Auto-sync: 2026-06-29 12:38:51 (part 2)

results/paper_analysis.json

@@ -1,6 +1,6 @@
 {
   "best_clean_key": "residual_k4_composemasked_grid035040045_noopbonus003",
-  "generated_utc": "2026-06-29T16:26:29+00:00",
+  "generated_utc": "2026-06-29T16:37:59+00:00",
   "mechanism_gap": {
     "best_clean_vs_direct_same_ckpt": 0.07246376811594196,
     "best_clean_vs_h16": 0.0579710144927536,
@@ -1153,6 +1153,16 @@
       "source": "results/h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_k4_composemasked_grid035040045_safe_margin0p20_noopbonus0p03_summary.json",
       "std_success": 0.010190374394925787
     },
+    "residual_k4_composemasked_l2comp002_grid035040045_noopbonus003": {
+      "label": "K4 composed type-consensus tangents, masked, composite L2 penalty 0.02",
+      "missing": true,
+      "source": "results/h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_k4_composemasked_l2comp002_grid035040045_safe_margin0p20_noopbonus0p03_summary.json"
+    },
+    "residual_k4_composemasked_l2comp005_grid035040045_noopbonus003": {
+      "label": "K4 composed type-consensus tangents, masked, composite L2 penalty 0.05",
+      "missing": true,
+      "source": "results/h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_k4_composemasked_l2comp005_grid035040045_safe_margin0p20_noopbonus0p03_summary.json"
+    },
     "residual_k4_consensus": {
       "ci95_success": 0.04490086956521744,
       "label": "K4 mean-by-type tangent consensus",

results/paper_analysis.md

@@ -1,6 +1,6 @@
 # Paper Analysis
 
-Generated: `2026-06-29T16:26:29+00:00`
+Generated: `2026-06-29T16:37:59+00:00`
 
 ## Main Seed Statistics
 
@@ -48,6 +48,8 @@ Generated: `2026-06-29T16:26:29+00:00`
 | residual_k4_composemasked_grid035040045 | K4 composed type-consensus tangents, masked, scales 0.35/0.40/0.45 | 3 | 35.30% +/- 1.22 | +/- 3.02 | 56.91% | 0.410 | +5.57 pp |
 | residual_k4_composemasked_grid035040045_noopbonus003 | K4 composed type-consensus tangents, masked, scales 0.35/0.40/0.45, no-op bonus 0.03 | 3 | 35.54% +/- 1.02 | +/- 2.53 | 57.02% | 0.411 | +5.80 pp |
 | residual_k4_composemasked_compbonus_grid035040045_noopbonus003 | K4 composed type-consensus tangents, masked, component no-op bonus 0.03 | 3 | 35.36% +/- 1.16 | +/- 2.88 | 56.98% | 0.413 | +5.62 pp |
+| residual_k4_composemasked_l2comp002_grid035040045_noopbonus003 | K4 composed type-consensus tangents, masked, composite L2 penalty 0.02 | 0 | missing | missing | missing | missing | missing |
+| residual_k4_composemasked_l2comp005_grid035040045_noopbonus003 | K4 composed type-consensus tangents, masked, composite L2 penalty 0.05 | 0 | missing | missing | missing | missing | missing |
 | repair_nearmiss_k4_grid025035050_margin020 | K4 near-miss-to-expert repair tangent, scales 0.25/0.35/0.50, margin 0.20 | 3 | 34.32% +/- 1.35 | +/- 3.36 | 55.97% | 0.394 | +4.58 pp |
 | repair_nearmiss_k4_grid035050075_margin020 | K4 near-miss-to-expert repair tangent, scales 0.35/0.50/0.75, margin 0.20 | 3 | 34.38% +/- 1.50 | +/- 3.73 | 56.05% | 0.394 | +4.64 pp |
 | repair_nearmiss_k4_grid025035050_margin010 | K4 near-miss-to-expert repair tangent, scales 0.25/0.35/0.50, margin 0.10 | 3 | 34.14% +/- 1.48 | +/- 3.67 | 56.01% | 0.393 | +4.41 pp |

