Auto-sync: 2026-06-28 17:43:12 (part 2)

results/paper_analysis.json

@@ -1,10 +1,11 @@
 {
-  "generated_utc": "2026-06-28T17:53:26+00:00",
+  "best_clean_key": "residual_k4_consensus_noopbonus003",
+  "generated_utc": "2026-06-28T21:52:36+00:00",
   "mechanism_gap": {
-    "best_clean_vs_direct_same_ckpt": 0.06724637681159418,
-    "best_clean_vs_h16": 0.05275362318840582,
+    "best_clean_vs_direct_same_ckpt": 0.06956521739130428,
+    "best_clean_vs_h16": 0.05507246376811592,
     "same_state_full_vs_no_expert": 0.12347826086956515,
-    "same_state_no_expert_vs_best_clean": 0.21971014492753632,
+    "same_state_no_expert_vs_best_clean": 0.21739130434782622,
     "same_state_no_expert_vs_h16": 0.27246376811594214
   },
   "methods": {
@@ -348,6 +349,237 @@
       "source": "results/h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_k4s040_safe_margin0p20_mean_by_type_summary.json",
       "std_success": 0.018073573644196997
     },
+    "residual_k4_consensus_noopbonus003": {
+      "ci95_success": 0.031847551522122,
+      "label": "K4 mean-by-type tangent consensus, no-op bonus 0.03",
+      "mean_action_mse_to_best": 0.3952702766551596,
+      "mean_progress": 0.56683575115927,
+      "mean_success": 0.352463768115942,
+      "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": 0.3244027510331859,
+          "std_success": 0.026822320875242476
+        },
+        "PullCube-v1": {
+          "mean_num_groups": 81.0,
+          "mean_success": 0.20103794840636946,
+          "std_success": 0.014667153114049147
+        },
+        "PushCube-v1": {
+          "mean_num_groups": 102.0,
+          "mean_success": 0.7480972651606738,
+          "std_success": 0.0690465171514111
+        },
+        "StackCube-v1": {
+          "mean_num_groups": 93.33333333333333,
+          "mean_success": 0.2011137011137011,
+          "std_success": 0.07105663963231741
+        }
+      },
+      "seed_action_mse_to_best": {
+        "0": 0.38182404620168003,
+        "1": 0.38781667656386676,
+        "2": 0.4161701071999319
+      },
+      "seed_progress": {
+        "0": 0.5526963438863016,
+        "1": 0.5665759359364925,
+        "2": 0.5812349736550159
+      },
+      "seed_success": {
+        "0": 0.34782608695652173,
+        "1": 0.3426086956521739,
+        "2": 0.36695652173913046
+      },
+      "selected_candidate_type_counts": {
+        "retrieval_residual_policy_residual": 1649,
+        "retrieval_residual_residual_no_op": 53,
+        "retrieval_residual_residual_wrong_gripper": 23
+      },
+      "selected_residual_scale_counts": {
+        "0.4": 1725
+      },
+      "selected_type_outcomes": {
+        "retrieval_residual_policy_residual": {
+          "count": 1649.0,
+          "mean_progress": 0.5612481112154409,
+          "success_rate": 0.34566403881140084
+        },
+        "retrieval_residual_residual_no_op": {
+          "count": 53.0,
+          "mean_progress": 0.7346579036763254,
+          "success_rate": 0.5283018867924528
+        },
+        "retrieval_residual_residual_wrong_gripper": {
+          "count": 23.0,
+          "mean_progress": 0.580724628723186,
+          "success_rate": 0.43478260869565216
+        }
+      },
+      "source": "results/h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_k4s040_safe_margin0p20_mean_by_type_noopbonus0p03_summary.json",
+      "std_success": 0.01281933007970784
+    },
+    "residual_k4_consensus_noopbonus005": {
+      "ci95_success": 0.03271496576240179,
+      "label": "K4 mean-by-type tangent consensus, no-op bonus 0.05",
+      "mean_action_mse_to_best": 0.39539043844421057,
+      "mean_progress": 0.5665814660472467,
+      "mean_success": 0.3518840579710145,
+      "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": 0.3228001869306218,
+          "std_success": 0.025410290531060663
+        },
+        "PullCube-v1": {
+          "mean_num_groups": 81.0,
+          "mean_success": 0.20103794840636946,
+          "std_success": 0.014667153114049147
+        },
+        "PushCube-v1": {
+          "mean_num_groups": 102.0,
+          "mean_success": 0.7480972651606738,
+          "std_success": 0.0690465171514111
+        },
+        "StackCube-v1": {
+          "mean_num_groups": 93.33333333333333,
+          "mean_success": 0.2011137011137011,
+          "std_success": 0.07105663963231741
+        }
+      },
+      "seed_action_mse_to_best": {
+        "0": 0.3818464175352584,
+        "1": 0.3878977092857594,
+        "2": 0.4164271885116139
+      },
+      "seed_progress": {
+        "0": 0.5519389306648594,
+        "1": 0.5665547833624094,
+        "2": 0.581250684114471
+      },
+      "seed_success": {
+        "0": 0.34608695652173915,
+        "1": 0.3426086956521739,
+        "2": 0.36695652173913046
+      },
+      "selected_candidate_type_counts": {
+        "retrieval_residual_policy_residual": 1639,
+        "retrieval_residual_residual_no_op": 64,
+        "retrieval_residual_residual_wrong_gripper": 22
+      },
+      "selected_residual_scale_counts": {
+        "0.4": 1725
+      },
+      "selected_type_outcomes": {
+        "retrieval_residual_policy_residual": {
+          "count": 1639.0,
+          "mean_progress": 0.5608266842868225,
+          "success_rate": 0.34533251982916413
+        },
+        "retrieval_residual_residual_no_op": {
+          "count": 64.0,
+          "mean_progress": 0.7030392430024222,
+          "success_rate": 0.484375
+        },
