Auto-sync: 2026-06-28 17:30:45 (part 2)

results/paper_table_status.json

@@ -612,6 +612,63 @@
       "best_config": null,
       "gain_vs_h16_policy": 0.05217391304347829
     },
+    {
+      "key": "retrieval_residual_k4_mean_noopbonus003",
+      "label": "K4 mean-by-type residual retrieval, scale 0.40, margin 0.20, no-op residual bonus 0.03",
+      "path": "h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_k4s040_safe_margin0p20_mean_by_type_noopbonus0p03_summary.json",
+      "clean_deployment": "yes",
+      "same_state_proposals": "no",
+      "expert_proposal": "no",
+      "story_role": "typed sparse-intervention prior diagnostic",
+      "fallback_success": null,
+      "pending_job": "14883370/14883371",
+      "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_k4_mean_noopbonus005",
+      "label": "K4 mean-by-type residual retrieval, scale 0.40, margin 0.20, no-op residual bonus 0.05",
+      "path": "h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_k4s040_safe_margin0p20_mean_by_type_noopbonus0p05_summary.json",
+      "clean_deployment": "yes",
+      "same_state_proposals": "no",
+      "expert_proposal": "no",
+      "story_role": "typed sparse-intervention prior diagnostic",
+      "fallback_success": null,
+      "pending_job": "14883372/14883373",
+      "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_k4_mean_noopbonus008",
+      "label": "K4 mean-by-type residual retrieval, scale 0.40, margin 0.20, no-op residual bonus 0.08",
+      "path": "h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_k4s040_safe_margin0p20_mean_by_type_noopbonus0p08_summary.json",
+      "clean_deployment": "yes",
+      "same_state_proposals": "no",
+      "expert_proposal": "no",
+      "story_role": "typed sparse-intervention prior diagnostic",
+      "fallback_success": null,
+      "pending_job": "14883374/14883375",
+      "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_policy_anchor_scale035_safe",
       "label": "Policy-relative train-state residual retrieval, scale 0.35, safe non-expert residuals",

results/paper_table_status.md

@@ -35,6 +35,9 @@ Baseline h=16 policy: 29.74%
 | retrieval_residual_k2grid_broad_safe_ray_margin020 | K2 train-state residual ray search, safe residuals, scales 0.20/0.35/0.50/0.65, advantage margin 0.20 | complete | 34.96% | +5.22 pp | yes | no | no | counterfactual tangent ray-search diagnostic |
 | retrieval_residual_k4grid_tight_safe_ray_margin020 | K4 train-state residual ray search, safe residuals, scales 0.30/0.40/0.50, advantage margin 0.20 | complete | 34.55% | +4.81 pp | yes | no | no | counterfactual tangent ray-search diagnostic |
 | retrieval_residual_k4_scale040_safe_margin020_mean_by_type | K4 train-state residual retrieval, scale 0.40, safe residuals, mean-by-type tangent consensus | complete | 34.96% | +5.22 pp | yes | no | no | counterfactual tangent consensus near-tie ablation |
+| retrieval_residual_k4_mean_noopbonus003 | K4 mean-by-type residual retrieval, scale 0.40, margin 0.20, no-op residual bonus 0.03 | pending 14883370/14883371 | pending | pending | yes | no | no | typed sparse-intervention prior diagnostic |
+| retrieval_residual_k4_mean_noopbonus005 | K4 mean-by-type residual retrieval, scale 0.40, margin 0.20, no-op residual bonus 0.05 | pending 14883372/14883373 | pending | pending | yes | no | no | typed sparse-intervention prior diagnostic |
+| retrieval_residual_k4_mean_noopbonus008 | K4 mean-by-type residual retrieval, scale 0.40, margin 0.20, no-op residual bonus 0.08 | pending 14883374/14883375 | pending | pending | yes | no | no | typed sparse-intervention prior diagnostic |
 | retrieval_residual_policy_anchor_scale035_safe | Policy-relative train-state residual retrieval, scale 0.35, safe non-expert residuals | complete | 33.74% | +4.00 pp | yes | no | no | policy-relative tangent anchor diagnostic |
 | retrieval_residual_scale030_safe_types | Train-state residual retrieval, scale 0.30, policy/no-op/wrong-gripper residuals | complete | 33.51% | +3.77 pp | yes | no | no | typed tangent scale zoom sweep |
 | retrieval_residual_scale0325_safe_types | Train-state residual retrieval, scale 0.325, policy/no-op/wrong-gripper residuals | complete | 33.74% | +4.00 pp | yes | no | no | typed tangent scale zoom sweep |

