: 0.6931467450606643 updated_parameter_count: 93 pooling_capacity: 3 readdition: scores: rmse: 0.7129609305304768 baseline_rmse: 0.8103532807277047 nrmse: 0.8798149492160151 only_a_distance_nrmse: 0.7605748091272082 source_normalization: mean: -0.0004043062508571893 variance: 0.3330785930156708 std: 0.5771382789381335 count: 65536 alpha: null fresh_normalization: mean: 0.0 variance: 1.0 std: 1.0000049999875 count: 0 alpha: null source_field_order: - record/a - record/b - record/y readded_field_order: - record/a - record/y - record/b test_rows: 2048 optimizer_policy: New AdamW factory and Trainer per phase; matched adaptation data and updates mutation: delete record/b; inactive control updates active=False; separate source fork deletes and re-adds b checks: Source learns both-input relationship below 0.25 nRMSE: true Source optimizer covers current parameters: true Deletion removes b from schema and runtime: true Deletion removes the selected state entries: true Deletion preserves all unselected state entries: true Deletion immediately loses the selected information: true 'deleted before: removed input values cannot affect predictions': true 'deleted: optimizer covers current parameters': true 'deleted: parameters actually learn': true 'deleted: hidden target values cannot affect predictions': true 'deleted: adapts near the remaining-information limit': true 'deleted after: removed input values cannot affect predictions': true Deleted checkpoint preserves schema and learned state: true Deleted checkpoint preserves adapted predictions: true Inactive control preserves all original state: true Deletion and deactivation initially give the same predictions: true 'inactive before: removed input values cannot affect predictions': true 'inactive: optimizer covers current parameters': true 'inactive: parameters actually learn': true 'inactive: hidden target values cannot affect predictions': true 'inactive: adapts near the remaining-information limit': true 'inactive after: removed input values cannot affect predictions': true Scratch schema and field order match deletion: true 'scratch_restricted before: removed input values cannot affect predictions': true 'scratch_restricted: optimizer covers current parameters': true 'scratch_restricted: parameters actually learn': true 'scratch_restricted: hidden target values cannot affect predictions': true 'scratch_restricted: adapts near the remaining-information limit': true 'scratch_restricted after: removed input values cannot affect predictions': true Unchanged checkpoint restores source predictions: true 'continuation: optimizer covers current parameters': true 'continuation: parameters actually learn': true 'continuation: hidden target values cannot affect predictions': true Continuation retains full-information accuracy below 0.25 nRMSE: true Re-adding the same name does not restore its learned state: true Re-added input has fresh normalization: true outcome: met duration_seconds: 30.909 - started_at: '2026-09-15T18:24:22.957999+00:00' mode: full seed: 7201 steps_override: null accelerator: gpu threads: 1 provenance: revision: null source_sha256: de366e85036478c26153e1ff1f82d3a7faec6fd712790249662f6d5fea48ec37 script_sha256: fe688bd87233e2bfcd5b1c5ca6916536a0049dfd012699881fa7eeb6edb1f487 code_sha256: 71ad78be90abfc30fba52b2e0182b4539ca9101f868befec33b05fdf349d023a python: 3.12.6 platform: Linux-6.8.0-138-generic-x86_64-with-glibc2.35 host: lab relflow_path: /home/grantham/relflow-removal-6b17okdx/src/relflow packages: relflow: 0.2.2 torch: 2.12.0 lightning: 2.6.4 numpy: 2.4.6 pyarrow: 24.0.0 polars: 1.44.1 torchmetrics: 1.9.0 tensordict: 0.12.4 pydantic: 2.13.4 cuda: '13.0' gpu: NVIDIA GeForce RTX 3090 metrics: source: rmse: 0.012387407114893877 baseline_rmse: 0.8120016401971133 nrmse: 0.015255396666300866 only_a_distance_nrmse: 0.7136397917074613 source_fit: steps: 512 optimizer_covers_current_parameters: true validation_curve: - step: 0 scores: rmse: 0.9848706373195013 baseline_rmse: 0.8396052345162109 nrmse: 1.173016313895415 only_a_distance_nrmse: 0.7962542109945072 - step: 32 scores: rmse: 0.052390351564096396 baseline_rmse: 0.8396052345162109 nrmse: 0.06239879101549938 only_a_distance_nrmse: 0.6842872372979514 - step: 128 scores: rmse: 0.015064316126240537 baseline_rmse: 0.8396052345162109 nrmse: 0.01794214174345965 only_a_distance_nrmse: 0.6847022180759422 - step: 512 scores: rmse: 0.01243701868450528 baseline_rmse: 0.8396052345162109 nrmse: 0.01481293609570171 only_a_distance_nrmse: 0.6891333235742876 source_prerequisite_met: true downstream_interpretable: true only_a_oracle: rmse: 0.5809035857842813 baseline_rmse: 0.8120016401971133 nrmse: 0.7153970595961691 only_a_distance_nrmse: 0.0 arms: deleted: before: rmse: 0.6978588819705238 baseline_rmse: 0.8120016401971133 nrmse: 0.8594303846493693 only_a_distance_nrmse: 0.4683118781791459 after: rmse: 0.5831621066597686 baseline_rmse: 0.8120016401971133 nrmse: 0.7181784836274544 only_a_distance_nrmse: 0.07859965746781736 steps: 256 optimizer_covers_current_parameters: true validation_curve: - step: 0 scores: rmse: 0.6901263734392206 baseline_rmse: 0.8256454947735132 nrmse: 0.8358628222497991 only_a_distance_nrmse: 0.4671020486917348 - step: 32 scores: rmse: 0.5880209811528249 baseline_rmse: 0.8256454947735132 nrmse: 0.7121954699384967 only_a_distance_nrmse: 0.09196330366955971 - step: 128 scores: rmse: 0.5903828080975378 baseline_rmse: 0.8256454947735132 nrmse: 0.7150560523066726 only_a_distance_nrmse: 0.11623862002519912 - step: 256 scores: rmse: 0.5808657463437441 baseline_rmse: 0.8256454947735132 nrmse: 0.7035292386632402 only_a_distance_nrmse: 0.07769698358159956 updated_parameter_count: 89 pooling_capacity: 2 removed_state_entries: - nodes.record/b.embedder.weights - nodes.record/b.embedder.max_fourier_input - nodes.record/b.embedder.embeddings.weight - nodes.record/b.embedder.counter.counts - nodes.record/b.embedder.linear.weight - nodes.record/b.embedder.linear.bias - nodes.record/b.embedder.scalar.weight - nodes.record/b.embedder.normalizer.mean - nodes.record/b.embedder.normalizer.var - nodes.record/b.embedder.normalizer.count changed_unselected_state_entries: [] inactive: before: rmse: 0.6978588819705238 baseline_rmse: 0.8120016401971133 nrmse: 0.8594303846493693 only_a_distance_nrmse: 0.4683118781791459 after: rmse: 0.5831621066597686 baseline_rmse: 0.8120016401971133 nrmse: 0.7181784836274544 only_a_distance_nrmse: 0.07859965746781736 steps: 256 optimizer_covers_current_parameters: true validation_curve: - step: 0 scores: rmse: 0.6901263734392206 baseline_rmse: 0.8256454947735132 nrmse: 0.8358628222497991 only_a_distance_nrmse: 0.4671020486917348 - step: 32 scores: rmse: 0.5880209811528249 baseline_rmse: 0.8256454947735132 nrmse: 0.7121954699384967 only_a_distance_nrmse: 0.09196330366955971 - step: 128 scores: rmse: 0.5903828080975378 baseline_rmse: 0.8256454947735132 nrmse: 0.7150560523066726 only_a_distance_nrmse: 0.11623862002519912 - step: 256 scores: rmse: 0.5808657463437441 baseline_rmse: 0.8256454947735132 