Auto-sync: 2026-06-29 11:49:11

results/paper_analysis.json

@@ -1,11 +1,11 @@
 {
-  "best_clean_key": "residual_k4_consensus_grid035040045_noopbonus003",
-  "generated_utc": "2026-06-29T13:46:02+00:00",
+  "best_clean_key": "residual_k4_composemasked_grid035040045_noopbonus003",
+  "generated_utc": "2026-06-29T15:56:21+00:00",
   "mechanism_gap": {
-    "best_clean_vs_direct_same_ckpt": 0.07130434782608691,
-    "best_clean_vs_h16": 0.05681159420289855,
+    "best_clean_vs_direct_same_ckpt": 0.07246376811594196,
+    "best_clean_vs_h16": 0.0579710144927536,
     "same_state_full_vs_no_expert": 0.12347826086956515,
-    "same_state_no_expert_vs_best_clean": 0.2156521739130436,
+    "same_state_no_expert_vs_best_clean": 0.21449275362318854,
     "same_state_no_expert_vs_h16": 0.27246376811594214
   },
   "methods": {
@@ -844,6 +844,212 @@
       "source": "results/h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_k4_compose_grid035040045_safe_margin0p20_noopbonus0p03_summary.json",
       "std_success": 0.01578047256674342
     },
+    "residual_k4_composemasked_grid035040045": {
+      "ci95_success": 0.03024411731871433,
+      "label": "K4 composed type-consensus tangents, masked, scales 0.35/0.40/0.45",
+      "mean_action_mse_to_best": 0.4104903701765706,
+      "mean_progress": 0.5690829302439838,
+      "mean_success": 0.35304347826086957,
+      "num_completed": 3,
+      "per_task_success": {
+        "LiftPegUpright-v1": {
+          "mean_num_groups": 102.0,
+          "mean_success": 0.25997222475645976,
+          "std_success": 0.06389511734286676
+        },
+        "PickCube-v1": {
+          "mean_num_groups": 196.66666666666666,
+          "mean_success": 0.32090765514678554,
+          "std_success": 0.028741814746710958
+        },
+        "PullCube-v1": {
+          "mean_num_groups": 81.0,
+          "mean_success": 0.20844535581377688,
+          "std_success": 0.024567915948941907
+        },
+        "PushCube-v1": {
+          "mean_num_groups": 102.0,
+          "mean_success": 0.757403601199683,
+          "std_success": 0.07781968814460603
+        },
+        "StackCube-v1": {
+          "mean_num_groups": 93.33333333333333,
+          "mean_success": 0.19707329707329704,
+          "std_success": 0.0831467196651622
+        }
+      },
+      "seed_action_mse_to_best": {
+        "0": 0.40412475434496353,
+        "1": 0.39919882555532715,
+        "2": 0.42814753062942107
+      },
+      "seed_progress": {
+        "0": 0.5549376096017659,
+        "1": 0.5699067040403252,
+        "2": 0.5824044770898605
+      },
+      "seed_success": {
+        "0": 0.34782608695652173,
+        "1": 0.3443478260869565,
+        "2": 0.36695652173913046
+      },
+      "selected_candidate_type_counts": {
+        "retrieval_residual_policy_residual": 1623,
+        "retrieval_residual_residual_near_miss": 7,
+        "retrieval_residual_residual_near_miss+residual_no_op": 16,
+        "retrieval_residual_residual_near_miss+residual_wrong_gripper": 5,
+        "retrieval_residual_residual_no_op": 31,
+        "retrieval_residual_residual_no_op+residual_wrong_gripper": 24,
+        "retrieval_residual_residual_wrong_gripper": 19
+      },
+      "selected_residual_scale_counts": {
+        "0.35": 1637,
+        "0.4": 10,
+        "0.45": 78
+      },
+      "selected_type_outcomes": {
+        "retrieval_residual_policy_residual": {
+          "count": 1623.0,
+          "mean_progress": 0.5605780961007093,
+          "success_rate": 0.34442390634627235
+        },
+        "retrieval_residual_residual_near_miss": {
+          "count": 7.0,
+          "mean_progress": 0.8583278954029083,
+          "success_rate": 0.7142857142857143
+        },
+        "retrieval_residual_residual_near_miss+residual_no_op": {
