Auto-sync: 2026-06-29 08:05:55 (part 3)

results/paper_analysis.json

@@ -1,6 +1,6 @@
 {
   "best_clean_key": "residual_k4_consensus_grid035040045_noopbonus003",
-  "generated_utc": "2026-06-29T11:38:14+00:00",
+  "generated_utc": "2026-06-29T12:01:00+00:00",
   "mechanism_gap": {
     "best_clean_vs_direct_same_ckpt": 0.07130434782608691,
     "best_clean_vs_h16": 0.05681159420289855,
@@ -274,6 +274,310 @@
       "source": "results/h16_policy_ckpt_near_miss_policy_bc5_summary.json",
       "std_success": 0.008032699397420892
     },
+    "repair_nearmiss_k4_grid025035050_margin010": {
+      "ci95_success": 0.03674617285159992,
+      "label": "K4 near-miss-to-expert repair tangent, scales 0.25/0.35/0.50, margin 0.10",
+      "mean_action_mse_to_best": 0.3929642635457002,
+      "mean_progress": 0.5600691433255407,
+      "mean_success": 0.34144927536231884,
+      "num_completed": 3,
+      "per_task_success": {
+        "LiftPegUpright-v1": {
+          "mean_num_groups": 102.0,
+          "mean_success": 0.26635850338067285,
+          "std_success": 0.05831834141997673
+        },
+        "PickCube-v1": {
+          "mean_num_groups": 196.66666666666666,
+          "mean_success": 0.3142074310552571,
+          "std_success": 0.02741955750008872
+        },
+        "PullCube-v1": {
+          "mean_num_groups": 81.0,
+          "mean_success": 0.15871370345054556,
+          "std_success": 0.036307042236911065
+        },
+        "PushCube-v1": {
+          "mean_num_groups": 102.0,
+          "mean_success": 0.7377047365461099,
+          "std_success": 0.07936301124798102
+        },
+        "StackCube-v1": {
+          "mean_num_groups": 93.33333333333333,
+          "mean_success": 0.19745069745069743,
+          "std_success": 0.07489140417757721
+        }
+      },
+      "seed_action_mse_to_best": {
+        "0": 0.3795984360802433,
+        "1": 0.3867724084287234,
+        "2": 0.41252194612813387
+      },
+      "seed_progress": {
+        "0": 0.5459818405164001,
+        "1": 0.5586519980981298,
+        "2": 0.5755735913620922
+      },
+      "seed_success": {
+        "0": 0.3356521739130435,
+        "1": 0.33043478260869563,
+        "2": 0.3582608695652174
+      },
+      "selected_candidate_type_counts": {
+        "retrieval_residual_policy_residual": 1616,
+        "retrieval_residual_residual_near_miss": 109
+      },
+      "selected_residual_scale_counts": {
+        "0.25": 1636,
+        "0.35": 11,
+        "0.5": 78
+      },
+      "selected_type_outcomes": {
+        "retrieval_residual_policy_residual": {
+          "count": 1616.0,
+          "mean_progress": 0.5759515361980464,
+          "success_rate": 0.35705445544554454
+        },
+        "retrieval_residual_residual_near_miss": {
+          "count": 109.0,
+          "mean_progress": 0.3246017407386675,
+          "success_rate": 0.11009174311926606
+        }
+      },
+      "source": "results/h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_repair_nearmiss_k4_grid025035050_margin0p10_summary.json",
+      "std_success": 0.014791131388026773
+    },
+    "repair_nearmiss_k4_grid025035050_margin020": {
+      "ci95_success": 0.03355996769539608,
+      "label": "K4 near-miss-to-expert repair tangent, scales 0.25/0.35/0.50, margin 0.20",
+      "mean_action_mse_to_best": 0.393933949969655,
+      "mean_progress": 0.5597380874861161,
+      "mean_success": 0.34318840579710147,
+      "num_completed": 3,
+      "per_task_success": {
+        "LiftPegUpright-v1": {
+          "mean_num_groups": 102.0,
+          "mean_success": 0.26635850338067285,
+          "std_success": 0.05831834141997673
+        },
+        "PickCube-v1": {
+          "mean_num_groups": 196.66666666666666,
+          "mean_success": 0.3142074310552571,