+        "retrieval_residual_residual_wrong_gripper": {
+          "count": 22.0,
+          "mean_progress": 0.5983446287837896,
+          "success_rate": 0.45454545454545453
+        }
+      },
+      "source": "results/h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_k4s040_safe_margin0p20_mean_by_type_noopbonus0p05_summary.json",
+      "std_success": 0.013168483120696304
+    },
+    "residual_k4_consensus_noopbonus008": {
+      "ci95_success": 0.029933913043478266,
+      "label": "K4 mean-by-type tangent consensus, no-op bonus 0.08",
+      "mean_action_mse_to_best": 0.39648806852518,
+      "mean_progress": 0.5659873814658595,
+      "mean_success": 0.35130434782608694,
+      "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": 0.31951412114455596,
+          "std_success": 0.022561949563808425
+        },
+        "PullCube-v1": {
+          "mean_num_groups": 81.0,
+          "mean_success": 0.20542391331865018,
+          "std_success": 0.007170913229566424
+        },
+        "PushCube-v1": {
+          "mean_num_groups": 102.0,
+          "mean_success": 0.7480972651606738,
+          "std_success": 0.0690465171514111
+        },
+        "StackCube-v1": {
+          "mean_num_groups": 93.33333333333333,
+          "mean_success": 0.2011137011137011,
+          "std_success": 0.07105663963231741
+        }
+      },
+      "seed_action_mse_to_best": {
+        "0": 0.38186485262992587,
+        "1": 0.38981235122065183,
+        "2": 0.4177870017249623
+      },
+      "seed_progress": {
+        "0": 0.5510500524022981,
+        "1": 0.568158886189694,
+        "2": 0.5787532058055861
+      },
+      "seed_success": {
+        "0": 0.3443478260869565,
+        "1": 0.3443478260869565,
+        "2": 0.3652173913043478
+      },
+      "selected_candidate_type_counts": {
+        "retrieval_residual_policy_residual": 1612,
+        "retrieval_residual_residual_no_op": 91,
+        "retrieval_residual_residual_wrong_gripper": 22
+      },
+      "selected_residual_scale_counts": {
+        "0.4": 1725
+      },
+      "selected_type_outcomes": {
+        "retrieval_residual_policy_residual": {
+          "count": 1612.0,
+          "mean_progress": 0.5615867632030066,
+          "success_rate": 0.34615384615384615
+        },
+        "retrieval_residual_residual_no_op": {
+          "count": 91.0,
+          "mean_progress": 0.6361185594738185,
+          "success_rate": 0.4175824175824176
+        },
+        "retrieval_residual_residual_wrong_gripper": {
+          "count": 22.0,
+          "mean_progress": 0.5983446287837896,
+          "success_rate": 0.45454545454545453
+        }
+      },
+      "source": "results/h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_k4s040_safe_margin0p20_mean_by_type_noopbonus0p08_summary.json",
+      "std_success": 0.012049049096131323
+    },
     "same_state_full": {
       "ci95_success": 0.0885808391099616,
       "label": "Same-state lattice, full",
@@ -687,38 +919,38 @@
   },
   "paired_deltas": {
     "best_clean - canonical_h16": {
-      "ci95_delta": 0.051546144179697446,
-      "left": "best_clean_residual_k2",
-      "mean_delta": 0.052753623188405784,
+      "ci95_delta": 0.04759188850663325,
+      "left": "residual_k4_consensus_noopbonus003",
+      "mean_delta": 0.05507246376811594,
       "right": "h16_policy_canonical",
       "seed_deltas": {
-        "0": 0.059130434782608654,
-        "1": 0.029565217391304355,
-        "2": 0.06956521739130433
+        "0": 0.0643478260869565,
+        "1": 0.03304347826086956,
+        "2": 0.06782608695652176
       },
       "seeds": [
         0,
         1,
         2
       ],
-      "std_delta": 0.020748440774693643
+      "std_delta": 0.01915676712099514
     },
     "best_clean - direct_same_ckpt": {
-      "ci95_delta": 0.054366209142189946,
-      "left": "best_clean_residual_k2",
-      "mean_delta": 0.06724637681159418,
+      "ci95_delta": 0.0475264700722654,
+      "left": "residual_k4_consensus_noopbonus003",
+      "mean_delta": 0.06956521739130433,
       "right": "near_miss_policy_bc5",
       "seed_deltas": {
-        "0": 0.0643478260869565,
-        "1": 0.0469565217391304,
-        "2": 0.09043478260869564
+        "0": 0.06956521739130433,
+        "1": 0.050434782608695605,
+        "2": 0.08869565217391306
       },
       "seeds": [
         0,
         1,
         2
       ],
-      "std_delta": 0.021883578073248564
+      "std_delta": 0.01913043478260873
     },
     "full_lattice - no_expert_lattice": {
       "ci95_delta": 0.02628111194117426,
@@ -775,17 +1007,17 @@
   "per_task_deltas": {
     "best_clean_vs_h16": {
       "LiftPegUpright-v1": 0.03278131303915899,
-      "PickCube-v1": 0.11673380966859229,
-      "PullCube-v1": 0.015488215488215523,
-      "PushCube-v1": -0.0027056759730027524,
-      "StackCube-v1": 0.039608539608539606
+      "PickCube-v1": 0.1185454038714909,
+      "PullCube-v1": 0.0031834637097795104,
+      "PushCube-v1": -0.002459906715225779,
+      "StackCube-v1": 0.060147260147260156
     },
     "no_expert_vs_best_clean": {
       "LiftPegUpright-v1": 0.35374851747103375,
-      "PickCube-v1": 0.2988877064964022,
-      "PullCube-v1": -0.014862914862914883,
-      "PushCube-v1": 0.05319148936170215,
-      "StackCube-v1": 0.30194250194250194
+      "PickCube-v1": 0.29707611229350356,
+      "PullCube-v1": -0.0025581630844788705,
+      "PushCube-v1": 0.052945720103925176,
+      "StackCube-v1": 0.2814037814037814
     }
   }
 }