+          "count": 16.0,
+          "mean_progress": 0.5099116703495383,
+          "success_rate": 0.25
+        },
+        "retrieval_residual_residual_near_miss+residual_wrong_gripper": {
+          "count": 5.0,
+          "mean_progress": 0.7557853102684021,
+          "success_rate": 0.6
+        },
+        "retrieval_residual_residual_no_op": {
+          "count": 31.0,
+          "mean_progress": 0.7701972944121207,
+          "success_rate": 0.5161290322580645
+        },
+        "retrieval_residual_residual_no_op+residual_wrong_gripper": {
+          "count": 24.0,
+          "mean_progress": 0.7473632637411356,
+          "success_rate": 0.5416666666666666
+        },
+        "retrieval_residual_residual_wrong_gripper": {
+          "count": 19.0,
+          "mean_progress": 0.6363769314791027,
+          "success_rate": 0.47368421052631576
+        }
+      },
+      "source": "results/h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_k4_composemasked_grid035040045_safe_margin0p20_summary.json",
+      "std_success": 0.012173913043478276
+    },
+    "residual_k4_composemasked_grid035040045_noopbonus003": {
+      "ci95_success": 0.025316336466430094,
+      "label": "K4 composed type-consensus tangents, masked, scales 0.35/0.40/0.45, no-op bonus 0.03",
+      "mean_action_mse_to_best": 0.41063271115952427,
+      "mean_progress": 0.570213930835129,
+      "mean_success": 0.35536231884057967,
+      "num_completed": 3,
+      "per_task_success": {
+        "LiftPegUpright-v1": {
+          "mean_num_groups": 102.0,
+          "mean_success": 0.25997222475645976,
+          "std_success": 0.06389511734286676
+        },
+        "PickCube-v1": {
+          "mean_num_groups": 196.66666666666666,
+          "mean_success": 0.3278169093386485,
+          "std_success": 0.02533117351097172
+        },
+        "PullCube-v1": {
+          "mean_num_groups": 81.0,
+          "mean_success": 0.20474165211007314,
+          "std_success": 0.019188581862611215
+        },
+        "PushCube-v1": {
+          "mean_num_groups": 102.0,
+          "mean_success": 0.757403601199683,
+          "std_success": 0.07781968814460603
+        },
+        "StackCube-v1": {
+          "mean_num_groups": 93.33333333333333,
+          "mean_success": 0.20044030044030045,
+          "std_success": 0.08896982358351117
+        }
+      },
+      "seed_action_mse_to_best": {
+        "0": 0.4049635051790139,
+        "1": 0.3985257065668702,
+        "2": 0.42840892173268874
+      },
+      "seed_progress": {
+        "0": 0.5575027554895243,
+        "1": 0.5714990638326044,
+        "2": 0.581639973183258
+      },
+      "seed_success": {
+        "0": 0.35130434782608694,
+        "1": 0.34782608695652173,
+        "2": 0.36695652173913046
+      },
+      "selected_candidate_type_counts": {
+        "retrieval_residual_policy_residual": 1608,
+        "retrieval_residual_residual_near_miss": 7,
+        "retrieval_residual_residual_near_miss+residual_no_op": 10,
+        "retrieval_residual_residual_near_miss+residual_wrong_gripper": 5,
+        "retrieval_residual_residual_no_op": 55,
+        "retrieval_residual_residual_no_op+residual_wrong_gripper": 21,
+        "retrieval_residual_residual_wrong_gripper": 19
+      },
+      "selected_residual_scale_counts": {
+        "0.35": 1625,
+        "0.4": 11,
+        "0.45": 89
+      },
+      "selected_type_outcomes": {
+        "retrieval_residual_policy_residual": {
+          "count": 1608.0,
+          "mean_progress": 0.5611789632519321,
+          "success_rate": 0.3451492537313433
+        },
+        "retrieval_residual_residual_near_miss": {
+          "count": 7.0,
+          "mean_progress": 0.8583278954029083,
+          "success_rate": 0.7142857142857143
+        },
