Auto-sync: 2026-06-30 03:41:59 (part 2)

results/paper_analysis.json

@@ -1,11 +1,11 @@
 {
-  "best_clean_key": "residual_k4_composemasked_dropnmnoop_grid035040045_noopbonus003_oraclek8",
-  "generated_utc": "2026-06-30T06:57:02+00:00",
+  "best_clean_key": "residual_k4_composemasked_dropnmnoop_grid035040045_noopbonus003_nmchallenger002",
+  "generated_utc": "2026-06-30T07:49:50+00:00",
   "mechanism_gap": {
-    "best_clean_vs_direct_same_ckpt": 0.07362318840579707,
-    "best_clean_vs_h16": 0.05913043478260871,
+    "best_clean_vs_direct_same_ckpt": 0.07710144927536233,
+    "best_clean_vs_h16": 0.06260869565217397,
     "same_state_full_vs_no_expert": 0.12347826086956515,
-    "same_state_no_expert_vs_best_clean": 0.21333333333333343,
+    "same_state_no_expert_vs_best_clean": 0.20985507246376817,
     "same_state_no_expert_vs_h16": 0.27246376811594214
   },
   "methods": {
@@ -1432,6 +1432,103 @@
       "source": "results/h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_k4_composemasked_dropnmnoop_grid035040045_safe_margin0p20_noopbonus0p03_nmbonus0p02_summary.json",
       "std_success": 0.009889114266221456
     },
+    "residual_k4_composemasked_dropnmnoop_grid035040045_noopbonus003_nmchallenger002": {
+      "ci95_success": 0.03115620450679224,
+      "label": "K4 compatible tangents, no-op bonus 0.03, near-miss challenger gate 0.02",
+      "mean_action_mse_to_best": 0.40733562724178896,
+      "mean_progress": 0.5741919005876812,
+      "mean_success": 0.36000000000000004,
+      "num_completed": 3,
+      "per_task_success": {
+        "LiftPegUpright-v1": {
+          "mean_num_groups": 102.0,
+          "mean_success": 0.26340865087329823,
+          "std_success": 0.06057881385012102
+        },
+        "PickCube-v1": {
+          "mean_num_groups": 196.66666666666666,
+          "mean_success": 0.3294194734412126,
+          "std_success": 0.02715752817589447
+        },
+        "PullCube-v1": {
+          "mean_num_groups": 81.0,
+          "mean_success": 0.22889673679147363,
+          "std_success": 0.029397613100483843
+        },
+        "PushCube-v1": {
+          "mean_num_groups": 102.0,
+          "mean_success": 0.7571578319419062,
+          "std_success": 0.0856403530080274
+        },
+        "StackCube-v1": {
+          "mean_num_groups": 93.33333333333333,
+          "mean_success": 0.20044030044030045,
+          "std_success": 0.08896982358351117
+        }
+      },
+      "seed_action_mse_to_best": {
+        "0": 0.3948699296703157,
+        "1": 0.39924815376005746,
+        "2": 0.4278887982949938
+      },
+      "seed_progress": {
+        "0": 0.5602596596268046,
+        "1": 0.5735099030577618,
+        "2": 0.5888061390784771
+      },
+      "seed_success": {
+        "0": 0.3565217391304348,
+        "1": 0.34956521739130436,
+        "2": 0.3739130434782609
+      },
+      "selected_candidate_type_counts": {
+        "retrieval_residual_policy_residual": 1556,
+        "retrieval_residual_residual_near_miss": 65,
+        "retrieval_residual_residual_near_miss+residual_wrong_gripper": 5,
+        "retrieval_residual_residual_no_op": 56,
+        "retrieval_residual_residual_no_op+residual_wrong_gripper": 24,
+        "retrieval_residual_residual_wrong_gripper": 19
+      },
+      "selected_residual_scale_counts": {
+        "0.35": 1587,
+        "0.4": 16,
+        "0.45": 122
+      },
+      "selected_type_outcomes": {
+        "retrieval_residual_policy_residual": {