results/paper_table_status.json

@@ -1315,6 +1315,44 @@
       "best_config": null,
       "gain_vs_h16_policy": 0.05623188405797103
     },
+    {
+      "key": "retrieval_residual_k4_composemasked_l2comp002_grid035040045_noopbonus003",
+      "label": "K4 composed type-consensus residual retrieval, masked, composite L2 penalty 0.02",
+      "path": "h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_k4_composemasked_l2comp002_grid035040045_safe_margin0p20_noopbonus0p03_summary.json",
+      "clean_deployment": "yes",
+      "same_state_proposals": "no",
+      "expert_proposal": "no",
+      "story_role": "trust-radius penalty on composed local tangent candidates",
+      "fallback_success": null,
+      "pending_job": "14913944/14913955",
+      "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_composemasked_l2comp005_grid035040045_noopbonus003",
+      "label": "K4 composed type-consensus residual retrieval, masked, composite L2 penalty 0.05",
+      "path": "h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_k4_composemasked_l2comp005_grid035040045_safe_margin0p20_noopbonus0p03_summary.json",
+      "clean_deployment": "yes",
+      "same_state_proposals": "no",
+      "expert_proposal": "no",
+      "story_role": "stronger trust-radius penalty on composed local tangent candidates",
+      "fallback_success": null,
+      "pending_job": "14913951/14913956",
+      "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_repair_nearmiss_k4_grid025035050_margin020",
       "label": "K4 near-miss-to-expert repair tangent, scales 0.25/0.35/0.50, margin 0.20",

results/paper_table_status.md

@@ -72,6 +72,8 @@ Baseline h=16 policy: 29.74%
 | retrieval_residual_k4_composemasked_grid035040045 | K4 composed type-consensus residual retrieval, masked, scales 0.35/0.40/0.45, margin 0.20 | complete | 35.30% | +5.57 pp | yes | no | no | local tangent composition with anti-goal composite masks |
 | retrieval_residual_k4_composemasked_grid035040045_noopbonus003 | K4 composed type-consensus residual retrieval, masked, scales 0.35/0.40/0.45, margin 0.20, no-op bonus 0.03 | complete | 35.54% | +5.80 pp | yes | no | no | local tangent composition with anti-goal composite masks on the current best typed prior |
 | retrieval_residual_k4_composemasked_compbonus_grid035040045_noopbonus003 | K4 composed type-consensus residual retrieval, masked, component no-op bonus 0.03 | complete | 35.36% | +5.62 pp | yes | no | no | component-wise sparse prior on the masked local tangent composition chart |
+| retrieval_residual_k4_composemasked_l2comp002_grid035040045_noopbonus003 | K4 composed type-consensus residual retrieval, masked, composite L2 penalty 0.02 | pending 14913944/14913955 | pending | pending | yes | no | no | trust-radius penalty on composed local tangent candidates |
+| retrieval_residual_k4_composemasked_l2comp005_grid035040045_noopbonus003 | K4 composed type-consensus residual retrieval, masked, composite L2 penalty 0.05 | pending 14913951/14913956 | pending | pending | yes | no | no | stronger trust-radius penalty on composed local tangent candidates |
 | retrieval_repair_nearmiss_k4_grid025035050_margin020 | K4 near-miss-to-expert repair tangent, scales 0.25/0.35/0.50, margin 0.20 | complete | 34.32% | +4.58 pp | yes | no | no | deployment-clean corrective tangent transport from train near-misses back toward expert actions |
 | retrieval_repair_nearmiss_k4_grid035050075_margin020 | K4 near-miss-to-expert repair tangent, scales 0.35/0.50/0.75, margin 0.20 | complete | 34.38% | +4.64 pp | yes | no | no | repair-tangent scale diagnostic for near-miss counterfactual geometry |
 | retrieval_repair_nearmiss_k4_grid025035050_margin010 | K4 near-miss-to-expert repair tangent, scales 0.25/0.35/0.50, margin 0.10 | complete | 34.14% | +4.41 pp | yes | no | no | repair-tangent abstention diagnostic for near-miss counterfactual geometry |