scripts/build_paper_table_status.py

@@ -335,6 +335,36 @@ SPECS = [
         story_role="counterfactual tangent consensus near-tie ablation",
         pending_job="14868699/14868700",
     ),
+    ResultSpec(
+        key="retrieval_residual_k4_mean_noopbonus003",
+        label="K4 mean-by-type residual retrieval, scale 0.40, margin 0.20, no-op residual bonus 0.03",
+        path="h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_k4s040_safe_margin0p20_mean_by_type_noopbonus0p03_summary.json",
+        clean_deployment="yes",
+        same_state_proposals="no",
+        expert_proposal="no",
+        story_role="typed sparse-intervention prior diagnostic",
+        pending_job="14883694/14883695",
+    ),
+    ResultSpec(
+        key="retrieval_residual_k4_mean_noopbonus005",
+        label="K4 mean-by-type residual retrieval, scale 0.40, margin 0.20, no-op residual bonus 0.05",
+        path="h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_k4s040_safe_margin0p20_mean_by_type_noopbonus0p05_summary.json",
+        clean_deployment="yes",
+        same_state_proposals="no",
+        expert_proposal="no",
+        story_role="typed sparse-intervention prior diagnostic",
+        pending_job="14883696/14883697",
+    ),
+    ResultSpec(
+        key="retrieval_residual_k4_mean_noopbonus008",
+        label="K4 mean-by-type residual retrieval, scale 0.40, margin 0.20, no-op residual bonus 0.08",
+        path="h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_k4s040_safe_margin0p20_mean_by_type_noopbonus0p08_summary.json",
+        clean_deployment="yes",
+        same_state_proposals="no",
+        expert_proposal="no",
+        story_role="typed sparse-intervention prior diagnostic",
+        pending_job="14883698/14883699",
+    ),
     ResultSpec(
         key="retrieval_residual_policy_anchor_scale035_safe",
         label="Policy-relative train-state residual retrieval, scale 0.35, safe non-expert residuals",

scripts/eval_maniskill_policy_rollout.py

@@ -158,10 +158,28 @@ def main(argv: list[str] | None = None) -> int:
         help="Comma-separated candidate_type values to mask in selection-mode=lattice, "
         "for example 'expert'.",
     )
+    parser.add_argument(
+        "--candidate-type-bonuses",
+        default="",
+        help="Comma-separated candidate_type=bonus priors added to field potentials before "
+        "selection, e.g. 'residual_no_op=0.05'. Empty preserves previous behavior.",
+    )
     args = parser.parse_args(argv)
     lattice_exclude_types = tuple(
         item.strip() for item in args.lattice_exclude_types.split(",") if item.strip()
     )
+    candidate_type_bonuses: dict[str, float] = {}
+    try:
+        for item in args.candidate_type_bonuses.split(","):
+            item = item.strip()
+            if not item:
+                continue
+            key, value = item.split("=", 1)
+            candidate_type_bonuses[key.strip()] = float(value.strip())
+    except ValueError as exc:
+        raise SystemExit(
+            "--candidate-type-bonuses must contain comma-separated candidate_type=bonus pairs"
+        ) from exc
     try:
         retrieval_residual_scales = tuple(
             float(item.strip())
@@ -198,6 +216,7 @@ def main(argv: list[str] | None = None) -> int:
         retrieval_residual_anchor=args.retrieval_residual_anchor,
         retrieval_residual_reduce=args.retrieval_residual_reduce,
         lattice_exclude_types=lattice_exclude_types,
+        candidate_type_bonuses=candidate_type_bonuses,
     )
     print(json.dumps({key: value for key, value in result.items() if key != "rows"}, indent=2))
     return 0

