Auto-sync: 2026-06-28 01:00:39 (part 2)

results/README.md

@@ -5,6 +5,7 @@ Use these summaries for the current paper tables:
 - `paper_story_memo.md`
 - `paper_table_status.md` and `.json` (generated by `scripts/build_paper_table_status.py`)
 - `nonexpert_proposal_target_census.md`
+- `field_selected_target_map_summary.md`
 - `h16_field_sweep_summary.md`
 - `h16_lattice_summary.md`
 - `h16_lattice_no_expert_summary.md`
@@ -19,6 +20,10 @@ Use these summaries for the current paper tables:
 - `h16_policy_ckpt_near_miss_policy_fieldckpt_field_k32_sigma0p35_summary.md`
 - `h16_policy_ckpt_near_miss_policy_bc5_fieldckpt_field_k32_sigma0p35_summary.md`
 - `h16_policy_ckpt_near_miss_policy_bc5_bestpt_field_sweep_summary.md`
+- `h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_v2_summary.md`
+- `h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_knn4_v2_summary.md`
+- `h16_policy_ckpt_nonexpert_policy_bc5_summary.md`
+- `h16_policy_ckpt_nonexpert_policy_bc5_bestpt_field_sweep_summary.md`
 
 Known stale or incomplete files:
 
@@ -30,85 +35,42 @@ Known stale or incomplete files:
   absent/incomplete because the final seed was canceled after the stronger `best.pt`
   field-checkpoint variant completed.
 
-Completed sweep:
+Completed deployment-clean diagnostics:
 
