Auto-sync: 2026-06-28 00:54:07 (part 3)

results/h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_v2_summary.md

@@ -0,0 +1,19 @@
+# h=16 Best-Policy Checkpoint Rollout
+
+Run root: `/scratch/knguy52/dovla/experiments/dovla_h16_policy_ckpt_runs`
+Objective: `near_miss_policy_bc5`
+Result file: `retrieval_residual_v2_rollout.json`
+Completed seeds: 3
+Baseline h=4 policy success: 29.67%
+Baseline h=16 rank-checkpoint success: 29.74%
+
+Mean success: 32.12% +/- 1.26%
+Gain vs h=16 rank checkpoint: +2.38%
+Mean progress: 54.83%
+Mean action MSE to best: 0.559
+
+| seed | mode | k | retrieval K | sigma | opt steps | trust | success | progress | oracle | action MSE |
+|---:|---|---:|---:|---:|---:|---:|---:|---:|---:|---:|
+| 0 | retrieval_residual | 16 | 1 | 0.00 | 0 | 0.00 | 31.48% | 53.24% | 85.74% | 0.633 |
+| 1 | retrieval_residual | 16 | 1 | 0.00 | 0 | 0.00 | 31.30% | 54.83% | 86.96% | 0.508 |
+| 2 | retrieval_residual | 16 | 1 | 0.00 | 0 | 0.00 | 33.57% | 56.41% | 87.65% | 0.538 |

results/paper_core_results.md

@@ -18,7 +18,9 @@ baseline is the h=16 rank-checkpoint online rollout (`29.74%`).
 | Trust-region field optimization | No | No | 25.39% | -4.35 pp | Differentiable field ascent is a negative diagnostic; the field is not a generic action optimizer |
 | Best non-expert proposal policy | No | No | 27.88% | -1.86 pp | Broadening BC targets beyond near-miss does not solve proposal generation |
 | Best non-expert proposal + field | No | No | 26.49% | -3.25 pp | The field still needs local counterfactual proposal geometry |
-| Train-state residual retrieval | No | No | Pending | Pending | Fixed v2 jobs are running; tests whether transferable counterfactual tangent directions bridge the clean gap |
+| Train-state residual retrieval | No | No | 32.12% | +2.38 pp | Transferred counterfactual residuals are a positive clean bridge but do not beat the near-miss proposal policy |
+| KNN train-state residual retrieval | No | No | 29.91% | +0.17 pp | Adding more retrieved tangent neighborhoods dilutes the signal |
+| Train-state near-miss residual retrieval | No | No | 14.06% smoke | -15.68 pp | Restricting to transferred near-miss residuals failed in smoke; full jobs canceled |
 | Lattice, no expert/no near-miss | Yes | No | 25.57% | -4.17 pp | Non-local negatives do not help |
 | Lattice, near-miss only | Yes | No | 55.94% | +26.20 pp | Local counterfactual proposals carry the gain |
 | Lattice, no expert | Yes | No | 56.99% | +27.25 pp | Reviewer-safe main result |
@@ -33,15 +35,17 @@ Suggested main-table rows:
 4. Near-miss proposal + field, BC x5 field checkpoint
 5. Trust-region field optimization
 6. Best non-expert proposal + field
-7. Lattice, near-miss only
-8. Lattice, no expert
-9. Lattice, full
-10. Oracle ceiling
+7. Train-state residual retrieval
+8. Lattice, near-miss only
+9. Lattice, no expert
+10. Lattice, full
+11. Oracle ceiling
 
 Suggested claim:
 
 > DoVLA-CIL is not a better behavior-cloning policy; it is a local counterfactual action
 > selection rule. A deployment-clean near-miss proposal policy plus the field gives a small
-> gain, while field-gradient ascent and broader non-expert BC targets both fail. The large
-> effect appears only when the field is queried on same-state intervention proposals, and
-> the mechanism is isolated to near-miss counterfactuals.
+> gain, and transferred counterfactual residuals nearly match it, while field-gradient ascent,
+> KNN residual retrieval, and broader non-expert BC targets fail. The large effect appears only
+> when the field is queried on same-state intervention proposals, and the mechanism is isolated
+> to near-miss counterfactuals.

results/paper_story_memo.md

@@ -62,7 +62,7 @@ test-time search. The cleaner novelty is:
 
 ## Active Jobs
 
-Last checked: `2026-06-28 04:44 UTC`.
+Last checked: `2026-06-28 04:51 UTC`.
 
