Update daily report with per-dim rollout evidence

workspace/reports/2026-07-03.md

@@ -134,7 +134,7 @@
   - first attempt `15130582`, `15130583`, `15130584`: train/validation/test arrays, seeds 0/1/2, `CLIP_ACTIONS=0`, identity transform, per-dimension train-only scale vector from `runs/action_scale_vector_train_base_branch_max`.
   - first attempt exposed a 7D vector vs adapted env action-dimension mismatch; fixed in `scripts/eval_ctt_generated_rollout.py`, canceled remaining tasks, and removed incomplete run dirs.
   - retry `15130732`, `15130733`, `15130734`: train/validation/test arrays, seeds 0/1/2, same convention. Canceled after validation seeds revealed Slurm comma splitting had truncated the vector to one dimension.
-  - clean retry using `EXECUTION_ACTION_SCALE_VECTOR_FILE` `15130968`, `15130969`, `15130970`: train/validation/test arrays, seeds 0/1/2, same convention. Current state at report update: submitted; waiting for scheduler/run completion.
+  - clean retry using `EXECUTION_ACTION_SCALE_VECTOR_FILE` `15130968`, `15130969`, `15130970`: train/validation/test arrays, seeds 0/1/2, same convention. Validation/test completed for all seeds; train seeds 1/2 completed, while train seed 0 and retry `15131363_0` were canceled after validation/test already showed the convention fails proposal support and generated action-bound validity.
 
 ## Jobs Completed
 
@@ -426,6 +426,14 @@
   - `runs/ctt_base_context_obs_nonlinear_dominance_context_tangent_positive_train_to_test`: selected success `0.2986`, coverage `0.4028`, selected model `hgb_classifier`, target `positive_margin`
   - `runs/ctt_base_context_obs_nonlinear_dominance_context_success_train_to_test`: selected success `0.2569`, coverage `0.3750`, selected model `hgb_regressor`, target `success`
   - interpretation: nonlinear selectors fit the calibration distribution better than ridge in some splits, but transfer worse than the best train-calibrated ridge row (`0.3125`). This is negative evidence that the current features are the bottleneck; adding a more flexible selector does not solve deployment-clean dominance.
+- Per-dimension train-max action-convention rollouts completed enough to reject the convention:
+  - scale artifact: `runs/action_scale_vector_train_base_branch_max`, fit from train-only `base_branch` per-dimension max-abs action bounds.
+  - validation seed artifacts: `runs/ctt_residual_base_context_obs_rollout_val69_perdim_trainmax_seed{0,1,2}`; aggregate `runs/ctt_base_context_obs_val_perdim_trainmax_rollout_comparison`.
+  - test seed artifacts: `runs/ctt_residual_base_context_obs_rollout_test_perdim_trainmax_seed{0,1,2}`; aggregate `runs/ctt_base_context_obs_test_perdim_trainmax_rollout_comparison`.
+  - train seed artifacts: `runs/ctt_residual_base_context_obs_rollout_train_cal_perdim_trainmax_seed{1,2}`; train seed 0 and retry `15131363_0` were canceled after the held-out result was already negative.
+  - validation, 207 rows: base success `0.2754`, selected success `0.0918`, proposal oracle `0.1691`, hidden oracle `0.6667`, OutcomePTR@8 `0.3043`, success support gap `0.4976`, success selector gap `0.0773`, generated action-bound violation rate `0.2742`, base action-bound violation rate `0.0000`.
+  - test, 144 rows: base success `0.3333`, selected success `0.1181`, proposal oracle `0.2222`, hidden oracle `0.7292`, OutcomePTR@8 `0.3611`, success support gap `0.5139`, success selector gap `0.1042`, generated action-bound violation rate `0.3030`, base action-bound violation rate `0.0000`.
+  - interpretation: per-dimension train-max scaling fixes base action-bound validity but makes proposal support much worse than bounded tanh and still leaves generated action-bound violations. It is a clean negative action-convention diagnostic; selector/dominance evaluation is not warranted until proposal support and generated validity improve.
 