nrmse: 0.7035292386632402 only_a_distance_nrmse: 0.07769698358159956 updated_parameter_count: 89 pooling_capacity: 2 scratch_restricted: before: rmse: 1.244657045306195 baseline_rmse: 0.8120016401971133 nrmse: 1.5328257773026845 only_a_distance_nrmse: 1.376798698767851 after: rmse: 0.581655971541918 baseline_rmse: 0.8120016401971133 nrmse: 0.7163236411698886 only_a_distance_nrmse: 0.05921633754258864 steps: 256 optimizer_covers_current_parameters: true validation_curve: - step: 0 scores: rmse: 1.2326355702152973 baseline_rmse: 0.8256454947735132 nrmse: 1.4929356219080776 only_a_distance_nrmse: 1.3516029008577253 - step: 32 scores: rmse: 0.5881523347689552 baseline_rmse: 0.8256454947735132 nrmse: 0.7123545619664455 only_a_distance_nrmse: 0.0967495088065333 - step: 128 scores: rmse: 0.5901936678822421 baseline_rmse: 0.8256454947735132 nrmse: 0.7148269706772166 only_a_distance_nrmse: 0.11287399450786655 - step: 256 scores: rmse: 0.5818092520166288 baseline_rmse: 0.8256454947735132 nrmse: 0.7046719877957157 only_a_distance_nrmse: 0.05865919066520842 updated_parameter_count: 89 pooling_capacity: 2 continuation: before: rmse: 0.012387407114893877 baseline_rmse: 0.8120016401971133 nrmse: 0.015255396666300866 only_a_distance_nrmse: 0.7136397917074613 after: rmse: 0.010233279181010285 baseline_rmse: 0.8120016401971133 nrmse: 0.012602535111291347 only_a_distance_nrmse: 0.7135734304229762 steps: 256 optimizer_covers_current_parameters: true validation_curve: - step: 0 scores: rmse: 0.012024774168433016 baseline_rmse: 0.8256454947735132 nrmse: 0.014564088636772117 only_a_distance_nrmse: 0.6996642910923084 - step: 32 scores: rmse: 0.042896468824362474 baseline_rmse: 0.8256454947735132 nrmse: 0.051955069210581246 only_a_distance_nrmse: 0.6805924397417519 - step: 128 scores: rmse: 0.018348379425840437 baseline_rmse: 0.8256454947735132 nrmse: 0.02222307218048064 only_a_distance_nrmse: 0.6965932513607549 - step: 256 scores: rmse: 0.009791922153065192 baseline_rmse: 0.8256454947735132 nrmse: 0.011859717293983735 only_a_distance_nrmse: 0.6998416283507768 updated_parameter_count: 93 pooling_capacity: 3 readdition: scores: rmse: 0.5735670565285867 baseline_rmse: 0.8120016401971133 nrmse: 0.7063619432952788 only_a_distance_nrmse: 0.22241200155948565 source_normalization: mean: 0.02379218488931656 variance: 0.33199870586395264 std: 0.5762019662097246 count: 65536 alpha: null fresh_normalization: mean: 0.0 variance: 1.0 std: 1.0000049999875 count: 0 alpha: null source_field_order: - record/a - record/b - record/y readded_field_order: - record/a - record/y - record/b test_rows: 2048 optimizer_policy: New AdamW factory and Trainer per phase; matched adaptation data and updates mutation: delete record/b; inactive control updates active=False; separate source fork deletes and re-adds b checks: Source learns both-input relationship below 0.25 nRMSE: true Source optimizer covers current parameters: true Deletion removes b from schema and runtime: true Deletion removes the selected state entries: true Deletion preserves all unselected state entries: true Deletion immediately loses the selected information: true 'deleted before: removed input values cannot affect predictions': true 'deleted: optimizer covers current parameters': true 'deleted: parameters actually learn': true 'deleted: hidden target values cannot affect predictions': true 'deleted: adapts near the remaining-information limit': true 'deleted after: removed input values cannot affect predictions': true Deleted checkpoint preserves