Auto-sync: 2026-06-27 15:04:36 (part 2)

results/README.md

@@ -2,6 +2,9 @@
 
 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`
 - `h16_field_sweep_summary.md`
 - `h16_lattice_summary.md`
 - `h16_lattice_no_expert_summary.md`
@@ -32,3 +35,66 @@ Completed sweep:
 - 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: `2026-06-27 21:12 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 is pending on unavailable CPU nodes.
+
+Paper table automation:
+
+- `scripts/build_paper_table_status.py` builds a paper-facing method table from
+  existing summary JSON files, uses canonical fallback values for already
+  verified rows, and marks pending experiments by job ID.
+- `scripts/slurm/build_paper_table_status.sbatch` reruns that table builder on a
+  CPU node with system `python3`.

results/dovla_cil_run_report_2026-06-27.md

@@ -73,6 +73,71 @@ Non-canonical/stale summaries:
 | 14836242 | Summarize near-miss BC x5 `best.pt` field selection | Completed |
 | 14837136 | Sweep near-miss BC x5 `best.pt` field selection over K/sigma | Completed |
 | 14837137 | Summarize near-miss BC x5 `best.pt` field sweep | Completed |
+| 14842523 | Smoke-test `field_optim` deployment-clean action optimization | Pending |
+| 14842533 | CPU smoke-test `field_optim` while GPU nodes are unavailable | Canceled |
+| 14842543 | CPU unit smoke-test for `field_optim` helper without ManiSkill | Canceled |
+| 14842557 | Low-resource CPU unit smoke-test for `field_optim` helper without ManiSkill | Pending |
+| 14842528 | Sweep `field_optim` over K/sigma after smoke succeeds | Pending |
+| 14842529 | Summarize `field_optim` sweep | Pending |
+| 14842551 | Fallback `afterany` summary for partial `field_optim` sweep results | Pending |
+| 14842574 | Train `nonexpert_policy_bc5` on best non-expert local interventions | Pending |
+| 14842575 | Evaluate `nonexpert_policy_bc5` direct `best_policy.pt` rollout | Pending |
+| 14842576 | Old summary for `nonexpert_policy_bc5` direct rollout | Canceled; replaced by `14842616` |
+| 14842577 | Sweep `nonexpert_policy_bc5` `best.pt` with Gaussian field selection | Pending |
+| 14842578 | Old summary for `nonexpert_policy_bc5` field sweep | Canceled; replaced by `14842617` |
+| 14842596 | Smoke-test train-split counterfactual residual retrieval around policy mean | Pending |
+| 14842597 | Evaluate full `retrieval_residual` rollout after smoke succeeds | Pending |
+| 14842598 | Old summary for `retrieval_residual` rollout | Canceled; replaced by `14842618` |
+| 14842609 | Smoke-test KNN train-split counterfactual residual retrieval (`RETRIEVAL_NEIGHBORS=4`) | Pending |
+| 14842610 | Evaluate full KNN `retrieval_residual` rollout after smoke succeeds | Pending |
+| 14842611 | Old summary for KNN `retrieval_residual` rollout | Canceled; replaced by `14842619` |
+| 14842616 | Patched summary for `nonexpert_policy_bc5` direct rollout | Pending |
+| 14842617 | Patched summary for `nonexpert_policy_bc5` field sweep | Pending |
+| 14842618 | Patched summary for `retrieval_residual` rollout | Pending |
+| 14842619 | Patched summary for KNN `retrieval_residual` rollout | Pending |
+| 14842646 | CPU unit smoke for KNN `retrieval_residual` helper | Pending |
+
+Last queue check: `2026-06-27 21:12 UTC`. GPU smoke `14842523` is pending on
+scheduler priority. CPU unit smoke `14842557` is pending because matching CPU
+nodes are down, drained, or reserved. No `field_optim` result file has been
+produced yet.
+
+The new `nonexpert_policy_bc5` branch is motivated by
+`results/nonexpert_proposal_target_census.md`: the best non-expert local action
+is not a `near_miss` in 888 of 2,873 states. This directly tests whether the
+proposal bottleneck can be reduced by imitating the whole successful non-expert
+intervention family rather than only near-miss proposals. Its training array
+`14842574` is pending because the requested GPU nodes are currently unavailable.
+
+The new `retrieval_residual` evaluator is a second deployment-clean proposal
+bridge. It retrieves counterfactual residuals from the nearest train-split state
+(`candidate_action - expert_action`), translates those residuals around the
+current policy mean, and lets the field select among the translated local
+proposal lattice. This directly tests whether the failed absolute-action
+retrieval result was caused by missing locality rather than by retrieval itself.
+The `retrieval_neighbors` parameter extends the same idea to K nearest
+train-state tangent neighborhoods; `14842609`-`14842611` evaluate K=4.
+The edited Python files passed a text-only `ast.parse` syntax check; local
+`py_compile` was avoided afterward because `.pyc` writes were hanging on the
+login filesystem.
+Summary Slurm scripts now default to system `python3` instead of the project
+`.venv`, because the summary jobs only need the standard library and `.venv`
+processes were observed stuck in login-node I/O wait.
+The KNN residual helper has a new unit-style test in
+`tests/test_maniskill_policy_rollout.py`. It passed Python `ast.parse`; direct
+local execution is deferred to the container/Slurm environment because bare
+`python3` on the login node lacks `numpy`.
+Job `14842646` is the matching Apptainer unit smoke for that helper. Slurm
+records `ExcNodeList=rc32303`, but the current pending reason still names
+`rc32303` as unavailable, so this appears to be the same cluster scheduling
+issue affecting the other CPU smoke jobs rather than a script failure.
+
+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,
+uses canonical fallback values for already verified rows, and labels pending
+experiments by job ID so the A*/paper story can be updated without hand-copying
+numbers after Slurm jobs complete.
 
 ## Interpretation
 
@@ -113,5 +178,7 @@ 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. Train a proposal model to imitate successful non-expert lattice candidates so deployment does not require an explicit lattice store.
-3. Add final paper tables separating `policy`, `proposal_field`, `field_lattice_no_expert`, `field_lattice_full`, and `oracle`.
+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`.