 - `14842523`: GPU smoke for `selection_mode=field_optim`.
 - `14842557`: low-resource CPU unit smoke for the pure action-optimization helper.
@@ -83,26 +83,24 @@ Last checked: `2026-06-28 04:44 UTC`.
 - `14842619`: KNN4 `retrieval_residual` summary.
 - `14842646`: CPU unit smoke for the KNN residual helper. Completed.
 - `14857111`: fixed nearest-1 `retrieval_residual` smoke.
-- `14857112`: fixed nearest-1 `retrieval_residual` full rollout.
-- `14857113`: fixed nearest-1 `retrieval_residual` summary.
-- `14857114`: fixed KNN4 `retrieval_residual` smoke.
-- `14857115`: fixed KNN4 `retrieval_residual` full rollout.
-- `14857116`: fixed KNN4 `retrieval_residual` summary.
-- `14857117`: rebuild `paper_table_status.*` after fixed residual summaries.
-- `14857692`: smoke nearest-1 transferred near-miss residual retrieval.
-- `14857693`: full nearest-1 transferred near-miss residual retrieval.
-- `14857694`: summary nearest-1 transferred near-miss residual retrieval.
-- `14857695`: smoke KNN4 transferred near-miss residual retrieval.
-- `14857696`: full KNN4 transferred near-miss residual retrieval.
-- `14857697`: summary KNN4 transferred near-miss residual retrieval.
-- `14857698`: rebuild `paper_table_status.*` after near-miss residual summaries.
-
-Current scheduler state: `field_optim` and `nonexpert_policy_bc5` jobs completed.
-The first residual rollout smokes failed on a missing `retrieval_neighbors`
-argument; fixed v2 residual jobs `14857111`-`14857117` are pending on
-priority/dependencies. The fixed smokes passed and full rollouts are running;
-near-miss-only residual diagnostics `14857692`-`14857698` are also running or
-dependency-held.
+- `14857112`: fixed nearest-1 `retrieval_residual` full rollout. Completed.
+- `14857113`: fixed nearest-1 `retrieval_residual` summary. Completed.
+- `14857114`: fixed KNN4 `retrieval_residual` smoke. Completed.
+- `14857115`: fixed KNN4 `retrieval_residual` full rollout. Completed.
+- `14857116`: fixed KNN4 `retrieval_residual` summary. Completed.
+- `14857117`: rebuild `paper_table_status.*` after fixed residual summaries. Completed.
+- `14857692`: smoke nearest-1 transferred near-miss residual retrieval. Completed.
+- `14857693`: full nearest-1 transferred near-miss residual retrieval. Canceled.
+- `14857694`: summary nearest-1 transferred near-miss residual retrieval. Canceled.
+- `14857695`: smoke KNN4 transferred near-miss residual retrieval. Completed.
+- `14857696`: full KNN4 transferred near-miss residual retrieval. Canceled.
+- `14857697`: summary KNN4 transferred near-miss residual retrieval. Canceled.
+- `14857698`: rebuild `paper_table_status.*` after near-miss residual summaries. Canceled.
+
+Current scheduler state: no tracked jobs are active. `field_optim`,
+`nonexpert_policy_bc5`, and residual v2 jobs completed. Residual nearest-1 is a
+positive clean bridge at 32.12%, KNN4 residual is 29.91%, and near-miss-only
+residual smoke was weak enough to cancel its full jobs.
 