+          "count": 1556.0,
+          "mean_progress": 0.565076002276384,
+          "success_rate": 0.34832904884318766
+        },
+        "retrieval_residual_residual_near_miss": {
+          "count": 65.0,
+          "mean_progress": 0.5945549370577702,
+          "success_rate": 0.4307692307692308
+        },
+        "retrieval_residual_residual_near_miss+residual_wrong_gripper": {
+          "count": 5.0,
+          "mean_progress": 0.7557869911193847,
+          "success_rate": 0.6
+        },
+        "retrieval_residual_residual_no_op": {
+          "count": 56.0,
+          "mean_progress": 0.7249722566921264,
+          "success_rate": 0.5
+        },
+        "retrieval_residual_residual_no_op+residual_wrong_gripper": {
+          "count": 24.0,
+          "mean_progress": 0.6711457151298722,
+          "success_rate": 0.4583333333333333
+        },
+        "retrieval_residual_residual_wrong_gripper": {
+          "count": 19.0,
+          "mean_progress": 0.6364115562878156,
+          "success_rate": 0.47368421052631576
+        }
+      },
+      "source": "results/h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_k4_composemasked_dropnmnoop_grid035040045_safe_margin0p20_noopbonus0p03_nmchallenger0p02_summary.json",
+      "std_success": 0.012541047914657353
+    },
     "residual_k4_composemasked_dropnmnoop_grid035040045_noopbonus003_oraclek8": {
       "candidate_oracle_rollouts": 8,
       "candidate_oracle_type_counts": {
@@ -1544,6 +1641,159 @@
       "source": "results/h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_k4_composemasked_dropnmnoop_grid035040045_safe_margin0p20_noopbonus0p03_oraclek8_summary.json",
       "std_success": 0.012173913043478276
     },
+    "residual_k4_composemasked_dropnmnoop_grid035040045_noopbonus003_oraclek8trace": {
+      "candidate_oracle_best_branch_rank_counts": {
+        "1": 931,
+        "2": 162,
+        "3": 88,
+        "4": 87,
+        "5": 102,
+        "6": 87,
+        "7": 119,
+        "8": 149
+      },
+      "candidate_oracle_rollouts": 8,
+      "candidate_oracle_type_counts": {
+        "retrieval_residual_policy_residual": 867,
+        "retrieval_residual_residual_near_miss": 201,
+        "retrieval_residual_residual_near_miss+residual_wrong_gripper": 99,
+        "retrieval_residual_residual_no_op": 182,
+        "retrieval_residual_residual_no_op+residual_wrong_gripper": 86,
+        "retrieval_residual_residual_wrong_gripper": 290
+      },
+      "candidate_oracle_unique_tolerance": 1e-06,
+      "ci95_success": 0.03024411731871433,
+      "label": "K4 compatible tangents, no-op bonus 0.03, unique candidate-oracle prefix K=8 branch trace",
+      "mean_action_mse_to_best": 0.4063519326314006,
+      "mean_candidate_oracle_best_branch_rank": 2.8544927536231883,
+      "mean_candidate_oracle_branch_progress": [
+        0.5663117068376537,
+        0.520170341766021,
+        0.4982426099374931,
+        0.48436801132148777,
+        0.4810738827217964,
+        0.4725861245269809,
+        0.45777137219419944,
+        0.4651103241433016
+      ],
+      "mean_candidate_oracle_branch_score_gains_over_selected": [
+        0.0,
+        -0.10121382852732812,
+        -0.139953154850299,
+        -0.16368282582816682,
+        -0.17625231607560662,
+        -0.1911168860916518,
+        -0.22332294284109622,
+        -0.19917239659924763
+      ],
+      "mean_candidate_oracle_branch_success_rates": [