 - Job `14837136` swept K/sigma for the current best deployment-clean route
   (`near_miss_policy_bc5` + `best.pt` + Gaussian field selection). Summary job
   `14837137` wrote `h16_policy_ckpt_near_miss_policy_bc5_bestpt_field_sweep_summary.*`.
-
-Pending field-optimization jobs:
-
-Last checked: `2026-06-28 04:51 UTC`.
-
-- Job `14842523` is a small runtime smoke test for `selection_mode=field_optim`
-  (`MAX_GROUPS=64`, `near_miss_policy_bc5` + `best.pt`); it is pending on
-  scheduler priority.
-- Jobs `14842533` and `14842543` were canceled after the queue could not place
-  them on the unavailable CPU nodes.
-- Job `14842557` is the current low-resource CPU unit smoke; it does not
-  instantiate ManiSkill, and only imports `_select_action_chunk` to check
-  `field_optim` action-space gradients. It is pending because available CPU
-  nodes are down, drained, or reserved.
-- Job `14842528` is a 4-config x 3-seed sweep of `field_optim` with
-  `FIELD_OPTIM_STEPS=4`, `FIELD_OPTIM_TRUST_RADIUS=0.5`, and
-  `FIELD_OPTIM_L2_PENALTY=0.02`. It depends on smoke job `14842523`.
-- Job `14842529` summarizes that sweep if `14842528` succeeds.
-- Job `14842551` is an `afterany` fallback summary for the same sweep, so partial
-  results are still collected if one array task fails.
-
-Pending non-expert proposal-policy jobs:
-
-- Job `14842574` trains `nonexpert_policy_bc5`, a policy target that imitates the
-  best non-expert candidate per state with BC weight 5.0. It is pending because
-  the requested GPU nodes are currently unavailable.
-- Job `14842575` evaluates direct rollout from `best_policy.pt`.
-- Job `14842616` summarizes direct rollout with tag
-  `policy_ckpt_nonexpert_policy_bc5`.
-- Job `14842577` sweeps Gaussian field selection around `best.pt`.
-- Job `14842617` summarizes that field sweep with tag
-  `policy_ckpt_nonexpert_policy_bc5_bestpt_field_sweep`.
-
-Pending counterfactual-residual retrieval jobs:
-
-- Job `14842596` smoke-tests `selection_mode=retrieval_residual` on
-  `near_miss_policy_bc5/best.pt` for 64 validation groups. It is pending
-  because the requested GPU nodes are currently unavailable.
-- Job `14842597` evaluates the 3-seed full validation rollout after the smoke
-  succeeds.
-- Job `14842618` summarizes the full rollout with tag
-  `policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual`.
-- Job `14842609` smoke-tests the KNN version with
-  `RETRIEVAL_NEIGHBORS=4`.
-- Job `14842610` evaluates the 3-seed KNN residual rollout after smoke.
-- Job `14842619` summarizes that rollout with tag
-  `policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_knn4`.
-
-The earlier summary jobs `14842576`, `14842578`, `14842598`, and `14842611`
-were canceled and replaced by the patched jobs above so summaries do not depend
-on the project `.venv` and include `retrieval_neighbors`.
-- Job `14842646` is a low-resource CPU unit smoke for the KNN residual helper.
-  It imports the helper in the Apptainer environment and checks that residuals
-  from multiple train states are attached while heldout candidates are ignored.
-  It completed successfully.
-- Jobs `14842596` and `14842609` were the first `retrieval_residual` rollout
-  smokes; they failed because the evaluator did not pass `retrieval_neighbors`
-  into `_attach_retrieved_residual_candidates`. The bug is fixed.
-- Job `14857111` reruns nearest-1 `retrieval_residual` smoke, `14857112` runs
-  full validation after that smoke, and `14857113` summarizes it.
-- Job `14857114` reruns KNN4 `retrieval_residual` smoke, `14857115` runs full
-  validation after that smoke, and `14857116` summarizes it.
-- Job `14857117` rebuilds `paper_table_status.*` after both residual summaries.
-  It completed successfully.
-- Jobs `14857692`/`14857693`/`14857694` evaluate the nearest-1
-  `retrieval_residual` diagnostic restricted to transferred `residual_near_miss`
-  proposals.
-- Jobs `14857695`/`14857696`/`14857697` evaluate the KNN4 version of the same
-  near-miss residual diagnostic.
-- Job `14857698` rebuilds `paper_table_status.*` after those near-miss residual
-  summaries.
-- Nearest-1 residual v2 completed at 32.12% success; KNN4 residual completed at
-  29.91%. The near-miss-only residual smokes reached only 14.06%, so the full
-  near-miss-only residual jobs were canceled.
+- Job `14842528` completed a partial trust-region `field_optim` sweep; the
+  after-any summary `14842551` shows a best observed result of 25.39%, so this is
+  a negative diagnostic rather than a clean bridge.
+- Jobs `14842574`/`14842575`/`14842616` completed the broad non-expert proposal
+  policy at 27.88%; jobs `14842577`/`14842617` completed its field sweep at
+  26.49%. Broad BC targets did not solve proposal generation.
+- Job `14842646` completed the KNN residual helper unit smoke.
+- Jobs `14857111`/`14857112`/`14857113` completed nearest-1 residual retrieval at
+  32.12%; jobs `14857114`/`14857115`/`14857116` completed KNN4 residual retrieval
+  at 29.91%. Nearest residual transfer is a positive bridge, while KNN dilutes it.
+- Jobs `14857692` and `14857695` completed near-miss-only residual smokes at
+  14.06%, so their full jobs were canceled.
+- Job `14858327` exported the train-split field-selected no-expert target map.
+  It selected `near_miss` for 1,820 of 2,298 train groups and non-near-miss
+  no-expert candidates for the rest.
+
+Active field-teacher distillation jobs:
+
+Last checked: `2026-06-28 05:03 UTC`.
+
+- Job `14858328` is running 3 seeds of `field_selected_noexpert_bc5` with the
+  train-split target map; jobs `14858329`/`14858330` and
+  `14858331`/`14858332` evaluate direct and field-guided rollouts afterward.
+- Job `14858333` rebuilds `paper_table_status.*` after the train-split
+  field-teacher summaries.
+- Job `14858449` completed the all-split target map export for aligned validation
+  checkpoint selection and seed-invariant student train coverage.
+- Job `14858450` trains 3 seeds of `field_selected_noexpert_bc5_allmap`; jobs
+  `14858451`/`14858452` and `14858453`/`14858454` evaluate direct and
+  field-guided rollouts afterward.
+- Job `14858455` rebuilds `paper_table_status.*` after the allmap summaries.
 
 Paper table automation:
 

results/dovla_cil_run_report_2026-06-27.md

@@ -20,6 +20,11 @@ The deployed h=16 behavior-cloning policy remains weak, but the interventional f
 | Lattice field selection, full lattice | 69.33% | +39.59 pp | Includes expert proposal in candidate set |
 | Retrieval lattice from nearest train state | 28.93% | -0.81 pp | Same-task action-library retrieval does not transfer |
 | Retrieval lattice, no expert | 27.13% | -2.61 pp | No same-state proposals, no gain |
+| Trust-region field optimization | 25.39% | -4.35 pp | Negative diagnostic; field is not a generic action optimizer |
+| Best non-expert proposal policy | 27.88% | -1.86 pp | Broad BC targets do not solve proposal generation |
+| Best non-expert proposal + field | 26.49% | -3.25 pp | Broad proposal-field bridge also fails |
+| Train-state residual retrieval | 32.12% | +2.38 pp | Positive clean bridge, below current clean best |
+| KNN train-state residual retrieval | 29.91% | +0.17 pp | More retrieved tangent neighborhoods dilute the signal |
 