schema and learned state: true Deleted checkpoint preserves adapted predictions: true Inactive control preserves all original state: true Deletion and deactivation initially give the same predictions: true 'inactive before: removed input values cannot affect predictions': true 'inactive: optimizer covers current parameters': true 'inactive: parameters actually learn': true 'inactive: hidden target values cannot affect predictions': true 'inactive: adapts near the remaining-information limit': true 'inactive after: removed input values cannot affect predictions': true Scratch schema and field order match deletion: true 'scratch_restricted before: removed input values cannot affect predictions': true 'scratch_restricted: optimizer covers current parameters': true 'scratch_restricted: parameters actually learn': true 'scratch_restricted: hidden target values cannot affect predictions': true 'scratch_restricted: adapts near the remaining-information limit': true 'scratch_restricted after: removed input values cannot affect predictions': true Unchanged checkpoint restores source predictions: true 'continuation: optimizer covers current parameters': true 'continuation: parameters actually learn': true 'continuation: hidden target values cannot affect predictions': true Continuation retains full-information accuracy below 0.25 nRMSE: true Re-adding the same name does not restore its learned state: true Re-added input has fresh normalization: true outcome: met duration_seconds: 31.006 - started_at: '2026-09-15T18:25:14.574219+00:00' mode: full seed: 7202 steps_override: null accelerator: gpu threads: 1 provenance: revision: null source_sha256: de366e85036478c26153e1ff1f82d3a7faec6fd712790249662f6d5fea48ec37 script_sha256: fe688bd87233e2bfcd5b1c5ca6916536a0049dfd012699881fa7eeb6edb1f487 code_sha256: 71ad78be90abfc30fba52b2e0182b4539ca9101f868befec33b05fdf349d023a python: 3.12.6 platform: Linux-6.8.0-138-generic-x86_64-with-glibc2.35 host: lab relflow_path: /home/grantham/relflow-removal-6b17okdx/src/relflow packages: relflow: 0.2.2 torch: 2.12.0 lightning: 2.6.4 numpy: 2.4.6 pyarrow: 24.0.0 polars: 1.44.1 torchmetrics: 1.9.0 tensordict: 0.12.4 pydantic: 2.13.4 cuda: '13.0' gpu: NVIDIA GeForce RTX 3090 metrics: source: rmse: 0.02008381101831335 baseline_rmse: 0.8352833232272557 nrmse: 0.02404430982856956 only_a_distance_nrmse: 0.7083488643135359 source_fit: steps: 512 optimizer_covers_current_parameters: true validation_curve: - step: 0 scores: rmse: 0.6233581094490702 baseline_rmse: 0.818524820858298 nrmse: 0.7615628672022704 only_a_distance_nrmse: 0.7208547812234832 - step: 32 scores: rmse: 0.1094425294716588 baseline_rmse: 0.818524820858298 nrmse: 0.1337070381774108 only_a_distance_nrmse: 0.7307140954775097 - step: 128 scores: rmse: 0.016519101916772697 baseline_rmse: 0.818524820858298 nrmse: 0.02018155283238804 only_a_distance_nrmse: 0.7153357098778558 - step: 512 scores: rmse: 0.019615138964158296 baseline_rmse: 0.818524820858298 nrmse: 0.02396401240904342 only_a_distance_nrmse: 0.7285638406273279 source_prerequisite_met: true downstream_interpretable: true only_a_oracle: rmse: 0.5780872291911047 baseline_rmse: 0.8352833232272557 nrmse: 0.6920852040449805 only_a_distance_nrmse: 0.0 arms: deleted: before: rmse: 0.8051448159247964 baseline_rmse: 0.8352833232272557 nrmse: 0.963918222159621 only_a_distance_nrmse: 0.7052061517161763 after: rmse: 0.5778075411583316 baseline_rmse: 0.8352833232272557 nrmse: 0.691750361932136 only_a_distance_nrmse: 0.06130398393420271 steps: 