 ## Failed Jobs And Log Excerpts
 
@@ -475,12 +483,14 @@ Interpretation: residual CTT beats local-atlas on mean positive distance while s
 | validation + base-context-observation no-clipping | 3 | 207 | 0.2899 | 0.2271 | 0.4106 | 0.6667 | 0.4541 | 0.2754 | 0.1836 | fail: action_bounds/selected/support/selector |
 | validation + base-context-observation scaled0.215 | 3 | 207 | 0.1594 | 0.0676 | 0.1208 | 0.6667 | 0.2754 | 0.5459 | 0.0531 | fail: scale/support/selected |
 | validation + base-context-observation tanh | 3 | 207 | 0.2754 | 0.2126 | 0.3671 | 0.6667 | 0.4251 | 0.3092 | 0.1546 | fail: selected/support/selector |
+| validation + base-context-observation per-dim trainmax | 3 | 207 | 0.2754 | 0.0918 | 0.1691 | 0.6667 | 0.3043 | 0.4976 | 0.0773 | fail: support/generated_bounds/selected |
 | train calibration + base-context-observation tanh | 3 | 432 | 0.3056 | 0.2245 | 0.3634 | 0.6597 | 0.3773 | 0.3333 | 0.1389 | calibration only: selected/support/selector |
 | test | 3 | 144 | 0.2847 | 0.2222 | 0.5069 | 0.7292 | 0.5278 | 0.2639 | 0.2847 | fail: selected/support/selector |
 | test + base-context-observation | 3 | 144 | 0.2917 | 0.2708 | 0.5139 | 0.7292 | 0.5347 | 0.2639 | 0.2431 | fail: selected/support/selector |
 | test + base-context-observation no-clipping | 3 | 144 | 0.2500 | 0.2500 | 0.5000 | 0.7292 | 0.5417 | 0.2708 | 0.2500 | fail: action_bounds/support/selector |
 | test + base-context-observation scaled0.215 | 3 | 144 | 0.2292 | 0.0903 | 0.1667 | 0.7292 | 0.3750 | 0.5625 | 0.0764 | fail: scale/support/selected |
 | test + base-context-observation tanh | 3 | 144 | 0.3333 | 0.2500 | 0.3819 | 0.7292 | 0.4167 | 0.3681 | 0.1319 | fail: selected/support/selector |
+| test + base-context-observation per-dim trainmax | 3 | 144 | 0.3333 | 0.1181 | 0.2222 | 0.7292 | 0.3611 | 0.5139 | 0.1042 | fail: support/generated_bounds/selected |
 | test + dominance auto | 3 | 144 | 0.2847 | 0.2500 | 0.5069 | 0.7292 | 0.5278 | 0.2639 | 0.2917 | fail: selected/support/selector |
 | test + dominance tau0 | 3 | 144 | 0.2847 | 0.2778 | 0.5069 | 0.7292 | 0.5278 | 0.2639 | 0.2778 | fail: selected/support/selector |
 | test + base-context-observation tau0 | 3 | 144 | 0.2917 | 0.3056 | 0.5139 | 0.7292 | 0.5347 | 0.2639 | 0.2639 | fail: selected/support/selector |
@@ -497,11 +507,11 @@ Interpretation: residual CTT beats local-atlas on mean positive distance while s
 | test + base-context-observation tanh learned context-tangent success dominance | 3 | 144 | 0.3333 | 0.3542 | 0.3819 | 0.7292 | 0.4167 | 0.3681 | 0.1528 | fail: support/selector |
 | test + base-context-observation tanh train learned context-success dominance | 3 | 144 | 0.3333 | 0.3819 | 0.3819 | 0.7292 | 0.4167 | 0.3681 | 0.1250 | fail: support/selector |
 