+        0.34840579710144925,
+        0.29333333333333333,
+        0.27652173913043476,
+        0.26666666666666666,
+        0.2573913043478261,
+        0.2510144927536232,
+        0.2336231884057971,
+        0.25043478260869567
+      ],
+      "mean_candidate_oracle_improvement_rate": 0.4602898550724637,
+      "mean_candidate_oracle_progress": 0.6443418554713328,
+      "mean_candidate_oracle_score_gain_over_selected": 0.16034898876064066,
+      "mean_candidate_oracle_success_rate": 0.4307246376811594,
+      "mean_candidate_oracle_unique_count": 8.0,
+      "mean_progress": 0.571575072284851,
+      "mean_success": 0.3565217391304348,
+      "num_completed": 3,
+      "per_task_success": {
+        "LiftPegUpright-v1": {
+          "mean_num_groups": 102.0,
+          "mean_success": 0.26340865087329823,
+          "std_success": 0.06057881385012102
+        },
+        "PickCube-v1": {
+          "mean_num_groups": 196.66666666666666,
+          "mean_success": 0.3278169093386485,
+          "std_success": 0.02533117351097172
+        },
+        "PullCube-v1": {
+          "mean_num_groups": 81.0,
+          "mean_success": 0.21716032505506191,
+          "std_success": 0.0182535211270701
+        },
+        "PushCube-v1": {
+          "mean_num_groups": 102.0,
+          "mean_success": 0.7541548289389031,
+          "std_success": 0.08244271787813129
+        },
+        "StackCube-v1": {
+          "mean_num_groups": 93.33333333333333,
+          "mean_success": 0.19707329707329704,
+          "std_success": 0.0831467196651622
+        }
+      },
+      "seed_action_mse_to_best": {
+        "0": 0.3930192190424904,
+        "1": 0.39824521353549286,
+        "2": 0.42779136531621864
+      },
+      "seed_progress": {
+        "0": 0.5579317773906919,
+        "1": 0.571580135051323,
+        "2": 0.5852133044125383
+      },
+      "seed_success": {
+        "0": 0.35130434782608694,
+        "1": 0.34782608695652173,
+        "2": 0.37043478260869567
+      },
+      "selected_candidate_type_counts": {
+        "retrieval_residual_policy_residual": 1613,
+        "retrieval_residual_residual_near_miss": 8,
+        "retrieval_residual_residual_near_miss+residual_wrong_gripper": 5,
+        "retrieval_residual_residual_no_op": 56,
+        "retrieval_residual_residual_no_op+residual_wrong_gripper": 24,
+        "retrieval_residual_residual_wrong_gripper": 19
+      },
+      "selected_residual_scale_counts": {
+        "0.35": 1631,
+        "0.4": 11,
+        "0.45": 83
+      },
+      "selected_type_outcomes": {
+        "retrieval_residual_policy_residual": {
+          "count": 1613.0,
+          "mean_progress": 0.561925975703087,
+          "success_rate": 0.34593924364538126
+        },
+        "retrieval_residual_residual_near_miss": {
+          "count": 8.0,
+          "mean_progress": 0.8754628524184227,
+          "success_rate": 0.75
+        },
+        "retrieval_residual_residual_near_miss+residual_wrong_gripper": {
+          "count": 5.0,
+          "mean_progress": 0.7557869911193847,
+          "success_rate": 0.6
+        },
+        "retrieval_residual_residual_no_op": {
+          "count": 56.0,
+          "mean_progress": 0.7249722566921264,
+          "success_rate": 0.5
+        },
+        "retrieval_residual_residual_no_op+residual_wrong_gripper": {
+          "count": 24.0,
+          "mean_progress": 0.6711457151298722,