results/paper_analysis.md

@@ -1,6 +1,6 @@
 # Paper Analysis
 
-Generated: `2026-06-28T17:53:26+00:00`
+Generated: `2026-06-28T21:52:36+00:00`
 
 ## Main Seed Statistics
 
@@ -11,6 +11,9 @@ Generated: `2026-06-28T17:53:26+00:00`
 | near_miss_policy_bc5 | Near-miss proposal policy, direct | 3 | 28.29% +/- 0.80 | +/- 2.00 | 51.99% | 0.394 | -1.45 pp |
 | best_clean_residual_k2 | K2 residual transport, safe + margin 0.20 | 3 | 35.01% +/- 1.62 | +/- 4.01 | 56.70% | 0.398 | +5.28 pp |
 | residual_k4_consensus | K4 mean-by-type tangent consensus | 3 | 34.96% +/- 1.81 | +/- 4.49 | 56.65% | 0.395 | +5.22 pp |
+| residual_k4_consensus_noopbonus003 | K4 mean-by-type tangent consensus, no-op bonus 0.03 | 3 | 35.25% +/- 1.28 | +/- 3.18 | 56.68% | 0.395 | +5.51 pp |
+| residual_k4_consensus_noopbonus005 | K4 mean-by-type tangent consensus, no-op bonus 0.05 | 3 | 35.19% +/- 1.32 | +/- 3.27 | 56.66% | 0.395 | +5.45 pp |
+| residual_k4_consensus_noopbonus008 | K4 mean-by-type tangent consensus, no-op bonus 0.08 | 3 | 35.13% +/- 1.20 | +/- 2.99 | 56.60% | 0.396 | +5.39 pp |
 | same_state_near_miss | Same-state lattice, near-miss only | 3 | 55.94% +/- 3.29 | +/- 8.18 | 75.15% | 0.347 | +26.20 pp |
 | same_state_no_expert | Same-state lattice, no expert | 3 | 56.99% +/- 4.62 | +/- 11.47 | 75.01% | 0.459 | +27.25 pp |
 | same_state_policy_baseline | Same-state no-expert + policy candidate | 3 | 40.70% +/- 4.91 | +/- 12.19 | 63.12% | 0.438 | +10.96 pp |
@@ -20,8 +23,8 @@ Generated: `2026-06-28T17:53:26+00:00`
 
 | comparison | seeds | mean delta | 95% CI | seed deltas |
 |---|---:|---:|---:|---|
-| best_clean - canonical_h16 | 3 | +5.28 pp | +/- 5.15 | 0:+5.91, 1:+2.96, 2:+6.96 |
-| best_clean - direct_same_ckpt | 3 | +6.72 pp | +/- 5.44 | 0:+6.43, 1:+4.70, 2:+9.04 |
+| best_clean - canonical_h16 | 3 | +5.51 pp | +/- 4.76 | 0:+6.43, 1:+3.30, 2:+6.78 |
+| best_clean - direct_same_ckpt | 3 | +6.96 pp | +/- 4.75 | 0:+6.96, 1:+5.04, 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 |
@@ -31,16 +34,16 @@ Generated: `2026-06-28T17:53:26+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% | 32.26% | 58.13% | 62.15% | 84.19% | +11.67 pp | +29.89 pp |
-| PullCube-v1 | 19.79% | 21.33% | 15.02% | 19.85% | 22.41% | +1.55 pp | -1.49 pp |
-| PushCube-v1 | 75.06% | 74.79% | 82.54% | 80.10% | 81.92% | -0.27 pp | +5.32 pp |
-| StackCube-v1 | 14.10% | 18.06% | 50.41% | 48.25% | 60.83% | +3.96 pp | +30.19 pp |
+| PickCube-v1 | 20.59% | 32.44% | 58.13% | 62.15% | 84.19% | +11.85 pp | +29.71 pp |
+| PullCube-v1 | 19.79% | 20.10% | 15.02% | 19.85% | 22.41% | +0.32 pp | -0.26 pp |
+| PushCube-v1 | 75.06% | 74.81% | 82.54% | 80.10% | 81.92% | -0.25 pp | +5.29 pp |
+| StackCube-v1 | 14.10% | 20.11% | 50.41% | 48.25% | 60.83% | +6.01 pp | +28.14 pp |
 
 ## Mechanism Gap
 
-- Best clean residual transport improves over canonical h16 by +5.28 pp.
+- Best clean residual transport improves over canonical h16 by +5.51 pp.
 - Same-state no-expert lattice improves over canonical h16 by +27.25 pp.
-- Remaining clean-to-same-state proposal gap is +21.97 pp.
+- Remaining clean-to-same-state proposal gap is +21.74 pp.
 - Full lattice adds expert proposals and reaches 69.33%, a +12.35 pp gain over no-expert.
 
 ## Selection Histograms
@@ -49,7 +52,8 @@ Generated: `2026-06-28T17:53:26+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%)
-- `best_clean_residual_k2`: retrieval_residual_policy_residual=1610 (93.3%), retrieval_residual_residual_no_op=84 (4.9%), retrieval_residual_residual_wrong_gripper=31 (1.8%)
+- `residual_k4_consensus_noopbonus003`: retrieval_residual_policy_residual=1649 (95.6%), retrieval_residual_residual_no_op=53 (3.1%), retrieval_residual_residual_wrong_gripper=23 (1.3%)
+- `residual_k4_consensus_noopbonus003` residual scale counts: {'0.4': 1725}
 
 ## Selected-Type Outcomes
 
@@ -63,6 +67,15 @@ These rows are measured from raw rollout rows. In residual retrieval, `policy_re
 | residual_k4_consensus | retrieval_residual_policy_residual | 1666 | 34.27% | 56.20% |
 | residual_k4_consensus | retrieval_residual_residual_no_op | 35 | 62.86% | 78.50% |
 | residual_k4_consensus | retrieval_residual_residual_wrong_gripper | 24 | 41.67% | 56.28% |
+| residual_k4_consensus_noopbonus003 | retrieval_residual_policy_residual | 1649 | 34.57% | 56.12% |
+| residual_k4_consensus_noopbonus003 | retrieval_residual_residual_no_op | 53 | 52.83% | 73.47% |
+| residual_k4_consensus_noopbonus003 | retrieval_residual_residual_wrong_gripper | 23 | 43.48% | 58.07% |
+| residual_k4_consensus_noopbonus005 | retrieval_residual_policy_residual | 1639 | 34.53% | 56.08% |
+| residual_k4_consensus_noopbonus005 | retrieval_residual_residual_no_op | 64 | 48.44% | 70.30% |
+| residual_k4_consensus_noopbonus005 | retrieval_residual_residual_wrong_gripper | 22 | 45.45% | 59.83% |
+| residual_k4_consensus_noopbonus008 | retrieval_residual_policy_residual | 1612 | 34.62% | 56.16% |
+| residual_k4_consensus_noopbonus008 | retrieval_residual_residual_no_op | 91 | 41.76% | 63.61% |
+| residual_k4_consensus_noopbonus008 | retrieval_residual_residual_wrong_gripper | 22 | 45.45% | 59.83% |
 | same_state_no_expert | lattice_near_miss | 1263 | 63.18% | 82.44% |
 | same_state_no_expert | lattice_no_op | 222 | 52.25% | 69.05% |
 | same_state_no_expert | lattice_random_negative | 144 | 13.89% | 28.42% |

results/paper_core_results.md

@@ -5,9 +5,9 @@ baseline is the h=16 rank-checkpoint online rollout (`29.74%`).
 