+        "retrieval_residual_residual_near_miss+residual_no_op": {
+          "count": 10.0,
+          "mean_progress": 0.4397697255015373,
+          "success_rate": 0.3
+        },
+        "retrieval_residual_residual_near_miss+residual_wrong_gripper": {
+          "count": 5.0,
+          "mean_progress": 0.7557853102684021,
+          "success_rate": 0.6
+        },
+        "retrieval_residual_residual_no_op": {
+          "count": 55.0,
+          "mean_progress": 0.7199717064811425,
+          "success_rate": 0.4909090909090909
+        },
+        "retrieval_residual_residual_no_op+residual_wrong_gripper": {
+          "count": 21.0,
+          "mean_progress": 0.7318444358451026,
+          "success_rate": 0.5238095238095238
+        },
+        "retrieval_residual_residual_wrong_gripper": {
+          "count": 19.0,
+          "mean_progress": 0.6363769314791027,
+          "success_rate": 0.47368421052631576
+        }
+      },
+      "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_consensus": {
       "ci95_success": 0.04490086956521744,
       "label": "K4 mean-by-type tangent consensus",
@@ -5801,13 +6007,13 @@
   },
   "paired_deltas": {
     "best_clean - canonical_h16": {
-      "ci95_delta": 0.043707043917985626,
-      "left": "residual_k4_consensus_grid035040045_noopbonus003",
-      "mean_delta": 0.05681159420289857,
+      "ci95_delta": 0.042406376811594176,
+      "left": "residual_k4_composemasked_grid035040045_noopbonus003",
+      "mean_delta": 0.057971014492753624,
       "right": "h16_policy_canonical",
       "seed_deltas": {
-        "0": 0.06608695652173913,
-        "1": 0.03652173913043483,
+        "0": 0.0678260869565217,
+        "1": 0.038260869565217404,
         "2": 0.06782608695652176
       },
       "seeds": [
@@ -5815,16 +6021,16 @@
         1,
         2
       ],
-      "std_delta": 0.017593032933905507
+      "std_delta": 0.01706948621951936
     },
     "best_clean - direct_same_ckpt": {
-      "ci95_delta": 0.043205881883877584,
-      "left": "residual_k4_consensus_grid035040045_noopbonus003",
-      "mean_delta": 0.07130434782608697,
+      "ci95_delta": 0.041064533365485906,
+      "left": "residual_k4_composemasked_grid035040045_noopbonus003",
+      "mean_delta": 0.07246376811594202,
       "right": "near_miss_policy_bc5",
       "seed_deltas": {
-        "0": 0.07130434782608697,
-        "1": 0.05391304347826087,
+        "0": 0.07304347826086954,
+        "1": 0.055652173913043446,
         "2": 0.08869565217391306
       },
       "seeds": [
@@ -5832,7 +6038,7 @@
         1,
         2
       ],
-      "std_delta": 0.017391304347826098
+      "std_delta": 0.01652936513551684
     },
     "full_lattice - no_expert_lattice": {
       "ci95_delta": 0.02628111194117426,
@@ -5888,18 +6094,18 @@
   },
   "per_task_deltas": {
     "best_clean_vs_h16": {
-      "LiftPegUpright-v1": 0.03278131303915899,
-      "PickCube-v1": 0.1203569980743894,
-      "PullCube-v1": 0.002501202501202532,
-      "PushCube-v1": 0.0035460992907800915,
-      "StackCube-v1": 0.05981055981055983
+      "LiftPegUpright-v1": 0.029344886922320512,
+      "PickCube-v1": 0.1219595621769535,
+      "PullCube-v1": 0.006887167413483192,
+      "PushCube-v1": 0.006846429323783432,
+      "StackCube-v1": 0.0594738594738595
     },
     "no_expert_vs_best_clean": {
-      "LiftPegUpright-v1": 0.35374851747103375,
-      "PickCube-v1": 0.29526451809060505,
-      "PullCube-v1": -0.001875901875901892,
-      "PushCube-v1": 0.046939714097919305,
-      "StackCube-v1": 0.2817404817404817
+      "LiftPegUpright-v1": 0.35718494358787223,
+      "PickCube-v1": 0.29366195398804096,
+      "PullCube-v1": -0.006261866788182552,
+      "PushCube-v1": 0.043639384064915965,
+      "StackCube-v1": 0.282077182077182
     }
   }
 }