 ## Decision Rule For Field Optim Jobs
 

results/paper_table_status.json

@@ -158,14 +158,14 @@
       "story_role": "pending transferable local tangent proposal",
       "fallback_success": null,
       "pending_job": "14857111/14857112/14857113",
-      "path_exists": false,
-      "status": "pending",
-      "success": null,
-      "std_success": null,
+      "path_exists": true,
+      "status": "complete",
+      "success": 0.32115942028985506,
+      "std_success": 0.012581179370556922,
       "completed_seeds": null,
-      "num_completed": null,
+      "num_completed": 3,
       "best_config": null,
-      "gain_vs_h16_policy": null
+      "gain_vs_h16_policy": 0.02376811594202899
     },
     {
       "key": "retrieval_residual_knn4",
@@ -177,52 +177,14 @@
       "story_role": "pending KNN tangent proposal",
       "fallback_success": null,
       "pending_job": "14857114/14857115/14857116",
-      "path_exists": false,
-      "status": "pending",
-      "success": null,
-      "std_success": null,
-      "completed_seeds": null,
-      "num_completed": null,
-      "best_config": null,
-      "gain_vs_h16_policy": null
-    },
-    {
-      "key": "retrieval_residual_nearmiss",
-      "label": "Train-state near-miss residual retrieval",
-      "path": "h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_nearmiss_v2_summary.json",
-      "clean_deployment": "yes",
-      "same_state_proposals": "no",
-      "expert_proposal": "no",
-      "story_role": "pending transferable near-miss tangent proposal",
-      "fallback_success": null,
-      "pending_job": "14857692/14857693/14857694",
-      "path_exists": false,
-      "status": "pending",
-      "success": null,
-      "std_success": null,
-      "completed_seeds": null,
-      "num_completed": null,
-      "best_config": null,
-      "gain_vs_h16_policy": null
-    },
-    {
-      "key": "retrieval_residual_nearmiss_knn4",
-      "label": "KNN near-miss residual retrieval",
-      "path": "h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_nearmiss_knn4_v2_summary.json",
-      "clean_deployment": "yes",
-      "same_state_proposals": "no",
-      "expert_proposal": "no",
-      "story_role": "pending KNN near-miss tangent proposal",
-      "fallback_success": null,
-      "pending_job": "14857695/14857696/14857697",
-      "path_exists": false,
-      "status": "pending",
-      "success": null,
-      "std_success": null,
+      "path_exists": true,
+      "status": "complete",
+      "success": 0.2991304347826087,
+      "std_success": 0.02005663059942746,
       "completed_seeds": null,
-      "num_completed": null,
+      "num_completed": 3,
       "best_config": null,
-      "gain_vs_h16_policy": null
+      "gain_vs_h16_policy": 0.001739130434782632
     },
     {
       "key": "near_miss_only_lattice",
@@ -349,9 +311,7 @@
     "Do not claim external SOTA from this table alone; add current external baselines separately.",
     "Current best clean deployment row is Near-miss proposal policy + field at 32.93%.",
     "Trust-region field optimization should be framed as a negative/diagnostic ablation.",
-    "Train-state counterfactual residual retrieval is pending (14857111/14857112/14857113).",
-    "KNN counterfactual residual retrieval is pending (14857114/14857115/14857116).",
-    "Train-state near-miss residual retrieval is pending (14857692/14857693/14857694).",
-    "KNN near-miss residual retrieval is pending (14857695/14857696/14857697)."
+    "Train-state counterfactual residual retrieval is a positive clean bridge but remains below the current clean best.",
+    "KNN counterfactual residual retrieval is a positive clean bridge but remains below the current clean best."
   ]
 }