 For paired seed deltas, per-task gaps, and selection histograms, regenerate and
 read `paper_analysis.md` with `python3 scripts/build_paper_analysis.py`. Current
-paired analysis: best clean K2 residual transport is `+5.28 pp` over canonical
-h=16, same-state no-expert lattice is `+27.25 pp`, and the remaining
-clean-to-same-state proposal gap is `+21.97 pp`.
+paired analysis: best clean K4 consensus with a typed no-op prior is `+5.51 pp`
+over canonical h=16, same-state no-expert lattice is `+27.25 pp`, and the
+remaining clean-to-same-state proposal gap is `+21.74 pp`.
 
 | Method | Uses same-state proposals | Uses expert proposal | Success | Gain vs policy | Interpretation |
 |---|---:|---:|---:|---:|---|
@@ -37,11 +37,12 @@ clean-to-same-state proposal gap is `+21.97 pp`.
 | Train-state residual retrieval, policy/no-op/wrong-gripper residuals | No | No | 33.68% | +3.94 pp | Typed family mask improves clean bridge |
 | Train-state residual retrieval, policy/no-op/wrong-gripper, scale 0.35 | No | No | 33.74% | +4.00 pp | Typed tangent transport before abstention |
 | Train-state residual retrieval, safe residuals + advantage margin 0.20 | No | No | 34.84% | +5.10 pp | Abstains unless field advantage beats policy |
-| K2 train-state residual retrieval, safe residuals + advantage margin 0.20 | No | No | 35.01% | +5.28 pp | Current best deployment-clean diagnostic; abstention makes a small train-neighborhood useful |
-| K4 train-state residual retrieval, safe residuals + mean-by-type tangent consensus | No | No | 34.96% | +5.22 pp | Near-tie clean diagnostic; consensus denoising does not beat raw K2 residuals |
-| K1 train-state residual ray-search, tight scales | No | No | 34.84% | +5.10 pp | Scale-grid ray-search is a near-tie but does not beat fixed K2 residual transport |
+| K2 train-state residual retrieval, safe residuals + advantage margin 0.20 | No | No | 35.01% | +5.28 pp | Previous best clean diagnostic; abstention makes a small train-neighborhood useful |
+| K4 train-state residual retrieval, safe residuals + mean-by-type tangent consensus | No | No | 34.96% | +5.22 pp | Near-tie clean diagnostic; consensus alone does not beat raw K2 residuals |
+| K4 mean-by-type residual retrieval + no-op prior 0.03 | No | No | 35.25% | +5.51 pp | Current best clean diagnostic; sparse typed prior nudges high-value no-op residuals without changing the core proposal family |
+| K1 train-state residual ray-search, tight scales | No | No | 34.84% | +5.10 pp | Scale-grid ray-search is a near-tie but does not beat the typed-prior clean row |
 | K2 train-state residual ray-search, tight scales | No | No | 34.84% | +5.10 pp | More scale choices along the same local rays do not improve the clean row |
-| K2 train-state residual ray-search, broad scales | No | No | 34.96% | +5.22 pp | Best ray-search row, still just below fixed K2 scale-0.40 |
+| K2 train-state residual ray-search, broad scales | No | No | 34.96% | +5.22 pp | Best ray-search row, still below the typed-prior clean row |
 | K4 train-state residual ray-search, tight scales | No | No | 34.55% | +4.81 pp | Larger neighborhood plus scale-grid dilutes the signal |
 | Policy-relative residual anchor, safe residuals | No | No | 33.74% | +4.00 pp | Policy-relative anchoring ties but does not improve expert-relative residuals |
 | Train-state residual retrieval, z-score metric | No | No | 32.23% | +2.49 pp | State normalization hurts nearest tangent retrieval here |
@@ -75,20 +76,21 @@ Suggested main-table rows:
 11. Train-state residual retrieval, typed safe families + advantage margin 0.20
 12. K2 train-state residual retrieval, typed safe families + advantage margin 0.20
 13. K4 train-state residual retrieval, mean-by-type tangent consensus
-14. K2 broad tangent ray-search
-15. Residual-tangent distillation policy
-16. Residual+Gaussian hybrid, K32 sigma0.35
-17. Lattice, near-miss only
-18. Lattice, no expert
-19. Lattice, no expert + policy baseline candidate
-20. Lattice, full
-21. Oracle ceiling
+14. K4 mean-by-type residual retrieval + no-op prior 0.03
+15. K2 broad tangent ray-search
+16. Residual-tangent distillation policy
+17. Residual+Gaussian hybrid, K32 sigma0.35
+18. Lattice, near-miss only
+19. Lattice, no expert
+20. Lattice, no expert + policy baseline candidate
+21. Lattice, full
+22. Oracle ceiling
 
 Suggested claim:
 
 > DoVLA-CIL is not a better behavior-cloning policy; it is a local counterfactual action
-> selection rule. Deployment-clean K2 typed counterfactual residual transport with advantage
-> abstention gives the strongest clean gain so far, while ungated KNN residual
+> selection rule. Deployment-clean K4 consensus residual transport with advantage
+> abstention and a small typed no-op prior gives the strongest clean gain so far, while ungated KNN residual
 > retrieval, field-gradient ascent, broader non-expert BC targets, field-teacher/tangent distillation, z-score retrieval,
 > train-family reliability priors, policy-relative anchoring, residual+Gaussian hybrids,
 > tangent consensus, tangent ray-search, and same-state policy-baseline fallback fail to improve the main rows.