results/paper_analysis.md

@@ -1,6 +1,6 @@
 # Paper Analysis
 
-Generated: `2026-06-29T13:46:02+00:00`
+Generated: `2026-06-29T15:56:21+00:00`
 
 ## Main Seed Statistics
 
@@ -45,6 +45,8 @@ Generated: `2026-06-29T13:46:02+00:00`
 | residual_k4_consensus_grid035040045_noopbonus003_consensus010 | K4 mean-by-type tangent consensus, scales 0.35/0.40/0.45, no-op bonus 0.03, consensus penalty 0.10 | 3 | 35.36% +/- 1.16 | +/- 2.88 | 56.75% | 0.397 | +5.62 pp |
 | residual_k4_compose_grid035040045 | K4 composed type-consensus tangents, scales 0.35/0.40/0.45 | 3 | 34.09% +/- 1.55 | +/- 3.84 | 55.96% | 0.482 | +4.35 pp |
 | residual_k4_compose_grid035040045_noopbonus003 | K4 composed type-consensus tangents, scales 0.35/0.40/0.45, no-op bonus 0.03 | 3 | 34.14% +/- 1.58 | +/- 3.92 | 56.00% | 0.482 | +4.41 pp |
+| 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 |
 | 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 |
@@ -84,8 +86,8 @@ Generated: `2026-06-29T13:46:02+00:00`
 
 | comparison | seeds | mean delta | 95% CI | seed deltas |
 |---|---:|---:|---:|---|
-| best_clean - canonical_h16 | 3 | +5.68 pp | +/- 4.37 | 0:+6.61, 1:+3.65, 2:+6.78 |
-| best_clean - direct_same_ckpt | 3 | +7.13 pp | +/- 4.32 | 0:+7.13, 1:+5.39, 2:+8.87 |
+| best_clean - canonical_h16 | 3 | +5.80 pp | +/- 4.24 | 0:+6.78, 1:+3.83, 2:+6.78 |
+| best_clean - direct_same_ckpt | 3 | +7.25 pp | +/- 4.11 | 0:+7.30, 1:+5.57, 2:+8.87 |
 | no_expert_lattice - canonical_h16 | 3 | +27.25 pp | +/- 8.58 | 0:+23.30, 1:+28.70, 2:+29.74 |
 | full_lattice - no_expert_lattice | 3 | +12.35 pp | +/- 2.63 | 0:+13.57, 1:+11.83, 2:+11.65 |
 | policy_candidate_lattice - no_expert_lattice | 3 | -16.29 pp | +/- 7.55 | 0:-15.48, 1:-13.74, 2:-19.65 |
@@ -94,17 +96,17 @@ Generated: `2026-06-29T13:46:02+00:00`
 
 | task | h16 policy | best clean | near-miss lattice | no-expert lattice | full lattice | clean-h16 delta | noexpert-clean gap |
 |---|---:|---:|---:|---:|---:|---:|---:|
-| LiftPegUpright-v1 | 23.06% | 26.34% | 62.90% | 61.72% | 73.19% | +3.28 pp | +35.37 pp |
-| PickCube-v1 | 20.59% | 32.62% | 58.13% | 62.15% | 84.19% | +12.04 pp | +29.53 pp |
-| PullCube-v1 | 19.79% | 20.04% | 15.02% | 19.85% | 22.41% | +0.25 pp | -0.19 pp |
-| PushCube-v1 | 75.06% | 75.41% | 82.54% | 80.10% | 81.92% | +0.35 pp | +4.69 pp |
-| StackCube-v1 | 14.10% | 20.08% | 50.41% | 48.25% | 60.83% | +5.98 pp | +28.17 pp |
+| LiftPegUpright-v1 | 23.06% | 26.00% | 62.90% | 61.72% | 73.19% | +2.93 pp | +35.72 pp |
+| PickCube-v1 | 20.59% | 32.78% | 58.13% | 62.15% | 84.19% | +12.20 pp | +29.37 pp |
+| PullCube-v1 | 19.79% | 20.47% | 15.02% | 19.85% | 22.41% | +0.69 pp | -0.63 pp |
+| PushCube-v1 | 75.06% | 75.74% | 82.54% | 80.10% | 81.92% | +0.68 pp | +4.36 pp |
+| StackCube-v1 | 14.10% | 20.04% | 50.41% | 48.25% | 60.83% | +5.95 pp | +28.21 pp |
 