results/paper_table_status.md

@@ -11,10 +11,8 @@ Baseline h=16 policy: 29.74%
 | field_optim | Trust-region field optimization | complete k32_sigma0.50 | 25.39% | -4.35 pp | yes | no | no | differentiable field-ascent diagnostic |
 | nonexpert_policy_bc5 | Best non-expert proposal policy | complete | 27.88% | -1.86 pp | yes | no | no | pending broader non-expert proposal model |
 | nonexpert_policy_bc5_field | Best non-expert proposal policy + field | complete k64_sigma0.50 | 26.49% | -3.25 pp | yes | no | no | pending broader proposal-field bridge |
-| retrieval_residual | Train-state counterfactual residual retrieval | pending 14857111/14857112/14857113 | pending | pending | yes | no | no | pending transferable local tangent proposal |
-| retrieval_residual_knn4 | KNN counterfactual residual retrieval | pending 14857114/14857115/14857116 | pending | pending | yes | no | no | pending KNN tangent proposal |
-| retrieval_residual_nearmiss | Train-state near-miss residual retrieval | pending 14857692/14857693/14857694 | pending | pending | yes | no | no | pending transferable near-miss tangent proposal |
-| retrieval_residual_nearmiss_knn4 | KNN near-miss residual retrieval | pending 14857695/14857696/14857697 | pending | pending | yes | no | no | pending KNN near-miss tangent proposal |
+| retrieval_residual | Train-state counterfactual residual retrieval | complete | 32.12% | +2.38 pp | yes | no | no | pending transferable local tangent proposal |
+| retrieval_residual_knn4 | KNN counterfactual residual retrieval | complete | 29.91% | +0.17 pp | yes | no | no | pending KNN tangent proposal |
 | near_miss_only_lattice | Same-state lattice, near-miss only | complete | 55.94% | +26.20 pp | no | yes | no | minimal mechanism result |
 | no_expert_lattice | Same-state lattice, no expert | complete | 56.99% | +27.25 pp | no | yes | no | main conservative mechanism result |
 | no_near_miss_no_expert_lattice | Same-state lattice, no expert/no near-miss | complete | 25.57% | -4.17 pp | no | yes | no | mechanism knockout |
@@ -27,7 +25,5 @@ Baseline h=16 policy: 29.74%
 - Do not claim external SOTA from this table alone; add current external baselines separately.
 - Current best clean deployment row is Near-miss proposal policy + field at 32.93%.
 - Trust-region field optimization should be framed as a negative/diagnostic ablation.
-- Train-state counterfactual residual retrieval is pending (14857111/14857112/14857113).
-- KNN counterfactual residual retrieval is pending (14857114/14857115/14857116).
-- Train-state near-miss residual retrieval is pending (14857692/14857693/14857694).
-- KNN near-miss residual retrieval is pending (14857695/14857696/14857697).
+- Train-state counterfactual residual retrieval is a positive clean bridge but remains below the current clean best.
+- KNN counterfactual residual retrieval is a positive clean bridge but remains below the current clean best.