+          "success_rate": 0.4583333333333333
+        },
+        "retrieval_residual_residual_wrong_gripper": {
+          "count": 19.0,
+          "mean_progress": 0.6364115562878156,
+          "success_rate": 0.47368421052631576
+        }
+      },
+      "source": "results/h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_k4_composemasked_dropnmnoop_grid035040045_safe_margin0p20_noopbonus0p03_oraclek8trace_summary.json",
+      "std_success": 0.012173913043478276
+    },
     "residual_k4_composemasked_dropnmnoop_grid035040045_noopbonus003_srcscorebonus002": {
       "ci95_success": 0.02131295542962397,
       "label": "K4 composed type-consensus tangents, masked, drop near-miss+no-op composite, no-op bonus 0.03, source-score bonus 0.02",
@@ -7204,38 +7454,38 @@
   },
   "paired_deltas": {
     "best_clean - canonical_h16": {
-      "ci95_delta": 0.04510826498554671,
-      "left": "residual_k4_composemasked_dropnmnoop_grid035040045_noopbonus003_oraclek8",
-      "mean_delta": 0.05913043478260869,
+      "ci95_delta": 0.048690556079097516,
+      "left": "residual_k4_composemasked_dropnmnoop_grid035040045_noopbonus003_nmchallenger002",
+      "mean_delta": 0.06260869565217392,
       "right": "h16_policy_canonical",
       "seed_deltas": {
-        "0": 0.0678260869565217,
-        "1": 0.038260869565217404,
-        "2": 0.07130434782608697
+        "0": 0.07304347826086954,
+        "1": 0.040000000000000036,
+        "2": 0.07478260869565218
       },
       "seeds": [
         0,
         1,
         2
       ],
-      "std_delta": 0.018157054798105293
+      "std_delta": 0.01959900464275588
     },
     "best_clean - direct_same_ckpt": {
-      "ci95_delta": 0.045383317930687055,
-      "left": "residual_k4_composemasked_dropnmnoop_grid035040045_noopbonus003_oraclek8",
-      "mean_delta": 0.07362318840579708,
+      "ci95_delta": 0.047591888506633255,
+      "left": "residual_k4_composemasked_dropnmnoop_grid035040045_noopbonus003_nmchallenger002",
+      "mean_delta": 0.07710144927536232,
       "right": "near_miss_policy_bc5",
       "seed_deltas": {
-        "0": 0.07304347826086954,
-        "1": 0.055652173913043446,
-        "2": 0.09217391304347827
+        "0": 0.07826086956521738,
+        "1": 0.05739130434782608,
+        "2": 0.09565217391304348
       },
       "seeds": [
         0,
         1,
         2
       ],
-      "std_delta": 0.018267769572856522
+      "std_delta": 0.019156767120995143
     },
     "full_lattice - no_expert_lattice": {
       "ci95_delta": 0.02628111194117426,
@@ -7292,17 +7542,17 @@
   "per_task_deltas": {
     "best_clean_vs_h16": {
       "LiftPegUpright-v1": 0.03278131303915899,
-      "PickCube-v1": 0.1219595621769535,
-      "PullCube-v1": 0.019305840358471965,
-      "PushCube-v1": 0.003597657063003523,
-      "StackCube-v1": 0.05610685610685609
+      "PickCube-v1": 0.1235621262795176,
+      "PullCube-v1": 0.031042252094883677,
+      "PushCube-v1": 0.0066006600660066805,
+      "StackCube-v1": 0.0594738594738595
     },
     "no_expert_vs_best_clean": {
       "LiftPegUpright-v1": 0.35374851747103375,
-      "PickCube-v1": 0.29366195398804096,
-      "PullCube-v1": -0.018680539733171325,
-      "PushCube-v1": 0.04688815632569587,
-      "StackCube-v1": 0.2854441854441855
+      "PickCube-v1": 0.29205938988547686,
+      "PullCube-v1": -0.030416951469583037,
+      "PushCube-v1": 0.043885153322692716,
+      "StackCube-v1": 0.282077182077182
     }
   }
 }