scripts/build_paper_analysis.py

@@ -361,6 +361,22 @@ METHODS = [
             "k4_composemasked_compbonus_grid035040045_safe_margin0p20_noopbonus0p03_summary.json"
         ),
     ),
+    MethodSpec(
+        key="residual_k4_composemasked_l2comp002_grid035040045_noopbonus003",
+        label="K4 composed type-consensus tangents, masked, composite L2 penalty 0.02",
+        summary_path=(
+            "h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_"
+            "k4_composemasked_l2comp002_grid035040045_safe_margin0p20_noopbonus0p03_summary.json"
+        ),
+    ),
+    MethodSpec(
+        key="residual_k4_composemasked_l2comp005_grid035040045_noopbonus003",
+        label="K4 composed type-consensus tangents, masked, composite L2 penalty 0.05",
+        summary_path=(
+            "h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_"
+            "k4_composemasked_l2comp005_grid035040045_safe_margin0p20_noopbonus0p03_summary.json"
+        ),
+    ),
     MethodSpec(
         key="repair_nearmiss_k4_grid025035050_margin020",
         label="K4 near-miss-to-expert repair tangent, scales 0.25/0.35/0.50, margin 0.20",

scripts/build_paper_table_status.py

@@ -705,6 +705,26 @@ SPECS = [
         story_role="component-wise sparse prior on the masked local tangent composition chart",
         pending_job="14912561/14912562",
     ),
+    ResultSpec(
+        key="retrieval_residual_k4_composemasked_l2comp002_grid035040045_noopbonus003",
+        label="K4 composed type-consensus residual retrieval, masked, composite L2 penalty 0.02",
+        path="h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_k4_composemasked_l2comp002_grid035040045_safe_margin0p20_noopbonus0p03_summary.json",
+        clean_deployment="yes",
+        same_state_proposals="no",
+        expert_proposal="no",
+        story_role="trust-radius penalty on composed local tangent candidates",
+        pending_job="14913944/14913955",
+    ),
+    ResultSpec(
+        key="retrieval_residual_k4_composemasked_l2comp005_grid035040045_noopbonus003",
+        label="K4 composed type-consensus residual retrieval, masked, composite L2 penalty 0.05",
+        path="h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_k4_composemasked_l2comp005_grid035040045_safe_margin0p20_noopbonus0p03_summary.json",
+        clean_deployment="yes",
+        same_state_proposals="no",
+        expert_proposal="no",
+        story_role="stronger trust-radius penalty on composed local tangent candidates",
+        pending_job="14913951/14913956",
+    ),
     ResultSpec(
         key="retrieval_repair_nearmiss_k4_grid025035050_margin020",
         label="K4 near-miss-to-expert repair tangent, scales 0.25/0.35/0.50, margin 0.20",

scripts/eval_maniskill_policy_rollout.py

@@ -178,6 +178,13 @@ def main(argv: list[str] | None = None) -> int:
         help="Scale for adding a train-source reward-score advantage prior relative to the "
         "source residual anchor action.",
     )
+    parser.add_argument(
+        "--retrieval-residual-composite-l2-penalty-scale",
+        type=float,
+        default=0.0,
+        help="Penalty scale subtracted from composed residual candidates according to "
+        "their mean squared residual energy.",
+    )
     parser.add_argument(
         "--retrieval-residual-action-l2-penalty",
         type=float,
@@ -307,6 +314,9 @@ def main(argv: list[str] | None = None) -> int:
         retrieval_residual_source_advantage_bonus_scale=(
             args.retrieval_residual_source_advantage_bonus_scale
         ),
+        retrieval_residual_composite_l2_penalty_scale=(
+            args.retrieval_residual_composite_l2_penalty_scale
+        ),
         retrieval_residual_action_l2_penalty=args.retrieval_residual_action_l2_penalty,
         retrieval_residual_scale=args.retrieval_residual_scale,
         retrieval_residual_scales=retrieval_residual_scales,