+          "std_success": 0.02741955750008872
+        },
+        "PullCube-v1": {
+          "mean_num_groups": 81.0,
+          "mean_success": 0.16304270777954988,
+          "std_success": 0.038804868325098954
+        },
+        "PushCube-v1": {
+          "mean_num_groups": 102.0,
+          "mean_success": 0.7379505058038869,
+          "std_success": 0.07221758584710138
+        },
+        "StackCube-v1": {
+          "mean_num_groups": 93.33333333333333,
+          "mean_success": 0.20481740481740482,
+          "std_success": 0.06861088450109236
+        }
+      },
+      "seed_action_mse_to_best": {
+        "0": 0.3804865829549406,
+        "1": 0.38731886736886656,
+        "2": 0.41399639958515766
+      },
+      "seed_progress": {
+        "0": 0.5470242628576639,
+        "1": 0.5587378957705653,
+        "2": 0.5734521038301191
+      },
+      "seed_success": {
+        "0": 0.3391304347826087,
+        "1": 0.33217391304347826,
+        "2": 0.3582608695652174
+      },
+      "selected_candidate_type_counts": {
+        "retrieval_residual_policy_residual": 1696,
+        "retrieval_residual_residual_near_miss": 29
+      },
+      "selected_residual_scale_counts": {
+        "0.25": 1698,
+        "0.35": 3,
+        "0.5": 24
+      },
+      "selected_type_outcomes": {
+        "retrieval_residual_policy_residual": {
+          "count": 1696.0,
+          "mean_progress": 0.564825300580208,
+          "success_rate": 0.3472877358490566
+        },
+        "retrieval_residual_residual_near_miss": {
+          "count": 29.0,
+          "mean_progress": 0.26222383205232946,
+          "success_rate": 0.10344827586206896
+        }
+      },
+      "source": "results/h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_repair_nearmiss_k4_grid025035050_margin0p20_summary.json",
+      "std_success": 0.013508614722007019
+    },
+    "repair_nearmiss_k4_grid035050075_margin020": {
+      "ci95_success": 0.03725072058131128,
+      "label": "K4 near-miss-to-expert repair tangent, scales 0.35/0.50/0.75, margin 0.20",
+      "mean_action_mse_to_best": 0.393982858941013,
+      "mean_progress": 0.5605357405742194,
+      "mean_success": 0.343768115942029,
+      "num_completed": 3,
+      "per_task_success": {
+        "LiftPegUpright-v1": {
+          "mean_num_groups": 102.0,
+          "mean_success": 0.2698307256028951,
+          "std_success": 0.06430281802636649
+        },
+        "PickCube-v1": {
+          "mean_num_groups": 196.66666666666666,
+          "mean_success": 0.31239583685235855,
+          "std_success": 0.03055175955315912
+        },
+        "PullCube-v1": {
+          "mean_num_groups": 81.0,
+          "mean_success": 0.1673717121085542,
+          "std_success": 0.042495599376252136
+        },
+        "PushCube-v1": {
+          "mean_num_groups": 102.0,
+          "mean_success": 0.7379505058038869,
+          "std_success": 0.07221758584710138
+        },
+        "StackCube-v1": {
+          "mean_num_groups": 93.33333333333333,
+          "mean_success": 0.20481740481740482,
+          "std_success": 0.06861088450109236
+        }
+      },
+      "seed_action_mse_to_best": {
+        "0": 0.3805449200680722,
+        "1": 0.38746632555418686,
+        "2": 0.4139373312007798
+      },
+      "seed_progress": {
+        "0": 0.5484290182841537,
+        "1": 0.5579161341779906,
+        "2": 0.5752620692605136
+      },
+      "seed_success": {
+        "0": 0.3408695652173913,
+        "1": 0.33043478260869563,
+        "2": 0.36
+      },
+      "selected_candidate_type_counts": {
+        "retrieval_residual_policy_residual": 1690,
+        "retrieval_residual_residual_near_miss": 35
+      },
+      "selected_residual_scale_counts": {
+        "0.35": 1695,