results/paper_story_memo.md

@@ -16,7 +16,7 @@ when queried on proposal geometry that matches those local counterfactuals.
 | Same-state local counterfactual proposals are the mechanism | near-miss-only lattice is 55.94%; removing expert+near_miss drops to 25.57% | Strongly supported |
 | Conservative same-state result is large | no-expert lattice is 56.99% vs 29.74% policy | Main result |
 | Full lattice gives upper result | full lattice is 69.33%, oracle is 86.78% | Strong but label expert proposal clearly |
-| Deployment-clean proposal is currently a bottleneck | best clean K2 residual transport with counterfactual-advantage abstention is 35.01%, far below 56.99% | Supported |
+| Deployment-clean proposal is currently a bottleneck | best clean K4 consensus with a small typed no-op prior is 35.25%, far below 56.99% | Supported |
 | Gradient-based field optimization does not solve the clean proposal gap | `field_optim` best observed result is 25.39% | Negative diagnostic |
 | A broader non-expert proposal target does not reduce the proposal gap | direct broad non-expert policy is 27.88%; with field scoring it is 26.49% | Negative diagnostic |
 | Counterfactual residuals transfer better than absolute retrieved actions | nearest residual retrieval is 32.12% vs absolute retrieval 28.93%; KNN4 residual drops to 29.91% | Supported as a clean bridge |
@@ -25,11 +25,12 @@ when queried on proposal geometry that matches those local counterfactuals.
 | Seed-0 train-split field-teacher distillation does not solve the proposal gap | direct student is 26.84%; with field scoring it is 27.65% | Negative diagnostic |
 | All-split field-teacher distillation does not fix checkpointing/coverage | allmap direct is 28.00%; field-guided best is 26.49% despite 100% target coverage | Negative diagnostic |
 | Residual family consistency improves clean transport | policy/no-op/wrong-gripper typed residuals reach 33.74%, above raw 33.33% | Supported as diagnostic |
-| Counterfactual advantage abstention improves clean transport | requiring field advantage over the zero-residual policy raises typed residual transport to 34.84%, and K2 retrieval reaches 35.01% | Current best clean result |
-| Clean residual transport behaves like sparse intervention | `paper_analysis.md` shows K2 residual retrieval abstains to zero-residual policy on 93.3% of states, while selected nonzero no-op/wrong-gripper residuals succeed at ~41.7-41.9% vs 34.5% for abstention | Stronger clean-mechanism framing |
-| Tangent consensus is close but does not beat raw K2 residuals | K4 mean-by-type residual consensus reaches 34.96%, just below the 35.01% K2 raw residual row | Near-tie diagnostic |
-| Tangent ray-search does not beat the fixed clean row | K1/K2 tight scale-grid ray search reach 34.84%; K2 broad reaches 34.96%; K4 tight reaches 34.55%, all below fixed K2 scale-0.40 at 35.01% | Near-tie/negative diagnostic |
-| The proposal gap is now quantified | `paper_analysis.md` reports best clean +5.28 pp over canonical h16, same-state no-expert +27.25 pp, leaving a +21.97 pp clean-to-same-state gap | Core paper tension |
+| Counterfactual advantage abstention improves clean transport | requiring field advantage over the zero-residual policy raises typed residual transport to 34.84%, and K2 retrieval reaches 35.01% | Supported as the previous clean best |
+| Clean residual transport behaves like sparse intervention | `paper_analysis.md` shows the best clean row abstains to zero-residual policy on 95.6% of states, while selected nonzero no-op residuals succeed at 52.83% vs 34.57% for abstention | Stronger clean-mechanism framing |
+| Tangent consensus is close but needs sparse typing | K4 mean-by-type residual consensus reaches 34.96%; adding a small no-op residual prior raises it to 35.25% | Current best clean result |
+| Tangent ray-search does not beat the typed-prior clean row | K1/K2 tight scale-grid ray search reach 34.84%; K2 broad reaches 34.96%; K4 tight reaches 34.55%, all below the no-op-prior row at 35.25% | Near-tie/negative diagnostic |
+| Typed no-op residual prior improves the clean bridge | CPU smoke `14883591` passed, GPU sweeps `14883919`/`14883921`/`14883923` completed, and bonus 0.03 is best at 35.25% | Current best clean diagnostic |
+| The proposal gap is now quantified | `paper_analysis.md` reports best clean +5.51 pp over canonical h16, same-state no-expert +27.25 pp, leaving a +21.74 pp clean-to-same-state gap | Core paper tension |
 | Policy fallback is not the same-state mechanism | adding a policy baseline candidate to the no-expert same-state lattice drops 56.99% to 40.70% even with margin 0.00 | Negative diagnostic |
 | Z-score retrieval metric does not help | z-score rows reach 32.23-32.81%, below raw retrieval | Negative diagnostic |
 | Train-split residual family reliability does not recover the typed mask | after fixing threshold pass-through, scale-0.35 thresholds 0.10/0.25 reach 33.33%/33.28%, below typed safe residuals | Negative diagnostic |
@@ -56,18 +57,19 @@ clean proposal result, the intended main rows are:
 12. Train-state residual retrieval, typed safe families + advantage margin: 34.84%
 13. K2 train-state residual retrieval, typed safe families + advantage margin: 35.01%
 14. K4 mean-by-type tangent consensus: 34.96%
-15. K2 broad tangent ray-search: 34.96%
-16. K1/K2 tight tangent ray-search: 34.84% / 34.84%
-17. K4 tight tangent ray-search: 34.55%
-18. Residual-tangent distillation policy: 28.87%
-19. Z-score residual retrieval: 32.23-32.81%
-20. Train-family reliability prior: 33.28-33.33%
-21. Residual+Gaussian hybrid K32/K64: 31.30% / 30.90%
-22. Lattice, near-miss only: 55.94%
-23. Lattice, no expert: 56.99%
-24. Lattice, no expert + policy baseline candidate: 40.70%
-25. Lattice, full: 69.33%
-26. Oracle ceiling: 86.78%
+15. K4 mean-by-type tangent consensus + typed no-op prior 0.03: 35.25%
+16. K2 broad tangent ray-search: 34.96%
+17. K1/K2 tight tangent ray-search: 34.84% / 34.84%
+18. K4 tight tangent ray-search: 34.55%
+19. Residual-tangent distillation policy: 28.87%
+20. Z-score residual retrieval: 32.23-32.81%
+21. Train-family reliability prior: 33.28-33.33%
+22. Residual+Gaussian hybrid K32/K64: 31.30% / 30.90%
+23. Lattice, near-miss only: 55.94%
+24. Lattice, no expert: 56.99%
+25. Lattice, no expert + policy baseline candidate: 40.70%
+26. Lattice, full: 69.33%
+27. Oracle ceiling: 86.78%
 
 ## Novelty Framing
 
@@ -95,8 +97,9 @@ test-time search. The cleaner novelty is:
 
 ## Job Status
 
-Last checked: `2026-06-28 17:54 UTC`. The counterfactual tangent ray-search batch
-completed and table/analysis outputs were rebuilt.
+Last checked: `2026-06-28 21:52 UTC`. The counterfactual tangent ray-search batch
+completed, and the typed no-op residual-prior clean sweep completed after a passing
+CPU smoke.
 