scripts/slurm/eval_maniskill_policy_rollout.sbatch

@@ -60,6 +60,7 @@ RETRIEVAL_RESIDUAL_MIN_SOURCE_ADVANTAGE="${RETRIEVAL_RESIDUAL_MIN_SOURCE_ADVANTA
 RETRIEVAL_RESIDUAL_SOURCE_PROGRESS_BONUS_SCALE="${RETRIEVAL_RESIDUAL_SOURCE_PROGRESS_BONUS_SCALE:-0.0}"
 RETRIEVAL_RESIDUAL_SOURCE_SCORE_BONUS_SCALE="${RETRIEVAL_RESIDUAL_SOURCE_SCORE_BONUS_SCALE:-0.0}"
 RETRIEVAL_RESIDUAL_SOURCE_ADVANTAGE_BONUS_SCALE="${RETRIEVAL_RESIDUAL_SOURCE_ADVANTAGE_BONUS_SCALE:-0.0}"
+RETRIEVAL_RESIDUAL_COMPOSITE_L2_PENALTY_SCALE="${RETRIEVAL_RESIDUAL_COMPOSITE_L2_PENALTY_SCALE:-0.0}"
 RETRIEVAL_RESIDUAL_ACTION_L2_PENALTY="${RETRIEVAL_RESIDUAL_ACTION_L2_PENALTY:-0.0}"
 RETRIEVAL_RESIDUAL_SCALE="${RETRIEVAL_RESIDUAL_SCALE:-1.0}"
 RETRIEVAL_RESIDUAL_SCALES="${RETRIEVAL_RESIDUAL_SCALES:-}"
@@ -138,6 +139,7 @@ apptainer exec --nv \
     --retrieval-residual-source-progress-bonus-scale "$RETRIEVAL_RESIDUAL_SOURCE_PROGRESS_BONUS_SCALE" \
     --retrieval-residual-source-score-bonus-scale "$RETRIEVAL_RESIDUAL_SOURCE_SCORE_BONUS_SCALE" \
     --retrieval-residual-source-advantage-bonus-scale "$RETRIEVAL_RESIDUAL_SOURCE_ADVANTAGE_BONUS_SCALE" \
+    --retrieval-residual-composite-l2-penalty-scale "$RETRIEVAL_RESIDUAL_COMPOSITE_L2_PENALTY_SCALE" \
     --retrieval-residual-action-l2-penalty "$RETRIEVAL_RESIDUAL_ACTION_L2_PENALTY" \
     --retrieval-residual-scale "$RETRIEVAL_RESIDUAL_SCALE" \
     --retrieval-residual-scales "$RETRIEVAL_RESIDUAL_SCALES" \