-Interpretation: the test proposal oracle exceeds 50% support for the clipped visual-stat row, and visual-stat CTT plus validation-calibrated learned context dominance gives the best clipped-convention held-out selected success (`0.3264`). The cleaner clipped-convention train-calibrated selector reaches `0.3125`, which still improves over base but does not beat the validation-calibrated diagnostic. Nonlinear train-calibrated selectors do not improve this number. The object-layout proxy pass does not survive measured validation rollout and is worse than the RGB-stat row on proposal oracle/selected success. No-clipping preserves measurable support but exposes widespread action-bound violations; global scaled replay fixes base bound validity but collapses proposal support. The bounded tanh convention eliminates action-bound violations on validation/test. With train-clean context-success dominance, bounded tanh reaches `0.3819` selected success on held-out test and matches the measured proposal oracle in micro success, but the row-level selector gap remains `0.1250`, its proposal oracle is also only `0.3819`, and the support gap to the hidden chart oracle remains `0.3681`. The current best bounded-action story is therefore a useful selector diagnostic and a transport/support failure. The next paper-critical work is richer learned visual-language/object-centric chart representation and a decoder/action convention that preserves proposal support, not another deterministic proxy descriptor.
+Interpretation: the test proposal oracle exceeds 50% support for the clipped visual-stat row, and visual-stat CTT plus validation-calibrated learned context dominance gives the best clipped-convention held-out selected success (`0.3264`). The cleaner clipped-convention train-calibrated selector reaches `0.3125`, which still improves over base but does not beat the validation-calibrated diagnostic. Nonlinear train-calibrated selectors do not improve this number. The object-layout proxy pass does not survive measured validation rollout and is worse than the RGB-stat row on proposal oracle/selected success. No-clipping preserves measurable support but exposes widespread action-bound violations; global scaled replay fixes base bound validity but collapses proposal support. The bounded tanh convention eliminates action-bound violations on validation/test, while per-dimension train-max replay fixes base validity but further collapses proposal support and still leaves generated action-bound violations. With train-clean context-success dominance, bounded tanh reaches `0.3819` selected success on held-out test and matches the measured proposal oracle in micro success, but the row-level selector gap remains `0.1250`, its proposal oracle is also only `0.3819`, and the support gap to the hidden chart oracle remains `0.3681`. The current best bounded-action story is therefore a useful selector diagnostic and a transport/support failure. The next paper-critical work is richer learned visual-language/object-centric chart representation and a decoder/action convention that preserves proposal support, not another deterministic proxy descriptor.
 
 Safety metric status: regenerated measured rollout reports now expose `base_safety_label_known`, `generated_safety_label_coverage_at_8`, and `selected_safety_label_known_at_8`. For the visual-stat, object-layout, held-out test, and train-calibration rollout rows generated before the action-bound hook, each coverage number is `0.0000`; the legacy validation aggregate without `candidate_outcomes` reports `n/a`. For action-bound smoke/no-clipping/scaled artifacts, coverage is `1.0000` because `safety_violation_source=action_bounds`. These are action-space validity labels only, not collision/contact labels.
 
-Action-bound audit status: new rollout artifacts now log action-bound safety labels. The first smoke has 100% known labels and 100% action-bound violation for base/generated candidates. The full chart DB audit shows base-action violation rate 1.0000 in train/val/test and branch-level violation rate about 0.308. The no-clipping refresh has 100% known labels and about 0.98 generated action-bound violation rate. The scaled0.215 refresh has 100% known labels and base action-bound violation rate `0.0000`, but generated violation rate remains `0.1920` on validation and `0.1997` on test while proposal support collapses. The tanh bounded-decoder refresh has 100% known labels and zero base/generated/selected action-bound violations on full validation/test, but support and selector gaps remain large. Collision/contact safety remains unmeasured.
+Action-bound audit status: new rollout artifacts now log action-bound safety labels. The first smoke has 100% known labels and 100% action-bound violation for base/generated candidates. The full chart DB audit shows base-action violation rate 1.0000 in train/val/test and branch-level violation rate about 0.308. The no-clipping refresh has 100% known labels and about 0.98 generated action-bound violation rate. The scaled0.215 refresh has 100% known labels and base action-bound violation rate `0.0000`, but generated violation rate remains `0.1920` on validation and `0.1997` on test while proposal support collapses. The tanh bounded-decoder refresh has 100% known labels and zero base/generated/selected action-bound violations on full validation/test, but support and selector gaps remain large. The per-dimension train-max refresh also has 100% known labels and base action-bound violation rate `0.0000`, but generated violation rate remains `0.2742` on validation and `0.3030` on test while support collapses further than tanh. Collision/contact safety remains unmeasured.
 