 ## Mechanism Gap
 
-- Best clean residual transport improves over canonical h16 by +5.68 pp.
+- Best clean residual transport improves over canonical h16 by +5.80 pp.
 - Same-state no-expert lattice improves over canonical h16 by +27.25 pp.
-- Remaining clean-to-same-state proposal gap is +21.57 pp.
+- Remaining clean-to-same-state proposal gap is +21.45 pp.
 - Full lattice adds expert proposals and reaches 69.33%, a +12.35 pp gain over no-expert.
 
 ## Selection Histograms
@@ -113,8 +115,8 @@ Generated: `2026-06-29T13:46:02+00:00`
 - `same_state_no_expert`: lattice_near_miss=1263 (73.2%), lattice_no_op=222 (12.9%), lattice_random_negative=144 (8.3%), lattice_wrong_gripper=62 (3.6%), lattice_wrong_direction=34 (2.0%)
 - `same_state_policy_baseline`: policy_continuous=1022 (59.2%), lattice_near_miss=448 (26.0%), lattice_no_op=119 (6.9%), lattice_random_negative=75 (4.3%), lattice_wrong_gripper=45 (2.6%), lattice_wrong_direction=16 (0.9%)
 - `same_state_full`: lattice_expert=977 (56.6%), lattice_near_miss=348 (20.2%), lattice_no_op=177 (10.3%), lattice_random_negative=138 (8.0%), lattice_wrong_gripper=55 (3.2%), lattice_wrong_direction=30 (1.7%)
-- `residual_k4_consensus_grid035040045_noopbonus003`: retrieval_residual_policy_residual=1631 (94.6%), retrieval_residual_residual_no_op=66 (3.8%), retrieval_residual_residual_wrong_gripper=28 (1.6%)
-- `residual_k4_consensus_grid035040045_noopbonus003` residual scale counts: {'0.35': 1643, '0.4': 10, '0.45': 72}
+- `residual_k4_composemasked_grid035040045_noopbonus003`: retrieval_residual_policy_residual=1608 (93.2%), retrieval_residual_residual_no_op=55 (3.2%), retrieval_residual_residual_no_op+residual_wrong_gripper=21 (1.2%), retrieval_residual_residual_wrong_gripper=19 (1.1%), retrieval_residual_residual_near_miss+residual_no_op=10 (0.6%), retrieval_residual_residual_near_miss=7 (0.4%), retrieval_residual_residual_near_miss+residual_wrong_gripper=5 (0.3%)
+- `residual_k4_composemasked_grid035040045_noopbonus003` residual scale counts: {'0.35': 1625, '0.4': 11, '0.45': 89}
 