results/paper_analysis.md

@@ -1,6 +1,6 @@
 # Paper Analysis
 
-Generated: `2026-06-30T06:57:02+00:00`
+Generated: `2026-06-30T07:49:50+00:00`
 
 ## Main Seed Statistics
 
@@ -51,6 +51,8 @@ Generated: `2026-06-30T06:57:02+00:00`
 | residual_k4_composemasked_dropnmnoop_grid035040045_noopbonus003 | K4 composed type-consensus tangents, masked, drop near-miss+no-op composite | 3 | 35.59% +/- 0.99 | +/- 2.46 | 57.07% | 0.406 | +5.86 pp |
 | residual_k4_composemasked_dropnmnoop_grid035040045_margin010_noopbonus003 | K4 compatible tangents, margin 0.10, no-op bonus 0.03 | 3 | 34.67% +/- 0.96 | +/- 2.38 | 56.45% | 0.418 | +4.93 pp |
 | residual_k4_composemasked_dropnmnoop_grid035040045_noopbonus003_oraclek8 | K4 compatible tangents, no-op bonus 0.03, unique candidate-oracle prefix K=8 diagnostic | 3 | 35.65% +/- 1.22 | +/- 3.02 | 57.16% | 0.406 | +5.91 pp |
+| residual_k4_composemasked_dropnmnoop_grid035040045_noopbonus003_oraclek8trace | K4 compatible tangents, no-op bonus 0.03, unique candidate-oracle prefix K=8 branch trace | 3 | 35.65% +/- 1.22 | +/- 3.02 | 57.16% | 0.406 | +5.91 pp |
+| residual_k4_composemasked_dropnmnoop_grid035040045_noopbonus003_nmchallenger002 | K4 compatible tangents, no-op bonus 0.03, near-miss challenger gate 0.02 | 3 | 36.00% +/- 1.25 | +/- 3.12 | 57.42% | 0.407 | +6.26 pp |
 | residual_k4_composemasked_dropnmnoop_grid035040045_noopbonus003_nmbonus001 | K4 composed compatible tangents, no-op bonus 0.03, singleton near-miss bonus 0.01 | 3 | 35.59% +/- 0.99 | +/- 2.46 | 57.10% | 0.406 | +5.86 pp |
 | residual_k4_composemasked_dropnmnoop_grid035040045_noopbonus003_nmbonus002 | K4 composed compatible tangents, no-op bonus 0.03, singleton near-miss bonus 0.02 | 3 | 35.59% +/- 0.99 | +/- 2.46 | 57.10% | 0.406 | +5.86 pp |
 | residual_k4_composemasked_dropnmnoop_grid035040045_srcscorebonus002 | K4 composed type-consensus tangents, masked, drop near-miss+no-op composite, source-score bonus 0.02 | 3 | 35.48% +/- 1.22 | +/- 3.02 | 57.02% | 0.408 | +5.74 pp |
@@ -98,8 +100,8 @@ Generated: `2026-06-30T06:57:02+00:00`
 
 | comparison | seeds | mean delta | 95% CI | seed deltas |
 |---|---:|---:|---:|---|
-| best_clean - canonical_h16 | 3 | +5.91 pp | +/- 4.51 | 0:+6.78, 1:+3.83, 2:+7.13 |
-| best_clean - direct_same_ckpt | 3 | +7.36 pp | +/- 4.54 | 0:+7.30, 1:+5.57, 2:+9.22 |
+| best_clean - canonical_h16 | 3 | +6.26 pp | +/- 4.87 | 0:+7.30, 1:+4.00, 2:+7.48 |
+| best_clean - direct_same_ckpt | 3 | +7.71 pp | +/- 4.76 | 0:+7.83, 1:+5.74, 2:+9.57 |
 | 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 |
@@ -109,24 +111,27 @@ Generated: `2026-06-30T06:57:02+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% | 32.78% | 58.13% | 62.15% | 84.19% | +12.20 pp | +29.37 pp |
-| PullCube-v1 | 19.79% | 21.72% | 15.02% | 19.85% | 22.41% | +1.93 pp | -1.87 pp |
-| PushCube-v1 | 75.06% | 75.42% | 82.54% | 80.10% | 81.92% | +0.36 pp | +4.69 pp |
-| StackCube-v1 | 14.10% | 19.71% | 50.41% | 48.25% | 60.83% | +5.61 pp | +28.54 pp |
+| PickCube-v1 | 20.59% | 32.94% | 58.13% | 62.15% | 84.19% | +12.36 pp | +29.21 pp |
+| PullCube-v1 | 19.79% | 22.89% | 15.02% | 19.85% | 22.41% | +3.10 pp | -3.04 pp |
+| PushCube-v1 | 75.06% | 75.72% | 82.54% | 80.10% | 81.92% | +0.66 pp | +4.39 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 +5.91 pp.
+- Best clean residual transport improves over canonical h16 by +6.26 pp.
 - Same-state no-expert lattice improves over canonical h16 by +27.25 pp.
-- Remaining clean-to-same-state proposal gap is +21.33 pp.
+- Remaining clean-to-same-state proposal gap is +20.99 pp.
 - Full lattice adds expert proposals and reaches 69.33%, a +12.35 pp gain over no-expert.
 