+        "0.5": 7,
+        "0.75": 23
+      },
+      "selected_type_outcomes": {
+        "retrieval_residual_policy_residual": {
+          "count": 1690.0,
+          "mean_progress": 0.5654949464062802,
+          "success_rate": 0.3479289940828402
+        },
+        "retrieval_residual_residual_near_miss": {
+          "count": 35.0,
+          "mean_progress": 0.3210769446832793,
+          "success_rate": 0.14285714285714285
+        }
+      },
+      "source": "results/h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_repair_nearmiss_k4_grid035050075_margin0p20_summary.json",
+      "std_success": 0.014994222790002973
+    },
+    "repair_safe_k4_grid025035050_margin020": {
+      "ci95_success": 0.03115620450679224,
+      "label": "K4 safe-family-to-expert repair tangent, scales 0.25/0.35/0.50, margin 0.20",
+      "mean_action_mse_to_best": 0.39449909915337744,
+      "mean_progress": 0.5602484929151746,
+      "mean_success": 0.3443478260869566,
+      "num_completed": 3,
+      "per_task_success": {
+        "LiftPegUpright-v1": {
+          "mean_num_groups": 102.0,
+          "mean_success": 0.27225820839542214,
+          "std_success": 0.05495571671473726
+        },
+        "PickCube-v1": {
+          "mean_num_groups": 196.66666666666666,
+          "mean_success": 0.31239583685235855,
+          "std_success": 0.03055175955315912
+        },
+        "PullCube-v1": {
+          "mean_num_groups": 81.0,
+          "mean_success": 0.1673717121085542,
+          "std_success": 0.042495599376252136
+        },
+        "PushCube-v1": {
+          "mean_num_groups": 102.0,
+          "mean_success": 0.7379505058038869,
+          "std_success": 0.07221758584710138
+        },
+        "StackCube-v1": {
+          "mean_num_groups": 93.33333333333333,
+          "mean_success": 0.20481740481740482,
+          "std_success": 0.06861088450109236
+        }
+      },
+      "seed_action_mse_to_best": {
+        "0": 0.38139278126475606,
+        "1": 0.38822281592728003,
+        "2": 0.41388170026809623
+      },
+      "seed_progress": {
+        "0": 0.5477940025257514,
+        "1": 0.5587685327847367,
+        "2": 0.574182943435036
+      },
+      "seed_success": {
+        "0": 0.3408695652173913,
+        "1": 0.3339130434782609,
+        "2": 0.3582608695652174
+      },
+      "selected_candidate_type_counts": {
+        "retrieval_residual_policy_residual": 1672,
+        "retrieval_residual_residual_near_miss": 24,
+        "retrieval_residual_residual_no_op": 27,
+        "retrieval_residual_residual_wrong_gripper": 2
+      },
+      "selected_residual_scale_counts": {
+        "0.25": 1679,
+        "0.35": 6,
+        "0.5": 40
+      },
+      "selected_type_outcomes": {
+        "retrieval_residual_policy_residual": {
+          "count": 1672.0,
+          "mean_progress": 0.5675147408519728,
+          "success_rate": 0.3492822966507177
+        },
+        "retrieval_residual_residual_near_miss": {
+          "count": 24.0,
+          "mean_progress": 0.2806516041358312,
+          "success_rate": 0.125
+        },
+        "retrieval_residual_residual_no_op": {
+          "count": 27.0,
+          "mean_progress": 0.3262357435154694,
+          "success_rate": 0.18518518518518517
+        },
+        "retrieval_residual_residual_wrong_gripper": {
+          "count": 2.0,
+          "mean_progress": 1.0,
+          "success_rate": 1.0
+        }
+      },
+      "source": "results/h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_repair_safe_k4_grid025035050_margin0p20_summary.json",
+      "std_success": 0.012541047914657353
+    },
     "residual_k4_consensus": {
       "ci95_success": 0.04490086956521744,
       "label": "K4 mean-by-type tangent consensus",