scripts/slurm/eval_maniskill_policy_rollout.sbatch

@@ -64,6 +64,10 @@ LATTICE_EXCLUDE_TYPES="${LATTICE_EXCLUDE_TYPES:-}"
 if [[ -n "${LATTICE_EXCLUDE_TYPES_COLON:-}" ]]; then
   LATTICE_EXCLUDE_TYPES="${LATTICE_EXCLUDE_TYPES_COLON//:/,}"
 fi
+CANDIDATE_TYPE_BONUSES="${CANDIDATE_TYPE_BONUSES:-}"
+if [[ -n "${CANDIDATE_TYPE_BONUSES_COLON:-}" ]]; then
+  CANDIDATE_TYPE_BONUSES="${CANDIDATE_TYPE_BONUSES_COLON//:/,}"
+fi
 
 module load StdEnv/2023 apptainer/1.4.5
 cd "$PROJECT_DIR"
@@ -119,4 +123,5 @@ apptainer exec --nv \
     --retrieval-residual-anchor "$RETRIEVAL_RESIDUAL_ANCHOR" \
     --retrieval-residual-reduce "$RETRIEVAL_RESIDUAL_REDUCE" \
     --lattice-exclude-types "$LATTICE_EXCLUDE_TYPES" \
+    --candidate-type-bonuses "$CANDIDATE_TYPE_BONUSES" \
     "${EXTRA_ARGS[@]}"

scripts/slurm/eval_maniskill_policy_rollout_cpu_smoke.sbatch

@@ -35,7 +35,7 @@ CPU_RENDER_LIBS="$SCRATCH_ROOT/cpu_render_libs"
 CA_BUNDLE="$SCRATCH_ROOT/ca-bundle.crt"
 VULKAN_ICD="$CPU_RENDER_LIBS/share/vulkan/icd.d/lvp_icd.x86_64.json"
 MAX_GROUPS="${MAX_GROUPS:-8}"
-GROUP_BATCH_SIZE="${GROUP_BATCH_SIZE:-2}"
+GROUP_BATCH_SIZE="${GROUP_BATCH_SIZE:-1}"
 SIM_BACKEND="${SIM_BACKEND:-physx_cpu}"
 RENDER_BACKEND="${RENDER_BACKEND:-cpu}"
 ALL_GROUPS="${ALL_GROUPS:-0}"
@@ -63,6 +63,10 @@ LATTICE_EXCLUDE_TYPES="${LATTICE_EXCLUDE_TYPES:-}"
 if [[ -n "${LATTICE_EXCLUDE_TYPES_COLON:-}" ]]; then
   LATTICE_EXCLUDE_TYPES="${LATTICE_EXCLUDE_TYPES_COLON//:/,}"
 fi
+CANDIDATE_TYPE_BONUSES="${CANDIDATE_TYPE_BONUSES:-}"
+if [[ -n "${CANDIDATE_TYPE_BONUSES_COLON:-}" ]]; then
+  CANDIDATE_TYPE_BONUSES="${CANDIDATE_TYPE_BONUSES_COLON//:/,}"
+fi
 
 module load StdEnv/2023 apptainer/1.4.5
 cd "$PROJECT_DIR"
@@ -115,4 +119,5 @@ apptainer exec \
     --retrieval-residual-anchor "$RETRIEVAL_RESIDUAL_ANCHOR" \
     --retrieval-residual-reduce "$RETRIEVAL_RESIDUAL_REDUCE" \
     --lattice-exclude-types "$LATTICE_EXCLUDE_TYPES" \
+    --candidate-type-bonuses "$CANDIDATE_TYPE_BONUSES" \
     "${EXTRA_ARGS[@]}"