scripts/build_paper_table_status.py

@@ -96,44 +96,44 @@ SPECS = [
         pending_job="14842574/14842577/14842617",
     ),
     ResultSpec(
-        key="retrieval_residual",
-        label="Train-state counterfactual residual retrieval",
-        path="h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_v2_summary.json",
+        key="field_selected_noexpert_policy",
+        label="Field-selected no-expert distillation policy",
+        path="h16_policy_ckpt_field_selected_noexpert_bc5_summary.json",
         clean_deployment="yes",
         same_state_proposals="no",
         expert_proposal="no",
-        story_role="pending transferable local tangent proposal",
-        pending_job="14857111/14857112/14857113",
+        story_role="pending student of field-on-lattice teacher",
+        pending_job="14858327/14858328/14858329/14858330",
     ),
     ResultSpec(
-        key="retrieval_residual_knn4",
-        label="KNN counterfactual residual retrieval",
-        path="h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_knn4_v2_summary.json",
+        key="field_selected_noexpert_policy_field",
+        label="Field-selected no-expert distillation + field",
+        path="h16_policy_ckpt_field_selected_noexpert_bc5_bestpt_field_sweep_summary.json",
         clean_deployment="yes",
         same_state_proposals="no",
         expert_proposal="no",
-        story_role="pending KNN tangent proposal",
-        pending_job="14857114/14857115/14857116",
+        story_role="pending student proposal with field scoring",
+        pending_job="14858327/14858328/14858331/14858332",
     ),
     ResultSpec(
-        key="retrieval_residual_nearmiss",
-        label="Train-state near-miss residual retrieval",
-        path="h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_nearmiss_v2_summary.json",
+        key="retrieval_residual",
+        label="Train-state counterfactual residual retrieval",
+        path="h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_v2_summary.json",
         clean_deployment="yes",
         same_state_proposals="no",
         expert_proposal="no",
-        story_role="pending transferable near-miss tangent proposal",
-        pending_job="14857692/14857693/14857694",
+        story_role="pending transferable local tangent proposal",
+        pending_job="14857111/14857112/14857113",
     ),
     ResultSpec(
-        key="retrieval_residual_nearmiss_knn4",
-        label="KNN near-miss residual retrieval",
-        path="h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_nearmiss_knn4_v2_summary.json",
+        key="retrieval_residual_knn4",
+        label="KNN counterfactual residual retrieval",
+        path="h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_knn4_v2_summary.json",
         clean_deployment="yes",
         same_state_proposals="no",
         expert_proposal="no",
-        story_role="pending KNN near-miss tangent proposal",
-        pending_job="14857695/14857696/14857697",
+        story_role="pending KNN tangent proposal",
+        pending_job="14857114/14857115/14857116",
     ),
     ResultSpec(
         key="near_miss_only_lattice",
@@ -287,8 +287,6 @@ def _decision_notes(rows: list[dict[str, Any]]) -> list[str]:
         "field_optim",
         "retrieval_residual",
         "retrieval_residual_knn4",
-        "retrieval_residual_nearmiss",
-        "retrieval_residual_nearmiss_knn4",
     ):
         row = by_key[key]
         if row["success"] is None:
@@ -297,6 +295,10 @@ def _decision_notes(rows: list[dict[str, Any]]) -> list[str]:
             notes.append(f"{row['label']} is strong enough to promote as a clean bridge.")
         elif row["success"] > 0.3293:
             notes.append(f"{row['label']} improves the clean bridge but is not yet the main result.")
+        elif row["success"] > BASELINE_H16_POLICY:
+            notes.append(
+                f"{row['label']} is a positive clean bridge but remains below the current clean best."
+            )
         else:
             notes.append(f"{row['label']} should be framed as a negative/diagnostic ablation.")
     return notes