 - `14858328`-`14858333`: completed train-split `field_selected_noexpert_bc5`;
   direct rollout is 26.84%, field-guided best is 27.65%.
@@ -134,24 +137,30 @@ completed and table/analysis outputs were rebuilt.
   Typed-safe residual transport at scale `0.35` with margin `0.20` or `0.22`
   reaches 34.84% mean success (+5.10 pp vs h=16).
 - `14862857`-`14862939`: completed KNN-with-abstention sweeps. K2 residual
-  retrieval at scale `0.40`, margin `0.20` is the current best clean row:
+  retrieval at scale `0.40`, margin `0.20` was the previous best clean row:
   35.01% mean success (+5.28 pp vs h=16).
 - `14868661`-`14868668`: completed same-state no-expert lattice with a prepended
   policy baseline candidate. The best setting, margin `0.00`, reaches only
   40.70%, far below the no-expert lattice's 56.99%; policy fallback should be
   framed as a negative diagnostic.
 - `14868693`-`14868700`: completed clean KNN residual mean-by-type consensus
-  sweep. K4, scale `0.40`, margin `0.20` reaches 34.96%, a near tie but still
-  below the 35.01% K2 raw residual best.
+  sweep. K4, scale `0.40`, margin `0.20` reaches 34.96%, a near tie below the
+  K2 raw residual row before adding the typed no-op prior.
 - `14868798`-`14868805`: completed consensus follow-up. K4 mean scales `0.425`
   and `0.45` reach 34.72% and 34.84%; K4 median and K8 mean at scale `0.40`
-  both reach 34.67%. The follow-up confirms K2 raw residual transport remains
-  the best clean row.
+  both reach 34.67%.
 - `14868993`/`14868995`/`14868997`/`14868999`: completed counterfactual tangent
   ray-search rollouts. Results are 34.84% for K1 tight, 34.84% for K2 tight,
   34.96% for K2 broad, and 34.55% for K4 tight. Summary jobs `14868994`/
   `14868996`/`14868998`/`14869000` and rebuild job `14869860` completed; the
   paper table and paired analysis outputs now include these rows.
+- `14883591`: completed CPU smoke for candidate-type potential bonuses with
+  `residual_no_op=0.05`. The smoke wrote valid rollout metadata and confirmed
+  the new CLI/Slurm path.
+- `14883919`/`14883921`/`14883923`: completed GPU clean sweeps for K4
+  mean-by-type residual retrieval with no-op residual bonuses `0.03`, `0.05`,
+  and `0.08`. Results are 35.25%, 35.19%, and 35.13%; summary jobs
+  `14883920`/`14883922`/`14883924` and rebuild job `14883926` completed.
 - `14869627`: completed CPU Apptainer smoke for the new residual scale-grid
   selector. It selected index `3` on a two-residual/two-scale toy case and
   returned the expected action `0.20`, validating the candidate expansion and
@@ -171,9 +180,10 @@ completed and table/analysis outputs were rebuilt.
 
 - Promote same-state no-expert lattice (56.99%) as the conservative mechanism
   result.
-- Use K2 typed safe residual transport with advantage abstention (35.01%) only as the current best clean
-  deployment diagnostic, not as a SOTA claim. The completed ray-search rows are
-  near-ties but do not replace it.
+- Use K4 mean-by-type residual transport with advantage abstention and a small
+  typed no-op prior (35.25%) as the current best clean deployment diagnostic,
+  not as a SOTA claim. The completed K2/ray-search rows are near-ties that
+  support the sparse-intervention story.
 - Use `results/paper_analysis.md` for paired seed deltas, per-task gaps, and
   selection histograms when writing reviewer-facing tables.
 - Treat z-score retrieval, repaired train-family reliability priors, Gaussian hybrids,

results/paper_table_status.json

@@ -505,7 +505,7 @@
       "clean_deployment": "yes",
       "same_state_proposals": "no",
       "expert_proposal": "no",
-      "story_role": "best clean counterfactual advantage abstention",
+      "story_role": "previous best counterfactual advantage abstention",
       "fallback_success": null,
       "pending_job": "14862936/14862937",
       "path_exists": true,
@@ -619,17 +619,17 @@
       "clean_deployment": "yes",
       "same_state_proposals": "no",
       "expert_proposal": "no",
-      "story_role": "typed sparse-intervention prior diagnostic",
+      "story_role": "current best clean typed sparse-intervention prior",
       "fallback_success": null,
-      "pending_job": "14883370/14883371",
-      "path_exists": false,
-      "status": "pending",
-      "success": null,
-      "std_success": null,
+      "pending_job": "14883919/14883920",
+      "path_exists": true,
+      "status": "complete",
+      "success": 0.35246376811594204,
+      "std_success": 0.01281933007970784,
       "completed_seeds": null,
-      "num_completed": null,
+      "num_completed": 3,
       "best_config": null,
-      "gain_vs_h16_policy": null
+      "gain_vs_h16_policy": 0.055072463768115976
     },
     {
       "key": "retrieval_residual_k4_mean_noopbonus005",
@@ -640,15 +640,15 @@
       "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,
+      "pending_job": "14883921/14883922",
+      "path_exists": true,
+      "status": "complete",
+      "success": 0.3518840579710145,
+      "std_success": 0.013168483120696304,
       "completed_seeds": null,
-      "num_completed": null,
+      "num_completed": 3,
       "best_config": null,
-      "gain_vs_h16_policy": null
+      "gain_vs_h16_policy": 0.05449275362318845
     },
     {
       "key": "retrieval_residual_k4_mean_noopbonus008",
@@ -659,15 +659,15 @@
       "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,
+      "pending_job": "14883923/14883924",
+      "path_exists": true,
+      "status": "complete",
+      "success": 0.35130434782608694,
+      "std_success": 0.01204904909613132,
       "completed_seeds": null,
-      "num_completed": null,
+      "num_completed": 3,
       "best_config": null,
-      "gain_vs_h16_policy": null
+      "gain_vs_h16_policy": 0.05391304347826087
     },
     {
       "key": "retrieval_residual_policy_anchor_scale035_safe",
@@ -1146,23 +1146,23 @@
     }
   ],
   "best_clean": {
-    "key": "retrieval_residual_knn2_scale040_safe_margin020",
-    "label": "K2 train-state residual retrieval, scale 0.40, safe residuals, advantage margin 0.20",
-    "path": "h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_knn2_scale0p40_safe_types_margin0p20_summary.json",
+    "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": "best clean counterfactual advantage abstention",
+    "story_role": "current best clean typed sparse-intervention prior",
     "fallback_success": null,
-    "pending_job": "14862936/14862937",
+    "pending_job": "14883919/14883920",
     "path_exists": true,
     "status": "complete",
-    "success": 0.3501449275362319,
-    "std_success": 0.016159257814221867,
+    "success": 0.35246376811594204,
+    "std_success": 0.01281933007970784,
     "completed_seeds": null,
     "num_completed": 3,
     "best_config": null,
-    "gain_vs_h16_policy": 0.05275362318840582
+    "gain_vs_h16_policy": 0.055072463768115976
   },
   "best_mechanism_no_expert": {
     "key": "no_expert_lattice",
@@ -1187,7 +1187,7 @@
     "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 K2 train-state residual retrieval, scale 0.40, safe residuals, advantage margin 0.20 at 35.01%.",
+    "Current best clean deployment row is K4 mean-by-type residual retrieval, scale 0.40, margin 0.20, no-op residual bonus 0.03 at 35.25%.",
     "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