 Canonical summaries:
 
@@ -37,13 +42,55 @@ Canonical summaries:
 - `results/h16_policy_ckpt_near_miss_policy_fieldckpt_field_k32_sigma0p35_summary.md`
 - `results/h16_policy_ckpt_near_miss_policy_bc5_fieldckpt_field_k32_sigma0p35_summary.md`
 - `results/h16_policy_ckpt_near_miss_policy_bc5_bestpt_field_sweep_summary.md`
+- `results/h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_v2_summary.md`
+- `results/h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_knn4_v2_summary.md`
+- `results/h16_policy_ckpt_nonexpert_policy_bc5_summary.md`
+- `results/h16_policy_ckpt_nonexpert_policy_bc5_bestpt_field_sweep_summary.md`
+- `results/h16_field_optim_near_miss_policy_bc5_bestpt_s4_trust05_afterany_summary.md`
+- `results/paper_table_status.md`
+- `results/field_selected_target_map_summary.md`
 
 Non-canonical/stale summaries:
 
 - `results/h16_lattice_no_near_miss_no_expert_summary.md` was produced before fixing Slurm comma handling for multi-type exclusion. Use the `_v2` file instead.
 - `results/h16_policy_ckpt_near_miss_policy_bc5_field_k32_sigma0p35_summary.md` is absent/incomplete because the last seed was canceled after the `best.pt` field-checkpoint variant superseded it.
 
-## Jobs Submitted
+## Current Update - 2026-06-28 05:03 UTC
+
+The old optimistic h=16 policy story is no longer the paper story. Direct BC
+rollout stayed near 30%, while the same-state no-expert lattice reached 56.99%.
+The strongest current paper claim is therefore a mechanism claim: the learned
+field works on local counterfactual proposal geometry, not on arbitrary
+off-manifold action search.
+
+Completed since the first report:
+
+- `field_optim` finished as a negative diagnostic: best observed 25.39%.
+- `nonexpert_policy_bc5` finished as a negative proposal-model ablation:
+  direct 27.88%, field-guided 26.49%.
+- `retrieval_residual` finished as a positive clean bridge: nearest-1 reached
+  32.12%, while KNN4 dropped to 29.91%.
+- Near-miss-only residual smokes reached only 14.06%, so their full jobs were
+  canceled.
+- Field-teacher target exports now exist:
+  - train split: 2,298 groups, 1,820 `near_miss`, 478 non-near-miss no-expert targets;
+  - all split: 2,873 groups, 2,273 `near_miss`, 600 non-near-miss no-expert targets.
+
+Active field-teacher jobs:
+
+- `14858328` is running train-split `field_selected_noexpert_bc5` training;
+  `14858329`/`14858330` and `14858331`/`14858332` evaluate direct and
+  field-guided rollouts.
+- `14858450` is pending all-split `field_selected_noexpert_bc5_allmap`
+  training; `14858451`/`14858452` and `14858453`/`14858454` evaluate direct and
+  field-guided rollouts.
+- `14858333` and `14858455` rebuild `paper_table_status.*` after the summaries.
+
+Use `results/paper_table_status.md` as the canonical current table. The job
+table below is historical and includes statuses that were true at earlier
+checkpoints.
+
+## Historical Jobs Submitted
 
 | Job | Purpose | Status |
 |---|---|---|
@@ -110,8 +157,15 @@ Non-canonical/stale summaries:
 | 14857696 | Evaluate KNN4 transferred near-miss residual retrieval | Canceled after weak smoke |
 | 14857697 | Summarize KNN4 transferred near-miss residual retrieval | Canceled |
 | 14857698 | Rebuild paper table after near-miss residual summaries | Canceled |
+| 14858327 | Export field-selected no-expert policy target map | Pending |
+| 14858328 | Train `field_selected_noexpert_bc5` policies | Pending |
+| 14858329 | Evaluate direct `field_selected_noexpert_bc5` policies | Pending |
+| 14858330 | Summarize direct `field_selected_noexpert_bc5` policies | Pending |
+| 14858331 | Evaluate `field_selected_noexpert_bc5` with field-guided proposals | Pending |
+| 14858332 | Summarize `field_selected_noexpert_bc5` field-guided proposals | Pending |
+| 14858333 | Rebuild paper table after field-teacher distillation | Pending |
 
-Last queue check: `2026-06-28 04:51 UTC`. The earlier queue backlog cleared.
+Last queue check: `2026-06-28 04:56 UTC`. The earlier queue backlog cleared.
 `field_optim`, `nonexpert_policy_bc5`, and their summaries completed; fixed
 `retrieval_residual` v2 jobs are now pending on priority/dependencies.
 