 ## Selected-Type Outcomes
 

results/paper_table_status.json

@@ -1258,6 +1258,44 @@
       "best_config": null,
       "gain_vs_h16_policy": 0.04405797101449277
     },
+    {
+      "key": "retrieval_residual_k4_composemasked_grid035040045",
+      "label": "K4 composed type-consensus residual retrieval, masked, scales 0.35/0.40/0.45, margin 0.20",
+      "path": "h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_k4_composemasked_grid035040045_safe_margin0p20_summary.json",
+      "clean_deployment": "yes",
+      "same_state_proposals": "no",
+      "expert_proposal": "no",
+      "story_role": "local tangent composition with anti-goal composite masks",
+      "fallback_success": null,
+      "pending_job": "14911979/14911982",
+      "path_exists": true,
+      "status": "complete",
+      "success": 0.35304347826086957,
+      "std_success": 0.012173913043478276,
+      "completed_seeds": null,
+      "num_completed": 3,
+      "best_config": null,
+      "gain_vs_h16_policy": 0.0556521739130435
+    },
+    {
+      "key": "retrieval_residual_k4_composemasked_grid035040045_noopbonus003",
+      "label": "K4 composed type-consensus residual retrieval, masked, scales 0.35/0.40/0.45, margin 0.20, no-op bonus 0.03",
+      "path": "h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_k4_composemasked_grid035040045_safe_margin0p20_noopbonus0p03_summary.json",
+      "clean_deployment": "yes",
+      "same_state_proposals": "no",
+      "expert_proposal": "no",
+      "story_role": "local tangent composition with anti-goal composite masks on the current best typed prior",
+      "fallback_success": null,
+      "pending_job": "14911980/14911983",
+      "path_exists": true,
+      "status": "complete",
+      "success": 0.3553623188405797,
+      "std_success": 0.010190374394925787,
+      "completed_seeds": null,
+      "num_completed": 3,
+      "best_config": null,
+      "gain_vs_h16_policy": 0.05797101449275366
+    },
     {
       "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",
@@ -2305,23 +2343,23 @@
     }
   ],
   "best_clean": {
-    "key": "retrieval_residual_k4_mean_grid035040045_noopbonus003",
-    "label": "K4 mean-by-type residual retrieval, scales 0.35/0.40/0.45, margin 0.20, no-op bonus 0.03",
-    "path": "h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_k4_grid035040045_safe_margin0p20_mean_by_type_noopbonus0p03_summary.json",
+    "key": "retrieval_residual_k4_composemasked_grid035040045_noopbonus003",
+    "label": "K4 composed type-consensus residual retrieval, masked, scales 0.35/0.40/0.45, margin 0.20, no-op bonus 0.03",
+    "path": "h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_k4_composemasked_grid035040045_safe_margin0p20_noopbonus0p03_summary.json",
     "clean_deployment": "yes",
     "same_state_proposals": "no",
     "expert_proposal": "no",
-    "story_role": "scale-grid diagnostic for sparse mean-consensus residual transport",
+    "story_role": "local tangent composition with anti-goal composite masks on the current best typed prior",
     "fallback_success": null,
-    "pending_job": "14897988/14897990",
+    "pending_job": "14911980/14911983",
     "path_exists": true,
     "status": "complete",
-    "success": 0.3542028985507247,
-    "std_success": 0.011181044360571544,
+    "success": 0.3553623188405797,
+    "std_success": 0.010190374394925787,
     "completed_seeds": null,
     "num_completed": 3,
     "best_config": null,
-    "gain_vs_h16_policy": 0.05681159420289861
+    "gain_vs_h16_policy": 0.05797101449275366
   },
   "best_mechanism_no_expert": {
     "key": "no_expert_lattice",
@@ -2346,7 +2384,7 @@
     "Use no-expert same-state lattice as the conservative mechanism result, not as deployment-clean inference.",
     "Use full lattice only as an upper result because it includes expert proposals.",
     "Do not claim external SOTA from this table alone; add current external baselines separately.",
-    "Current best clean deployment row is K4 mean-by-type residual retrieval, scales 0.35/0.40/0.45, margin 0.20, no-op bonus 0.03 at 35.42%.",
+    "Current best clean deployment row is K4 composed type-consensus residual retrieval, masked, scales 0.35/0.40/0.45, margin 0.20, no-op bonus 0.03 at 35.54%.",
     "Trust-region field optimization should be framed as a negative/diagnostic ablation.",
     "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