@@ -29,15 +29,15 @@ Baseline h=16 policy: 29.74%
 | retrieval_residual_scale035_safe_types | Train-state residual retrieval, scale 0.35, policy/no-op/wrong-gripper residuals | complete | 33.74% | +4.00 pp | yes | no | no | typed tangent scale fine sweep |
 | retrieval_residual_scale035_safe_margin020 | Train-state residual retrieval, scale 0.35, safe residuals, advantage margin 0.20 | complete | 34.84% | +5.10 pp | yes | no | no | counterfactual advantage abstention |
 | retrieval_residual_scale050_safe_margin020 | Train-state residual retrieval, scale 0.50, safe residuals, advantage margin 0.20 | complete | 34.84% | +5.10 pp | yes | no | no | counterfactual advantage abstention scale tie |
-| retrieval_residual_knn2_scale040_safe_margin020 | K2 train-state residual retrieval, scale 0.40, safe residuals, advantage margin 0.20 | complete | 35.01% | +5.28 pp | yes | no | no | best clean counterfactual advantage abstention |
+| retrieval_residual_knn2_scale040_safe_margin020 | K2 train-state residual retrieval, scale 0.40, safe residuals, advantage margin 0.20 | complete | 35.01% | +5.28 pp | yes | no | no | previous best counterfactual advantage abstention |
 | retrieval_residual_k1grid_tight_safe_ray_margin020 | K1 train-state residual ray search, safe residuals, scales 0.30/0.40/0.50, advantage margin 0.20 | complete | 34.84% | +5.10 pp | yes | no | no | counterfactual tangent ray-search diagnostic |
 | retrieval_residual_k2grid_tight_safe_ray_margin020 | K2 train-state residual ray search, safe residuals, scales 0.30/0.40/0.50, advantage margin 0.20 | complete | 34.84% | +5.10 pp | yes | no | no | counterfactual tangent ray-search diagnostic |
 | 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_k4_mean_noopbonus003 | K4 mean-by-type residual retrieval, scale 0.40, margin 0.20, no-op residual bonus 0.03 | complete | 35.25% | +5.51 pp | yes | no | no | current best clean typed sparse-intervention prior |
+| retrieval_residual_k4_mean_noopbonus005 | K4 mean-by-type residual retrieval, scale 0.40, margin 0.20, no-op residual bonus 0.05 | complete | 35.19% | +5.45 pp | 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 | complete | 35.13% | +5.39 pp | 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 |
@@ -69,7 +69,7 @@ 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 K2 train-state residual retrieval, scale 0.40, safe residuals, advantage margin 0.20 at 35.01%.
+- Current best clean deployment row is K4 mean-by-type residual retrieval, scale 0.40, margin 0.20, no-op residual bonus 0.03 at 35.25%.
 - 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

@@ -14,6 +14,7 @@ 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_noopbonus003"
 