@@ -155,6 +209,12 @@ close to but below the clean best 32.93%. KNN4 residual retrieval completed at
 signal. The near-miss-only residual smoke was only 14.06%, so the full
 near-miss-only residual jobs were canceled.
 
+The next proposal-generator attempt is field-teacher distillation. Instead of
+imitating reward-best non-expert candidates, job `14858327` exports train-split
+targets chosen by the learned field on no-expert local lattices. Jobs
+`14858328`-`14858333` then train/evaluate a policy student of that local
+counterfactual decision rule.
+
 Paper table automation now lives in `scripts/build_paper_table_status.py` with a
 CPU wrapper at `scripts/slurm/build_paper_table_status.sbatch`. The table builder
 is intentionally standard-library-only: it consumes result summary JSON files,
@@ -203,8 +263,11 @@ as the main result.
 
 ## Next Best Experiments
 
-1. Generate an on-the-fly non-expert proposal lattice that does not depend on stored validation candidates, then evaluate with `selection_mode=lattice`.
-2. Finish `field_optim` jobs `14842523`/`14842557`/`14842528` to test whether differentiable trust-region field ascent beats Gaussian proposal selection.
-3. Monitor `nonexpert_policy_bc5` jobs `14842574`-`14842578`, which train and evaluate a proposal model for best non-expert local interventions.
-4. Monitor `retrieval_residual` jobs `14842596`-`14842598` and KNN4 jobs `14842609`-`14842611`; these test transferable counterfactual tangent directions without validation-lattice leakage.
-4. Add final paper tables separating `policy`, `proposal_field`, `field_lattice_no_expert`, `field_lattice_full`, and `oracle`.
+1. Finish train-split and all-split field-teacher distillation jobs
+   (`14858328`-`14858333`, `14858450`-`14858455`).
+2. Promote allmap field-teacher distillation only if it beats the current clean
+   best 32.93%; otherwise keep residual retrieval as the strongest clean bridge.
+3. Keep the main paper table separated into direct policy, clean proposal+field,
+   same-state no-expert lattice, same-state full lattice, and oracle ceiling.
+4. Run or cite external baselines only after matching protocol; do not claim SOTA
+   from the internal table alone.

results/field_selected_target_map_summary.md

@@ -0,0 +1,46 @@
+# Field-Selected Policy Target Map Summary
+
+Train-split target map:
+`/scratch/knguy52/dovla/experiments/field_selected_targets/near_miss_policy_bc5_seed0_bestpt_noexpert_train.json`
+
+All-split target map:
+`/scratch/knguy52/dovla/experiments/field_selected_targets/near_miss_policy_bc5_seed0_bestpt_noexpert_all.json`
+
+Teacher checkpoint:
+`/scratch/knguy52/dovla/experiments/dovla_h16_policy_ckpt_runs/near_miss_policy_bc5/seed_0/best.pt`
+
+Train split: `train`
+
+All split: `all`
+
+Excluded proposal types: `expert`
+
+Total train groups: 2,298
+
+Total all-split groups: 2,873
+
+## Train-Split Counts
+
+| selected candidate type | groups |
+|---|---:|
+| near_miss | 1,820 |
+| no_op | 227 |
+| random_negative | 115 |
+| wrong_gripper | 103 |
+| wrong_direction | 33 |
+
+## All-Split Counts
+
+| selected candidate type | groups |
+|---|---:|
+| near_miss | 2,273 |
+| no_op | 275 |
+| random_negative | 146 |
+| wrong_gripper | 135 |
+| wrong_direction | 44 |
+
+Interpretation: the learned field's no-expert lattice teacher mostly selects
+`near_miss` actions, but not exclusively. The train-split map is useful as a
+strict seed-0 train target export. The all-split map gives every student seed a
+field-selected target for its own train and validation groups, avoiding fallback
+to fixed `near_miss` targets when split seeds differ.