256 optimizer_covers_current_parameters: true validation_curve: - step: 0 scores: rmse: 0.8142301306723146 baseline_rmse: 0.8056991213085883 nrmse: 1.0105883314727593 only_a_distance_nrmse: 0.721816357907311 - step: 32 scores: rmse: 0.5715751093486549 baseline_rmse: 0.8056991213085883 nrmse: 0.7094150834126796 only_a_distance_nrmse: 0.13606648887478617 - step: 128 scores: rmse: 0.5753770291476329 baseline_rmse: 0.8056991213085883 nrmse: 0.714133867011206 only_a_distance_nrmse: 0.1312310084402827 - step: 256 scores: rmse: 0.5649223332311453 baseline_rmse: 0.8056991213085883 nrmse: 0.7011579363691227 only_a_distance_nrmse: 0.0625918032746263 updated_parameter_count: 89 pooling_capacity: 2 removed_state_entries: - nodes.record/b.embedder.weights - nodes.record/b.embedder.max_fourier_input - nodes.record/b.embedder.embeddings.weight - nodes.record/b.embedder.counter.counts - nodes.record/b.embedder.linear.weight - nodes.record/b.embedder.linear.bias - nodes.record/b.embedder.scalar.weight - nodes.record/b.embedder.normalizer.mean - nodes.record/b.embedder.normalizer.var - nodes.record/b.embedder.normalizer.count changed_unselected_state_entries: [] inactive: before: rmse: 0.8051448159247964 baseline_rmse: 0.8352833232272557 nrmse: 0.963918222159621 only_a_distance_nrmse: 0.7052061517161763 after: rmse: 0.5778075411583316 baseline_rmse: 0.8352833232272557 nrmse: 0.691750361932136 only_a_distance_nrmse: 0.06130398393420271 steps: 256 optimizer_covers_current_parameters: true validation_curve: - step: 0 scores: rmse: 0.8142301306723146 baseline_rmse: 0.8056991213085883 nrmse: 1.0105883314727593 only_a_distance_nrmse: 0.721816357907311 - step: 32 scores: rmse: 0.5715751093486549 baseline_rmse: 0.8056991213085883 nrmse: 0.7094150834126796 only_a_distance_nrmse: 0.13606648887478617 - step: 128 scores: rmse: 0.5753770291476329 baseline_rmse: 0.8056991213085883 nrmse: 0.714133867011206 only_a_distance_nrmse: 0.1312310084402827 - step: 256 scores: rmse: 0.5649223332311453 baseline_rmse: 0.8056991213085883 nrmse: 0.7011579363691227 only_a_distance_nrmse: 0.0625918032746263 updated_parameter_count: 89 pooling_capacity: 2 scratch_restricted: before: rmse: 0.8868556317484789 baseline_rmse: 0.8352833232272557 nrmse: 1.06174229400626 only_a_distance_nrmse: 0.7709742570931406 after: rmse: 0.5782203378198156 baseline_rmse: 0.8352833232272557 nrmse: 0.69224456150491 only_a_distance_nrmse: 0.07019150170318522 steps: 256 optimizer_covers_current_parameters: true validation_curve: - step: 0 scores: rmse: 0.8572424072001292 baseline_rmse: 0.8056991213085883 nrmse: 1.0639733673878484 only_a_distance_nrmse: 0.7993693131078533 - step: 32 scores: rmse: 0.5694346192640745 baseline_rmse: 0.8056991213085883 nrmse: 0.7067583967811938 only_a_distance_nrmse: 0.0866815169412168 - step: 128 scores: rmse: 0.5711789456390196 baseline_rmse: 0.8056991213085883 nrmse: 0.7089233816109055 only_a_distance_nrmse: 0.10832556341325922 - step: 256 scores: rmse: 0.5654904967131181 baseline_rmse: 0.8056991213085883 nrmse: 0.7018631170835439 only_a_distance_nrmse: 0.07370084116310446 updated_parameter_count: 89 pooling_capacity: 2 continuation: before: rmse: 0.02008381101831335 baseline_rmse: 0.8352833232272557 nrmse: 0.02404430982856956 only_a_distance_nrmse: 0.7083488643135359 after: rmse: 0.01599579466581891 baseline_rmse: 0.8352833232272557 nrmse: 0.019150142497777287 only_a_distance_nrmse: 0.6791647199298226 steps: 256 optimizer_covers_current_parameters: true validation_curve: - step: 0 scores: rmse: 