 ## Candidate-Oracle Diagnostic
 
 - Oracle over the deployed candidate prefix reaches 43.07% with mean progress 64.43%; this is diagnostic-only because it uses measured rollout outcomes after generating candidates.
 - Mean oracle-prefix score gain over the selected branch is +0.160, which isolates ranking/abstention headroom inside the clean proposal set.
-- Mean unique candidates in the prefix: nan.
+- Mean unique candidates in the prefix: 8.00.
 - Candidate-oracle best type counts: {'retrieval_residual_policy_residual': 867, 'retrieval_residual_residual_near_miss': 201, 'retrieval_residual_residual_near_miss+residual_wrong_gripper': 99, 'retrieval_residual_residual_no_op': 182, 'retrieval_residual_residual_no_op+residual_wrong_gripper': 86, 'retrieval_residual_residual_wrong_gripper': 290}.
+- Mean best branch rank in the field-ordered prefix: 2.85; rank histogram {'1': 931, '2': 162, '3': 88, '4': 87, '5': 102, '6': 87, '7': 119, '8': 149}.
+- Branch success by prefix rank: 34.84%, 29.33%, 27.65%, 26.67%, 25.74%, 25.10%, 23.36%, 25.04%.
+- Branch score gain by prefix rank: +0.000, -0.101, -0.140, -0.164, -0.176, -0.191, -0.223, -0.199.
 
 ## Selection Histograms
 
@@ -134,8 +139,8 @@ Generated: `2026-06-30T06:57:02+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_composemasked_dropnmnoop_grid035040045_noopbonus003_oraclek8`: retrieval_residual_policy_residual=1613 (93.5%), retrieval_residual_residual_no_op=56 (3.2%), retrieval_residual_residual_no_op+residual_wrong_gripper=24 (1.4%), retrieval_residual_residual_wrong_gripper=19 (1.1%), retrieval_residual_residual_near_miss=8 (0.5%), retrieval_residual_residual_near_miss+residual_wrong_gripper=5 (0.3%)
-- `residual_k4_composemasked_dropnmnoop_grid035040045_noopbonus003_oraclek8` residual scale counts: {'0.35': 1631, '0.4': 11, '0.45': 83}
+- `residual_k4_composemasked_dropnmnoop_grid035040045_noopbonus003_nmchallenger002`: retrieval_residual_policy_residual=1556 (90.2%), retrieval_residual_residual_near_miss=65 (3.8%), retrieval_residual_residual_no_op=56 (3.2%), retrieval_residual_residual_no_op+residual_wrong_gripper=24 (1.4%), retrieval_residual_residual_wrong_gripper=19 (1.1%), retrieval_residual_residual_near_miss+residual_wrong_gripper=5 (0.3%)
+- `residual_k4_composemasked_dropnmnoop_grid035040045_noopbonus003_nmchallenger002` residual scale counts: {'0.35': 1587, '0.4': 16, '0.45': 122}
 