scripts/export_field_selected_policy_targets.py

@@ -0,0 +1,159 @@
+#!/usr/bin/env python
+from __future__ import annotations
+
+import argparse
+import json
+import sys
+from pathlib import Path
+from typing import Any
+
+PROJECT_ROOT = Path(__file__).resolve().parents[1]
+if str(PROJECT_ROOT) not in sys.path:
+    sys.path.insert(0, str(PROJECT_ROOT))
+
+from dovla_cil.data.datasets import CILDataset  # noqa: E402
+from dovla_cil.eval.lattice_eval import _validation_group_ids  # noqa: E402
+from dovla_cil.eval.maniskill_policy_rollout import _numeric_action_values  # noqa: E402
+from dovla_cil.models.dovla import (  # noqa: E402
+    DoVLAConfig,
+    DoVLAModel,
+    load_model_state,
+    vectorize_toy_observation,
+)
+
+
+def main(argv: list[str] | None = None) -> int:
+    parser = argparse.ArgumentParser(
+        description="Export policy BC targets chosen by a trained field on CIL action lattices."
+    )
+    parser.add_argument("--checkpoint", type=Path, required=True)
+    parser.add_argument("--dataset", type=Path, required=True)
+    parser.add_argument("--out", type=Path, required=True)
+    parser.add_argument("--device", default="auto")
+    parser.add_argument("--split", choices=("train", "val", "all"), default="train")
+    parser.add_argument("--batch-groups", type=int, default=32)
+    parser.add_argument(
+        "--exclude-types",
+        default="expert",
+        help="Comma-separated candidate_type values to exclude before field selection.",
+    )
+    parser.add_argument("--max-groups", type=int, default=None)
+    args = parser.parse_args(argv)
+
+    if args.batch_groups <= 0:
+        parser.error("--batch-groups must be positive")
+    if args.max_groups is not None and args.max_groups <= 0:
+        parser.error("--max-groups must be positive when provided")
+
+    try:
+        import torch
+    except ImportError as exc:  # pragma: no cover
+        raise ImportError("export_field_selected_policy_targets.py requires torch") from exc
+
+    checkpoint = torch.load(
+        args.checkpoint,
+        map_location=_resolve_device(args.device),
+        weights_only=False,
+    )
+    model_config = DoVLAConfig(**checkpoint["model_config"])
+    if model_config.observation_mode != "state":
+        raise ValueError("field-selected target export currently supports state observations only")
+    device = _resolve_device(args.device)
+    model = DoVLAModel(model_config).to(device)
+    load_model_state(model, checkpoint)
+    model.eval()
+
+    dataset = CILDataset(args.dataset)
+    trainer_config = checkpoint.get("trainer_config", {})
+    val_ids = set(
+        _validation_group_ids(
+            dataset.group_ids,
+            val_fraction=float(trainer_config.get("val_fraction", 0.2)),
+            seed=int(trainer_config.get("seed", 0)),
+        )
+    )
+    if args.split == "train":
+        group_ids = [group_id for group_id in dataset.group_ids if group_id not in val_ids]
+    elif args.split == "val":
+        group_ids = [group_id for group_id in dataset.group_ids if group_id in val_ids]
+    else:
+        group_ids = list(dataset.group_ids)
+    if args.max_groups is not None:
+        group_ids = group_ids[: args.max_groups]
+
+    excluded = {item.strip() for item in args.exclude_types.split(",") if item.strip()}
+    targets: dict[str, dict[str, Any]] = {}
+    counts: dict[str, int] = {}
+    with torch.no_grad():
+        for start in range(0, len(group_ids), args.batch_groups):
+            for group_id in group_ids[start : start + args.batch_groups]:
+                records = [
+                    record
+                    for record in dataset.get_group(group_id)
+                    if record.candidate_type not in excluded
+                ]
+                if not records:
+                    records = dataset.get_group(group_id)
+                if not records:
+                    continue
+                obs = torch.tensor(
+                    [
+                        vectorize_toy_observation(
+                            records[0].observation_inline or {},
+                            obs_dim=model_config.obs_dim,
+                        )
+                    ]
+                    * len(records),
+                    dtype=torch.float32,
+                    device=device,
+                )
+                actions = torch.tensor(
+                    [_numeric_action_values(record) for record in records],
+                    dtype=torch.float32,
+                    device=device,
+                )
+                field = model.forward_field(
+                    obs,
+                    [record.instruction for record in records],
+                    actions,
+                )
+                best_idx = int(torch.argmax(field["potential"].reshape(len(records))).item())
+                best = records[best_idx]
+                counts[best.candidate_type] = counts.get(best.candidate_type, 0) + 1
+                targets[group_id] = {
+                    "record_id": best.record_id,
+                    "candidate_type": best.candidate_type,
+                    "task_id": best.task_id,
+                    "score": float(best.reward.score),
+                    "rank_within_group": best.rank_within_group,
+                }
+
+    payload = {
+        "checkpoint": str(args.checkpoint),
+        "dataset": str(args.dataset),
+        "split": args.split,
+        "excluded_candidate_types": sorted(excluded),
+        "num_groups": len(group_ids),
+        "num_targets": len(targets),
+        "selected_candidate_type_counts": counts,
+        "targets": targets,
+    }
+    args.out.parent.mkdir(parents=True, exist_ok=True)
+    args.out.write_text(json.dumps(payload, indent=2) + "\n")
+    print(json.dumps({k: v for k, v in payload.items() if k != "targets"}, indent=2))
+    print(f"Wrote {args.out}")
+    return 0
+
+
+def _resolve_device(device: str) -> str:
+    if device != "auto":
+        return device
+    try:
+        import torch
+    except ImportError:  # pragma: no cover
+        return "cpu"
+    return "cuda" if torch.cuda.is_available() else "cpu"
+
+
+if __name__ == "__main__":
+    raise SystemExit(main())