@@ -69,6 +69,8 @@ Baseline h=16 policy: 29.74%
 | retrieval_residual_k4_mean_grid035040045_noopbonus003_consensus010 | K4 mean-by-type residual retrieval, scales 0.35/0.40/0.45, margin 0.20, no-op bonus 0.03, consensus penalty 0.10 | complete | 35.36% | +5.62 pp | yes | no | no | train-neighbor tangent-consensus confidence on the current best typed prior |
 | retrieval_residual_k4_compose_grid035040045 | K4 composed type-consensus residual retrieval, scales 0.35/0.40/0.45, margin 0.20 | complete | 34.09% | +4.35 pp | yes | no | no | local tangent composition without typed priors |
 | retrieval_residual_k4_compose_grid035040045_noopbonus003 | K4 composed type-consensus residual retrieval, scales 0.35/0.40/0.45, margin 0.20, no-op bonus 0.03 | complete | 34.14% | +4.41 pp | yes | no | no | local tangent composition on the current best typed prior |
+| 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_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 |
@@ -130,7 +132,7 @@ Baseline h=16 policy: 29.74%
 - Use no-expert same-state lattice as the conservative mechanism result, not as deployment-clean inference.
 - Use full lattice only as an upper result because it includes expert proposals.
 - Do not claim external SOTA from this table alone; add current external baselines separately.
-- Current best clean deployment row is K4 mean-by-type residual retrieval, scales 0.35/0.40/0.45, margin 0.20, no-op bonus 0.03 at 35.42%.
+- Current best clean deployment row is K4 composed type-consensus residual retrieval, masked, scales 0.35/0.40/0.45, margin 0.20, no-op bonus 0.03 at 35.54%.
 - Trust-region field optimization should be framed as a negative/diagnostic ablation.
 - 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_analysis.py

@@ -14,7 +14,13 @@ RESULTS_DIR = Path("results")
 OUT_JSON = RESULTS_DIR / "paper_analysis.json"
 OUT_MD = RESULTS_DIR / "paper_analysis.md"
 CANONICAL_H16_ROLLOUT = Path("/scratch/knguy52/dovla/experiments/dovla_h16_rollout_runs")
-BEST_CLEAN_KEY = "residual_k4_consensus_grid035040045_noopbonus003"
+FALLBACK_BEST_CLEAN_KEY = "residual_k4_consensus_grid035040045_noopbonus003"
+NON_DEPLOYMENT_KEYS = {
+    "same_state_near_miss",
+    "same_state_no_expert",
+    "same_state_policy_baseline",
+    "same_state_full",
+}
 
 
 @dataclass(frozen=True)
@@ -807,6 +813,22 @@ def _load_methods() -> dict[str, dict[str, Any]]:
     return methods
 
 
+def _best_clean_key(methods: dict[str, dict[str, Any]]) -> str:
+    candidates: list[tuple[float, str]] = []
+    for spec in METHODS:
+        if spec.key in NON_DEPLOYMENT_KEYS:
+            continue
+        method = methods.get(spec.key, {})
+        if method.get("missing"):
+            continue
+        success = method.get("mean_success")
+        if isinstance(success, (int, float)) and math.isfinite(float(success)):
+            candidates.append((float(success), spec.key))
+    if not candidates:
+        return FALLBACK_BEST_CLEAN_KEY
+    return max(candidates)[1]
+
+
 def _paired_delta(
     methods: dict[str, dict[str, Any]],
     left: str,
@@ -854,6 +876,7 @@ def _pp(value: float) -> str:
 
 def _render_markdown(report: dict[str, Any]) -> str:
     methods = report["methods"]
+    best_clean_key = report["best_clean_key"]
     lines = [
         "# Paper Analysis",
         "",
@@ -911,7 +934,7 @@ def _render_markdown(report: dict[str, Any]) -> str:
     tasks = sorted(methods["h16_policy_canonical"].get("per_task_success", {}))
     for task in tasks:
         h16 = methods["h16_policy_canonical"]["per_task_success"][task]["mean_success"]
-        clean = methods[BEST_CLEAN_KEY]["per_task_success"][task]["mean_success"]
+        clean = methods[best_clean_key]["per_task_success"][task]["mean_success"]
         near = methods["same_state_near_miss"]["per_task_success"][task]["mean_success"]
         noexpert = methods["same_state_no_expert"]["per_task_success"][task]["mean_success"]
         full = methods["same_state_full"]["per_task_success"][task]["mean_success"]
@@ -933,7 +956,7 @@ def _render_markdown(report: dict[str, Any]) -> str:
             "",
         ]
     )
-    for key in ["same_state_near_miss", "same_state_no_expert", "same_state_policy_baseline", "same_state_full", BEST_CLEAN_KEY]:
+    for key in ["same_state_near_miss", "same_state_no_expert", "same_state_policy_baseline", "same_state_full", best_clean_key]:
         counts = methods[key].get("selected_candidate_type_counts", {})
         if counts:
             total = sum(int(value) for value in counts.values())
@@ -944,9 +967,9 @@ def _render_markdown(report: dict[str, Any]) -> str:
         else:
             summary = "not recorded"
         lines.append(f"- `{key}`: {summary}")
-    scale_counts = methods[BEST_CLEAN_KEY].get("selected_residual_scale_counts", {})
+    scale_counts = methods[best_clean_key].get("selected_residual_scale_counts", {})
     if scale_counts:
-        lines.append(f"- `{BEST_CLEAN_KEY}` residual scale counts: {scale_counts}")
+        lines.append(f"- `{best_clean_key}` residual scale counts: {scale_counts}")
     lines.extend(
         [
             "",
@@ -994,9 +1017,10 @@ def _render_markdown(report: dict[str, Any]) -> str:
 