 ## Leakage Audit Status
 
@@ -561,6 +571,7 @@ Action-bound audit status: new rollout artifacts now log action-bound safety lab
 - Added scaled0.215 validation/test aggregate tables via `\input{../runs/ctt_base_context_obs_val_scaled0215_rollout_comparison/table}` and `\input{../runs/ctt_base_context_obs_test_scaled0215_rollout_comparison/table}`. The paper treats global scaling as a negative diagnostic: it fixes base action-bound validity but collapses proposal support and leaves nonzero generated action-bound violations.
 - Added bounded-decoder tanh support to the rollout artifact path, plus full validation/test tanh aggregate tables and bounded-tanh dominance diagnostics. The draft treats tanh as an action-bound-clean diagnostic convention with lower support than clipped/no-clipping visual-stat rows, not as a final deployment method.
 - Added the train-calibrated bounded-tanh dominance table via `\input{../runs/ctt_base_context_obs_learned_dominance_context_success_tanh_train_to_test/table}` and updated the text: train-clean bounded dominance matches the proposal oracle in micro selected success, but the row-level selector gap remains nonzero and support remains too weak.
+- Added per-dimension train-max validation/test aggregate tables via `\input{../runs/ctt_base_context_obs_val_perdim_trainmax_rollout_comparison/table}` and `\input{../runs/ctt_base_context_obs_test_perdim_trainmax_rollout_comparison/table}`. The draft treats per-dimension train-max scaling as a negative diagnostic: it fixes base action-bound validity, but proposal support collapses and generated action-bound violations remain.
 - Verification after the bounded-tanh/dominance paper patch:
   - `.venv/bin/python -m py_compile scripts/eval_ctt_generated_rollout.py tests/test_ctt.py scripts/eval_dominance_selector.py scripts/eval_learned_dominance_selector.py`
   - `.venv/bin/python -m pytest tests/test_ctt.py tests/test_metrics.py tests/test_cil_metrics.py tests/test_dominance_selector.py tests/test_chart_features.py -q`: 36 passed.
@@ -572,6 +583,9 @@ Action-bound audit status: new rollout artifacts now log action-bound safety lab
   - `bash -n scripts/slurm/eval_ctt_generated_rollout.sbatch`
   - `bash -n scripts/slurm/eval_perdim_trainmax_dominance.sbatch`
   - HF forced-download SHA verification matched local files for the new scale-vector builder, rollout evaluator, Slurm wrappers, `tests/test_ctt.py`, daily report, and `runs/action_scale_vector_train_base_branch_max/{vector.json,vector_env.txt}`.
+- Verification after adding final per-dimension train-max rollout evidence to the paper/report:
+  - `latexmk -pdf -interaction=nonstopmode main.tex`: PDF rebuilt successfully at 27 pages; only existing overfull/underfull layout warnings remain.
+  - Overclaim scan for selector-gap closure, submission-tier, leaderboard-style, and internal acceptance phrasing in `latex/main.tex`, `reports/2026-07-03.md`, `reports/draft_audit_ctt.md`, and `docs/cil_format.md` returned no matches.
 
 ## Hugging Face Sync Status
 
@@ -614,7 +628,7 @@ Action-bound audit status: new rollout artifacts now log action-bound safety lab
 1. Fix the action representation/decoder, not with a global scalar: scaled0.215 fixes base bounds but collapses generated support.
 2. Stop treating deterministic object-layout as a likely win: it passed proxy but failed measured support relative to RGB-stat.
 3. Replace deterministic RGB/object descriptors with learned visual-language or task/object tokens that distinguish target, distractor, gripper/contact region, and task stage.
-4. Improve bounded-action support/transport: train-clean tanh dominance reaches `0.3819`, matching the tanh proposal oracle in micro success, so the next bottleneck is proposal support under a valid action convention. Per-dimension train-max scaled identity rollouts are now queued as the next support-preservation diagnostic.
+4. Improve bounded-action support/transport: train-clean tanh dominance reaches `0.3819`, matching the tanh proposal oracle in micro success, so the next bottleneck is proposal support under a valid action convention. Per-dimension train-max scaled identity rollouts are now negative evidence, not the path forward.
 5. Train utility/dominance on richer outcome-vector features only after support improves; under bounded tanh, current train-clean context-success dominance matches proposal-oracle micro success but leaves a nonzero row-level selector gap.
 6. Rerun measured support/selection once under exactly one declared deployment action convention after the decoder/action representation fix; tanh is action-bound-clean but currently sacrifices too much proposal support.
 7. Add actual collision/contact labels after action-bound validity is resolved.