 
 @dataclass(frozen=True)
@@ -58,6 +59,62 @@ METHODS = [
             "k4s040_safe_margin0p20_mean_by_type_summary.json"
         ),
     ),
+    MethodSpec(
+        key="residual_k4_consensus_noopbonus003",
+        label="K4 mean-by-type tangent consensus, no-op bonus 0.03",
+        summary_path=(
+            "h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_"
+            "k4s040_safe_margin0p20_mean_by_type_noopbonus0p03_summary.json"
+        ),
+    ),
+    MethodSpec(
+        key="residual_k4_consensus_noopbonus001",
+        label="K4 mean-by-type tangent consensus, no-op bonus 0.01",
+        summary_path=(
+            "h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_"
+            "k4s040_safe_margin0p20_mean_by_type_noopbonus0p01_summary.json"
+        ),
+    ),
+    MethodSpec(
+        key="residual_k4_consensus_noopbonus002",
+        label="K4 mean-by-type tangent consensus, no-op bonus 0.02",
+        summary_path=(
+            "h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_"
+            "k4s040_safe_margin0p20_mean_by_type_noopbonus0p02_summary.json"
+        ),
+    ),
+    MethodSpec(
+        key="residual_k4_consensus_noopbonus0025",
+        label="K4 mean-by-type tangent consensus, no-op bonus 0.025",
+        summary_path=(
+            "h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_"
+            "k4s040_safe_margin0p20_mean_by_type_noopbonus0p025_summary.json"
+        ),
+    ),
+    MethodSpec(
+        key="residual_k4_consensus_noopbonus0035",
+        label="K4 mean-by-type tangent consensus, no-op bonus 0.035",
+        summary_path=(
+            "h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_"
+            "k4s040_safe_margin0p20_mean_by_type_noopbonus0p035_summary.json"
+        ),
+    ),
+    MethodSpec(
+        key="residual_k4_consensus_noopbonus005",
+        label="K4 mean-by-type tangent consensus, no-op bonus 0.05",
+        summary_path=(
+            "h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_"
+            "k4s040_safe_margin0p20_mean_by_type_noopbonus0p05_summary.json"
+        ),
+    ),
+    MethodSpec(
+        key="residual_k4_consensus_noopbonus008",
+        label="K4 mean-by-type tangent consensus, no-op bonus 0.08",
+        summary_path=(
+            "h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_"
+            "k4s040_safe_margin0p20_mean_by_type_noopbonus0p08_summary.json"
+        ),
+    ),
     MethodSpec(
         key="same_state_near_miss",
         label="Same-state lattice, near-miss only",
@@ -382,7 +439,7 @@ 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_residual_k2"]["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"]
@@ -404,7 +461,7 @@ 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_residual_k2"]:
+    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())
@@ -415,9 +472,9 @@ def _render_markdown(report: dict[str, Any]) -> str:
         else:
             summary = "not recorded"
         lines.append(f"- `{key}`: {summary}")
-    scale_counts = methods["best_clean_residual_k2"].get("selected_residual_scale_counts", {})
+    scale_counts = methods[BEST_CLEAN_KEY].get("selected_residual_scale_counts", {})
     if scale_counts:
-        lines.append(f"- `best_clean_residual_k2` residual scale counts: {scale_counts}")
+        lines.append(f"- `{BEST_CLEAN_KEY}` residual scale counts: {scale_counts}")
     lines.extend(
         [
             "",
@@ -429,7 +486,19 @@ def _render_markdown(report: dict[str, Any]) -> str:
             "|---|---|---:|---:|---:|",
         ]
     )
-    for key in ["best_clean_residual_k2", "residual_k4_consensus", "same_state_no_expert", "same_state_policy_baseline"]:
+    for key in [
+        "best_clean_residual_k2",
+        "residual_k4_consensus",
+        "residual_k4_consensus_noopbonus003",
+        "residual_k4_consensus_noopbonus001",
+        "residual_k4_consensus_noopbonus002",
+        "residual_k4_consensus_noopbonus0025",
+        "residual_k4_consensus_noopbonus0035",
+        "residual_k4_consensus_noopbonus005",
+        "residual_k4_consensus_noopbonus008",
+        "same_state_no_expert",
+        "same_state_policy_baseline",
+    ]:
         for candidate_type, values in methods[key].get("selected_type_outcomes", {}).items():
             lines.append(
                 f"| {key} | {candidate_type} | {int(values['count'])} | {_pct(values['success_rate'])} | {_pct(values['mean_progress'])} |"
@@ -440,8 +509,8 @@ def _render_markdown(report: dict[str, Any]) -> str:
 def build_report() -> dict[str, Any]:
     methods = _load_methods()
     paired_deltas = {
-        "best_clean - canonical_h16": _paired_delta(methods, "best_clean_residual_k2", "h16_policy_canonical"),
-        "best_clean - direct_same_ckpt": _paired_delta(methods, "best_clean_residual_k2", "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(
@@ -451,14 +520,14 @@ def build_report() -> dict[str, Any]:
         ),
     }
     mechanism_gap = {
-        "best_clean_vs_h16": methods["best_clean_residual_k2"]["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_residual_k2"]["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_residual_k2"]["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"],
     }
@@ -467,10 +536,11 @@ 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_residual_k2", "h16_policy_canonical"),
-            "no_expert_vs_best_clean": _per_task_delta(methods, "same_state_no_expert", "best_clean_residual_k2"),
+            "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,
     }
 
 

scripts/build_paper_table_status.py

@@ -282,7 +282,7 @@ SPECS = [
         clean_deployment="yes",
         same_state_proposals="no",
         expert_proposal="no",
-        story_role="best clean counterfactual advantage abstention",
+        story_role="previous best counterfactual advantage abstention",
         pending_job="14862936/14862937",
     ),
     ResultSpec(
@@ -342,8 +342,44 @@ SPECS = [
         clean_deployment="yes",
         same_state_proposals="no",
         expert_proposal="no",
-        story_role="typed sparse-intervention prior diagnostic",
-        pending_job="14883694/14883695",
+        story_role="current best clean typed sparse-intervention prior",
+        pending_job="14883919/14883920",
+    ),
+    ResultSpec(
+        key="retrieval_residual_k4_mean_noopbonus001",
+        label="K4 mean-by-type residual retrieval, scale 0.40, margin 0.20, no-op residual bonus 0.01",
+        path="h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_k4s040_safe_margin0p20_mean_by_type_noopbonus0p01_summary.json",
+        clean_deployment="yes",
+        same_state_proposals="no",
+        expert_proposal="no",
+        story_role="typed sparse-intervention prior fine sweep",
+    ),
+    ResultSpec(
+        key="retrieval_residual_k4_mean_noopbonus002",
+        label="K4 mean-by-type residual retrieval, scale 0.40, margin 0.20, no-op residual bonus 0.02",
+        path="h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_k4s040_safe_margin0p20_mean_by_type_noopbonus0p02_summary.json",
+        clean_deployment="yes",
+        same_state_proposals="no",
+        expert_proposal="no",
+        story_role="typed sparse-intervention prior fine sweep",
+    ),
+    ResultSpec(
+        key="retrieval_residual_k4_mean_noopbonus0025",
+        label="K4 mean-by-type residual retrieval, scale 0.40, margin 0.20, no-op residual bonus 0.025",
+        path="h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_k4s040_safe_margin0p20_mean_by_type_noopbonus0p025_summary.json",
+        clean_deployment="yes",
+        same_state_proposals="no",
+        expert_proposal="no",
+        story_role="typed sparse-intervention prior fine sweep",
+    ),
+    ResultSpec(
+        key="retrieval_residual_k4_mean_noopbonus0035",
+        label="K4 mean-by-type residual retrieval, scale 0.40, margin 0.20, no-op residual bonus 0.035",
+        path="h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_k4s040_safe_margin0p20_mean_by_type_noopbonus0p035_summary.json",
+        clean_deployment="yes",
+        same_state_proposals="no",
+        expert_proposal="no",
+        story_role="typed sparse-intervention prior fine sweep",
     ),
     ResultSpec(
         key="retrieval_residual_k4_mean_noopbonus005",
@@ -353,7 +389,7 @@ SPECS = [
         same_state_proposals="no",
         expert_proposal="no",
         story_role="typed sparse-intervention prior diagnostic",
-        pending_job="14883696/14883697",
+        pending_job="14883921/14883922",
     ),
     ResultSpec(
         key="retrieval_residual_k4_mean_noopbonus008",
@@ -363,7 +399,7 @@ SPECS = [
         same_state_proposals="no",
         expert_proposal="no",
         story_role="typed sparse-intervention prior diagnostic",
-        pending_job="14883698/14883699",
+        pending_job="14883923/14883924",
     ),
     ResultSpec(
         key="retrieval_residual_policy_anchor_scale035_safe",