0.018716987765503217 baseline_rmse: 0.8056991213085883 nrmse: 0.023230741191704093 only_a_distance_nrmse: 0.7158475211177245 - step: 32 scores: rmse: 0.0270836781610786 baseline_rmse: 0.8056991213085883 nrmse: 0.033615126844237134 only_a_distance_nrmse: 0.7217665791058488 - step: 128 scores: rmse: 0.006284964283198546 baseline_rmse: 0.8056991213085883 nrmse: 0.007800634401823259 only_a_distance_nrmse: 0.7005309659747299 - step: 256 scores: rmse: 0.014907214874622854 baseline_rmse: 0.8056991213085883 nrmse: 0.01850221066446129 only_a_distance_nrmse: 0.6880186986846835 updated_parameter_count: 93 pooling_capacity: 3 readdition: scores: rmse: 0.5636660707421731 baseline_rmse: 0.8352833232272557 nrmse: 0.6748202137741188 only_a_distance_nrmse: 0.24856142796021527 source_normalization: mean: -0.019764553755521774 variance: 0.3272486925125122 std: 0.5720652869319307 count: 65536 alpha: null fresh_normalization: mean: 0.0 variance: 1.0 std: 1.0000049999875 count: 0 alpha: null source_field_order: - record/a - record/b - record/y readded_field_order: - record/a - record/y - record/b test_rows: 2048 optimizer_policy: New AdamW factory and Trainer per phase; matched adaptation data and updates mutation: delete record/b; inactive control updates active=False; separate source fork deletes and re-adds b checks: Source learns both-input relationship below 0.25 nRMSE: true Source optimizer covers current parameters: true Deletion removes b from schema and runtime: true Deletion removes the selected state entries: true Deletion preserves all unselected state entries: true Deletion immediately loses the selected information: true 'deleted before: removed input values cannot affect predictions': true 'deleted: optimizer covers current parameters': true 'deleted: parameters actually learn': true 'deleted: hidden target values cannot affect predictions': true 'deleted: adapts near the remaining-information limit': true 'deleted after: removed input values cannot affect predictions': true Deleted checkpoint preserves schema and learned state: true Deleted checkpoint preserves adapted predictions: true Inactive control preserves all original state: true Deletion and deactivation initially give the same predictions: true 'inactive before: removed input values cannot affect predictions': true 'inactive: optimizer covers current parameters': true 'inactive: parameters actually learn': true 'inactive: hidden target values cannot affect predictions': true 'inactive: adapts near the remaining-information limit': true 'inactive after: removed input values cannot affect predictions': true Scratch schema and field order match deletion: true 'scratch_restricted before: removed input values cannot affect predictions': true 'scratch_restricted: optimizer covers current parameters': true 'scratch_restricted: parameters actually learn': true 'scratch_restricted: hidden target values cannot affect predictions': true 'scratch_restricted: adapts near the remaining-information limit': true 'scratch_restricted after: removed input values cannot affect predictions': true Unchanged checkpoint restores source predictions: true 'continuation: optimizer covers current parameters': true 'continuation: parameters actually learn': true 'continuation: hidden target values cannot affect predictions': true Continuation retains full-information accuracy below 0.25 nRMSE: true Re-adding the same name does not restore its learned state: true Re-added input has fresh normalization: true outcome: met duration_seconds: 31.192