scripts/slurm/eval_maniskill_policy_rollout_cpu_smoke.sbatch

@@ -59,6 +59,7 @@ RETRIEVAL_RESIDUAL_MIN_SOURCE_ADVANTAGE="${RETRIEVAL_RESIDUAL_MIN_SOURCE_ADVANTA
 RETRIEVAL_RESIDUAL_SOURCE_PROGRESS_BONUS_SCALE="${RETRIEVAL_RESIDUAL_SOURCE_PROGRESS_BONUS_SCALE:-0.0}"
 RETRIEVAL_RESIDUAL_SOURCE_SCORE_BONUS_SCALE="${RETRIEVAL_RESIDUAL_SOURCE_SCORE_BONUS_SCALE:-0.0}"
 RETRIEVAL_RESIDUAL_SOURCE_ADVANTAGE_BONUS_SCALE="${RETRIEVAL_RESIDUAL_SOURCE_ADVANTAGE_BONUS_SCALE:-0.0}"
+RETRIEVAL_RESIDUAL_COMPOSITE_L2_PENALTY_SCALE="${RETRIEVAL_RESIDUAL_COMPOSITE_L2_PENALTY_SCALE:-0.0}"
 RETRIEVAL_RESIDUAL_ACTION_L2_PENALTY="${RETRIEVAL_RESIDUAL_ACTION_L2_PENALTY:-0.0}"
 RETRIEVAL_RESIDUAL_SCALE="${RETRIEVAL_RESIDUAL_SCALE:-1.0}"
 RETRIEVAL_RESIDUAL_SCALES="${RETRIEVAL_RESIDUAL_SCALES:-}"
@@ -134,6 +135,7 @@ apptainer exec \
     --retrieval-residual-source-progress-bonus-scale "$RETRIEVAL_RESIDUAL_SOURCE_PROGRESS_BONUS_SCALE" \
     --retrieval-residual-source-score-bonus-scale "$RETRIEVAL_RESIDUAL_SOURCE_SCORE_BONUS_SCALE" \
     --retrieval-residual-source-advantage-bonus-scale "$RETRIEVAL_RESIDUAL_SOURCE_ADVANTAGE_BONUS_SCALE" \
+    --retrieval-residual-composite-l2-penalty-scale "$RETRIEVAL_RESIDUAL_COMPOSITE_L2_PENALTY_SCALE" \
     --retrieval-residual-action-l2-penalty "$RETRIEVAL_RESIDUAL_ACTION_L2_PENALTY" \
     --retrieval-residual-scale "$RETRIEVAL_RESIDUAL_SCALE" \
     --retrieval-residual-scales "$RETRIEVAL_RESIDUAL_SCALES" \

scripts/slurm/summarize_h16_policy_ckpt.sbatch

@@ -91,6 +91,9 @@ for result_path in sorted(base_dir.glob(f"seed_*/{out_name}")):
             "retrieval_residual_source_advantage_bonus_scale": data.get(
                 "retrieval_residual_source_advantage_bonus_scale", 0.0
             ),
+            "retrieval_residual_composite_l2_penalty_scale": data.get(
+                "retrieval_residual_composite_l2_penalty_scale", 0.0
+            ),
             "retrieval_residual_scale": data.get("retrieval_residual_scale", 0.0),
             "retrieval_residual_scales": data.get("retrieval_residual_scales", []),
             "retrieval_residual_anchor": data.get("retrieval_residual_anchor", "none"),

tests/test_maniskill_policy_rollout.py

@@ -1022,6 +1022,35 @@ def test_retrieval_residual_reducer_builds_composed_type_tangents() -> None:
     )
 
 
+def test_retrieval_residual_reducer_penalizes_composed_tangent_energy() -> None:
+    residuals, candidate_types, bonuses = _reduce_residual_candidates_by_type(
+        [
+            [[0.0, 0.0]],
+            [[0.4, 0.2]],
+            [[-0.2, 0.0]],
+        ],
+        [
+            "policy_residual",
+            "residual_no_op",
+            "residual_wrong_gripper",
+        ],
+        mode="compose_mean_by_type",
+        bonuses=[0.0, 0.04, 0.02],
+        composite_l2_penalty_scale=0.5,
+    )
+
+    assert candidate_types == [
+        "policy_residual",
+        "residual_no_op",
+        "residual_wrong_gripper",
+        "residual_no_op+residual_wrong_gripper",
+    ]
+    assert np.allclose(np.asarray(residuals[-1], dtype=np.float32), [[0.2, 0.2]])
+    # Base composite bonus is mean(0.04, 0.02)=0.03; residual energy is
+    # mean([0.2^2, 0.2^2])=0.04, so a 0.5 penalty subtracts 0.02.
+    assert np.allclose(bonuses, [0.0, 0.04, 0.02, 0.01])
+
+
 def test_retrieval_residual_reducer_penalizes_low_consensus_tangents() -> None:
     residuals, candidate_types, bonuses = _reduce_residual_candidates_by_type(
         [