 def build_report() -> dict[str, Any]:
     methods = _load_methods()
+    best_clean_key = _best_clean_key(methods)
     paired_deltas = {
-        "best_clean - canonical_h16": _paired_delta(methods, BEST_CLEAN_KEY, "h16_policy_canonical"),
-        "best_clean - direct_same_ckpt": _paired_delta(methods, BEST_CLEAN_KEY, "near_miss_policy_bc5"),
+        "best_clean - canonical_h16": _paired_delta(methods, best_clean_key, "h16_policy_canonical"),
+        "best_clean - direct_same_ckpt": _paired_delta(methods, best_clean_key, "near_miss_policy_bc5"),
         "no_expert_lattice - canonical_h16": _paired_delta(methods, "same_state_no_expert", "h16_policy_canonical"),
         "full_lattice - no_expert_lattice": _paired_delta(methods, "same_state_full", "same_state_no_expert"),
         "policy_candidate_lattice - no_expert_lattice": _paired_delta(
@@ -1006,14 +1030,14 @@ def build_report() -> dict[str, Any]:
         ),
     }
     mechanism_gap = {
-        "best_clean_vs_h16": methods[BEST_CLEAN_KEY]["mean_success"]
+        "best_clean_vs_h16": methods[best_clean_key]["mean_success"]
         - methods["h16_policy_canonical"]["mean_success"],
-        "best_clean_vs_direct_same_ckpt": methods[BEST_CLEAN_KEY]["mean_success"]
+        "best_clean_vs_direct_same_ckpt": methods[best_clean_key]["mean_success"]
         - methods["near_miss_policy_bc5"]["mean_success"],
         "same_state_no_expert_vs_h16": methods["same_state_no_expert"]["mean_success"]
         - methods["h16_policy_canonical"]["mean_success"],
         "same_state_no_expert_vs_best_clean": methods["same_state_no_expert"]["mean_success"]
-        - methods[BEST_CLEAN_KEY]["mean_success"],
+        - methods[best_clean_key]["mean_success"],
         "same_state_full_vs_no_expert": methods["same_state_full"]["mean_success"]
         - methods["same_state_no_expert"]["mean_success"],
     }
@@ -1022,11 +1046,11 @@ def build_report() -> dict[str, Any]:
         "methods": methods,
         "paired_deltas": paired_deltas,
         "per_task_deltas": {
-            "best_clean_vs_h16": _per_task_delta(methods, BEST_CLEAN_KEY, "h16_policy_canonical"),
-            "no_expert_vs_best_clean": _per_task_delta(methods, "same_state_no_expert", BEST_CLEAN_KEY),
+            "best_clean_vs_h16": _per_task_delta(methods, best_clean_key, "h16_policy_canonical"),
+            "no_expert_vs_best_clean": _per_task_delta(methods, "same_state_no_expert", best_clean_key),
         },
         "mechanism_gap": mechanism_gap,
-        "best_clean_key": BEST_CLEAN_KEY,
+        "best_clean_key": best_clean_key,
     }