 ## Selected-Type Outcomes
 

results/paper_table_status.json

@@ -1372,6 +1372,44 @@
       "best_config": null,
       "gain_vs_h16_policy": 0.13333333333333336
     },
+    {
+      "key": "retrieval_residual_k4_composemasked_dropnmnoop_grid035040045_noopbonus003_oraclek8trace",
+      "label": "K4 compatible residual retrieval, unique candidate-oracle prefix K=8 branch trace",
+      "path": "h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_k4_composemasked_dropnmnoop_grid035040045_safe_margin0p20_noopbonus0p03_oraclek8trace_summary.json",
+      "clean_deployment": "diagnostic",
+      "same_state_proposals": "no",
+      "expert_proposal": "no",
+      "story_role": "diagnostic branch trace for deployable selector calibration over the local tangent chart",
+      "fallback_success": null,
+      "pending_job": "14953960/14953961",
+      "path_exists": true,
+      "status": "complete",
+      "success": 0.43072463768115943,
+      "std_success": 0.012173913043478276,
+      "completed_seeds": null,
+      "num_completed": 3,
+      "best_config": null,
+      "gain_vs_h16_policy": 0.13333333333333336
+    },
+    {
+      "key": "retrieval_residual_k4_composemasked_dropnmnoop_grid035040045_noopbonus003_nmchallenger002",
+      "label": "K4 compatible residual retrieval, near-miss challenger gate 0.02",
+      "path": "h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_k4_composemasked_dropnmnoop_grid035040045_safe_margin0p20_noopbonus0p03_nmchallenger0p02_summary.json",
+      "clean_deployment": "yes",
+      "same_state_proposals": "no",
+      "expert_proposal": "no",
+      "story_role": "trace-motivated two-stage selector calibration over the compatible local tangent chart",
+      "fallback_success": null,
+      "pending_job": "14954280/14954281",
+      "path_exists": true,
+      "status": "complete",
+      "success": 0.36,
+      "std_success": 0.012541047914657353,
+      "completed_seeds": null,
+      "num_completed": 3,
+      "best_config": null,
+      "gain_vs_h16_policy": 0.06260869565217392
+    },
     {
       "key": "retrieval_residual_k4_composemasked_dropnmnoop_grid035040045_noopbonus003_nmbonus001",
       "label": "K4 composed compatible residual retrieval, no-op bonus 0.03, singleton near-miss bonus 0.01",
@@ -2571,23 +2609,23 @@
     }
   ],
   "best_clean": {
-    "key": "retrieval_residual_k4_composemasked_dropnmnoop_grid035040045_noopbonus003",
-    "label": "K4 composed type-consensus residual retrieval, masked, drop near-miss+no-op composite",
-    "path": "h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_k4_composemasked_dropnmnoop_grid035040045_safe_margin0p20_noopbonus0p03_summary.json",
+    "key": "retrieval_residual_k4_composemasked_dropnmnoop_grid035040045_noopbonus003_nmchallenger002",
+    "label": "K4 compatible residual retrieval, near-miss challenger gate 0.02",
+    "path": "h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_k4_composemasked_dropnmnoop_grid035040045_safe_margin0p20_noopbonus0p03_nmchallenger0p02_summary.json",
     "clean_deployment": "yes",
     "same_state_proposals": "no",
     "expert_proposal": "no",
-    "story_role": "exact incompatibility mask for one weak composed tangent pair",
+    "story_role": "trace-motivated two-stage selector calibration over the compatible local tangent chart",
     "fallback_success": null,
-    "pending_job": "14915009/14915013",
+    "pending_job": "14954280/14954281",
     "path_exists": true,
     "status": "complete",
-    "success": 0.35594202898550725,
-    "std_success": 0.009889114266221456,
+    "success": 0.36,
+    "std_success": 0.012541047914657353,
     "completed_seeds": null,
     "num_completed": 3,
     "best_config": null,
-    "gain_vs_h16_policy": 0.058550724637681184
+    "gain_vs_h16_policy": 0.06260869565217392
   },
   "best_mechanism_no_expert": {
     "key": "no_expert_lattice",
@@ -2612,13 +2650,13 @@
     "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 composed type-consensus residual retrieval, masked, drop near-miss+no-op composite at 35.59%.",
+    "Current best clean deployment row is K4 compatible residual retrieval, near-miss challenger gate 0.02 at 36.00%.",
     "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.",
     "K4 train-state residual retrieval, scale 0.40, safe residuals, mean-by-type tangent consensus improves the clean bridge but is not yet the main result.",
     "Policy-baseline abstention inside the same-state no-expert lattice drops from 56.99% to 40.70%, so the mechanism result should emphasize counterfactual proposal geometry rather than policy fallback.",
-    "Singleton near-miss priors 0.01/0.02 tie the clean best rather than improving it; the compatible chart preserves high-precision near-miss tangents, but they remain sparse.",
-    "Candidate-oracle prefix K=8 is diagnostic-only; compare it to the clean selected row to decide whether the next paper move should be proposal generation or field ranking."
+    "Candidate-oracle prefix K=8 is diagnostic-only; compare it to the clean selected row to decide whether the next paper move should be proposal generation or field ranking.",
+    "Candidate-oracle branch trace is diagnostic-only; use its rank histogram and branch success profile to design one clean selector-calibration run rather than a broad sweep."
   ]
 }