scripts/slurm/train_dovla_h16_policy_ckpt.sbatch

@@ -26,6 +26,7 @@ DATASET="${DATASET:-$SCRATCH_ROOT/experiments/h16_merged_dataset}"
 RUN_ROOT="${RUN_ROOT:-$SCRATCH_ROOT/experiments/dovla_h16_policy_ckpt_runs}"
 OBJECTIVE="${OBJECTIVE:-base}"
 POLICY_TARGET_TYPES="${POLICY_TARGET_TYPES:-}"
+POLICY_TARGET_MAP="${POLICY_TARGET_MAP:-}"
 SEED=$SLURM_ARRAY_TASK_ID
 OUT_DIR="$RUN_ROOT/$OBJECTIVE/seed_$SEED"
 
@@ -50,6 +51,9 @@ TRAIN_EXTRA_ARGS=()
 if [[ -n "$POLICY_TARGET_TYPES" ]]; then
   TRAIN_EXTRA_ARGS+=(--policy-target-types "$POLICY_TARGET_TYPES")
 fi
+if [[ -n "$POLICY_TARGET_MAP" ]]; then
+  TRAIN_EXTRA_ARGS+=(--policy-target-map "$POLICY_TARGET_MAP")
+fi
 if [[ -n "${EXTRA_TRAIN_ARGS:-}" ]]; then
   # shellcheck disable=SC2206
   EXTRA_SPLIT=($EXTRA_TRAIN_ARGS)
@@ -62,6 +66,7 @@ echo "Seed: $SEED"
 echo "Dataset: $DATASET"
 echo "Output: $OUT_DIR"
 echo "Policy target types: ${POLICY_TARGET_TYPES:-<best-any>}"
+echo "Policy target map: ${POLICY_TARGET_MAP:-<none>}"
 echo "=================================================="
 
 "${PYTHON_CMD[@]}" -c "

scripts/train_dovla.py

@@ -85,6 +85,13 @@ def main(argv: list[str] | None = None) -> int:
         help="Comma-separated candidate_type filter for policy BC targets. "
         "Empty means best action from every group.",
     )
+    parser.add_argument(
+        "--policy-target-map",
+        type=Path,
+        default=None,
+        help="JSON mapping group_id to policy BC target record_id. Missing groups fall back "
+        "to --policy-target-types or best-in-group.",
+    )
     parser.add_argument(
         "--loss-weight",
         action="append",
@@ -129,6 +136,7 @@ def main(argv: list[str] | None = None) -> int:
         policy_target_types=tuple(
             item.strip() for item in args.policy_target_types.split(",") if item.strip()
         ),
+        policy_target_map=args.policy_target_map,
         losses=loss_weights,
     )
     result = DoVLATrainer(config).train()

tests/test_trainer.py

@@ -141,3 +141,40 @@ def test_policy_target_type_filter_selects_best_allowed_candidate() -> None:
         "g0": "near",
         "g1": "fallback",
     }
+
+
+def test_policy_target_map_overrides_group_target_with_fallback() -> None:
+    records = [
+        SimpleNamespace(
+            group_id="g0",
+            candidate_type="expert",
+            reward=SimpleNamespace(score=2.0),
+            rank_within_group=0,
+            record_id="expert",
+        ),
+        SimpleNamespace(
+            group_id="g0",
+            candidate_type="near_miss",
+            reward=SimpleNamespace(score=1.5),
+            rank_within_group=1,
+            record_id="field_choice",
+        ),
+        SimpleNamespace(
+            group_id="g1",
+            candidate_type="near_miss",
+            reward=SimpleNamespace(score=0.7),
+            rank_within_group=1,
+            record_id="fallback",
+        ),
+    ]
+
+    selected = _best_records_by_group(
+        records,
+        candidate_types=("near_miss",),
+        target_record_ids={"g0": "field_choice"},
+    )
+
+    assert {record.group_id: record.record_id for record in selected} == {
+        "g0": "field_choice",
+        "g1": "fallback",
+    }