results/paper_table_status.md

@@ -75,6 +75,8 @@ Baseline h=16 policy: 29.74%
 | retrieval_residual_k4_composemasked_dropnmnoop_grid035040045_noopbonus003 | K4 composed type-consensus residual retrieval, masked, drop near-miss+no-op composite | complete | 35.59% | +5.86 pp | yes | no | no | exact incompatibility mask for one weak composed tangent pair |
 | retrieval_residual_k4_composemasked_dropnmnoop_grid035040045_margin010_noopbonus003 | K4 compatible residual retrieval, margin 0.10, no-op bonus 0.03 | complete | 34.67% | +4.93 pp | yes | no | no | oracle-motivated abstention sensitivity on the compatible local tangent chart |
 | retrieval_residual_k4_composemasked_dropnmnoop_grid035040045_noopbonus003_oraclek8 | K4 compatible residual retrieval, unique candidate-oracle prefix K=8 diagnostic | complete | 43.07% | +13.33 pp | diagnostic | no | no | diagnostic oracle over generated clean candidate prefix to separate proposal headroom from field ranking |
+| retrieval_residual_k4_composemasked_dropnmnoop_grid035040045_noopbonus003_oraclek8trace | K4 compatible residual retrieval, unique candidate-oracle prefix K=8 branch trace | complete | 43.07% | +13.33 pp | diagnostic | no | no | diagnostic branch trace for deployable selector calibration over the local tangent chart |
+| retrieval_residual_k4_composemasked_dropnmnoop_grid035040045_noopbonus003_nmchallenger002 | K4 compatible residual retrieval, near-miss challenger gate 0.02 | complete | 36.00% | +6.26 pp | yes | no | no | trace-motivated two-stage selector calibration over the compatible local tangent chart |
 | retrieval_residual_k4_composemasked_dropnmnoop_grid035040045_noopbonus003_nmbonus001 | K4 composed compatible residual retrieval, no-op bonus 0.03, singleton near-miss bonus 0.01 | complete | 35.59% | +5.86 pp | yes | no | no | revive high-precision singleton near-miss tangents without boosting toxic near-miss+no-op composites |
 | retrieval_residual_k4_composemasked_dropnmnoop_grid035040045_noopbonus003_nmbonus002 | K4 composed compatible residual retrieval, no-op bonus 0.03, singleton near-miss bonus 0.02 | complete | 35.59% | +5.86 pp | yes | no | no | stronger singleton near-miss revival prior on the compatible local tangent chart |
 | retrieval_residual_k4_composemasked_dropnmnoop_grid035040045_srcscorebonus002 | K4 composed type-consensus residual retrieval, masked, drop near-miss+no-op composite, source-score bonus 0.02 | complete | 35.48% | +5.74 pp | yes | no | no | train-measured source-score prior on the compatible local tangent chart |
@@ -144,11 +146,11 @@ 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 composed type-consensus residual retrieval, masked, drop near-miss+no-op composite at 35.59%.
+- Current best clean deployment row is K4 compatible residual retrieval, near-miss challenger gate 0.02 at 36.00%.
 - 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.
 - K4 train-state residual retrieval, scale 0.40, safe residuals, mean-by-type tangent consensus improves the clean bridge but is not yet the main result.
 - Policy-baseline abstention inside the same-state no-expert lattice drops from 56.99% to 40.70%, so the mechanism result should emphasize counterfactual proposal geometry rather than policy fallback.
-- Singleton near-miss priors 0.01/0.02 tie the clean best rather than improving it; the compatible chart preserves high-precision near-miss tangents, but they remain sparse.
 - Candidate-oracle prefix K=8 is diagnostic-only; compare it to the clean selected row to decide whether the next paper move should be proposal generation or field ranking.
+- Candidate-oracle branch trace is diagnostic-only; use its rank histogram and branch success profile to design one clean selector-calibration run rather than a broad sweep.