results/paper_analysis.md

@@ -1,6 +1,6 @@
 # Paper Analysis
 
-Generated: `2026-06-29T11:38:14+00:00`
+Generated: `2026-06-29T12:01:00+00:00`
 
 ## Main Seed Statistics
 
@@ -43,6 +43,10 @@ Generated: `2026-06-29T11:38:14+00:00`
 | residual_k4_consensus_grid035040045_noopbonus003_consensus002 | K4 mean-by-type tangent consensus, scales 0.35/0.40/0.45, no-op bonus 0.03, consensus penalty 0.02 | 3 | 35.36% +/- 1.16 | +/- 2.88 | 56.82% | 0.398 | +5.62 pp |
 | residual_k4_consensus_grid035040045_noopbonus003_consensus005 | K4 mean-by-type tangent consensus, scales 0.35/0.40/0.45, no-op bonus 0.03, consensus penalty 0.05 | 3 | 35.36% +/- 1.16 | +/- 2.88 | 56.78% | 0.398 | +5.62 pp |
 | 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 |
+| 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 |
+| repair_safe_k4_grid025035050_margin020 | K4 safe-family-to-expert repair tangent, scales 0.25/0.35/0.50, margin 0.20 | 3 | 34.43% +/- 1.25 | +/- 3.12 | 56.02% | 0.394 | +4.70 pp |
 | residual_k4_consensus_grid035040045_noopbonus003_l2penalty005 | K4 mean-by-type tangent consensus, scales 0.35/0.40/0.45, no-op bonus 0.03, action L2 penalty 0.05 | 3 | 35.42% +/- 1.12 | +/- 2.78 | 56.87% | 0.397 | +5.68 pp |
 | residual_k4_consensus_grid035040045_noopbonus003_l2penalty010 | K4 mean-by-type tangent consensus, scales 0.35/0.40/0.45, no-op bonus 0.03, action L2 penalty 0.10 | 3 | 35.36% +/- 1.16 | +/- 2.88 | 56.80% | 0.397 | +5.62 pp |
 | residual_k4_consensus_grid035040045_noopbonus003_l2penalty020 | K4 mean-by-type tangent consensus, scales 0.35/0.40/0.45, no-op bonus 0.03, action L2 penalty 0.20 | 3 | 35.36% +/- 1.16 | +/- 2.88 | 56.78% | 0.397 | +5.62 pp |

results/paper_core_results.md

@@ -52,6 +52,7 @@ no-op prior is `+5.68 pp` over canonical h=16, same-state no-expert lattice is
 | K4 mean-by-type residual retrieval + source-advantage prior/gate | No | No | 35.13-35.30% | +5.39-5.57 pp | Measuring local train-source utility lift over the anchor does not replace the typed no-op prior; positive-advantage gates over-filter useful residual geometry |
 | K4 mean-by-type residual retrieval + train-family success bonus | No | No | 35.25-35.42% | +5.51-5.68 pp | A continuous train terminal-success prior is below the best by itself and only ties when added to the no-op row; train outcome reliability does not add the gain |
 | K4 mean-by-type residual retrieval + train-neighbor consensus penalty | No | No | 35.19-35.36% | +5.45-5.62 pp | Penalizing high-dispersion local tangent families is coherent but over-abstains by one success; geometric confidence does not improve the sparse no-op scale-grid row |
+| K4 repair-tangent residual transport | No | No | 34.14-34.43% | +4.41-4.70 pp | Reversing residuals into failure-to-expert repair tangents is a clean negative diagnostic; the current gain is not recovered by transporting near-miss-to-expert vectors |
 | K4 mean-by-type residual retrieval + source-score prior 0.025 | No | No | 35.19% | +5.45 pp | A stronger reward-score prior drops below the plateau |
 | K4 mean-by-type residual retrieval, no-op-only residuals | No | No | 35.19% | +5.45 pp | Removing wrong-gripper residuals loses one success versus the fixed-scale safe-family plateau; the core gain is sparse no-op/tangent repair, with wrong-gripper acting only as a marginal helper |
 | K4 mean-by-type residual retrieval + margin sweep around 0.20 | No | No | 34.84-35.25% | +5.10-5.51 pp | Margin 0.20 is a local abstention optimum for both typed no-op and source-score priors; 0.15 and 0.25 drop below the plateau |
@@ -102,21 +103,22 @@ Suggested main-table rows:
 17. K4 mean-by-type residual retrieval + source-progress/source-score/source-advantage prior diagnostics
 18. K4 mean-by-type residual retrieval + train-family success bonus diagnostics
 19. K4 mean-by-type residual retrieval + train-neighbor consensus-confidence diagnostics
-20. K4 mean-by-type residual retrieval + no-op-only family diagnostic
-21. K4 mean-by-type residual retrieval + abstention margin fine sweep
-22. Source-progress viability gate diagnostics
-23. K2/K4 task-relative retrieval metric diagnostics
-24. K4 kernel-weighted residual consensus + no-op prior diagnostics
-25. K4 field-softmax residual barycenter + margin diagnostics
-26. K4 mean-by-type residual retrieval + wrong-gripper typed-prior diagnostics
-27. K2 broad tangent ray-search
-28. Residual-tangent distillation policy
-29. Residual+Gaussian hybrid, K32 sigma0.35
-30. Lattice, near-miss only
-31. Lattice, no expert
-32. Lattice, no expert + policy baseline candidate
-33. Lattice, full
-34. Oracle ceiling
+20. K4 repair-tangent residual transport diagnostics
+21. K4 mean-by-type residual retrieval + no-op-only family diagnostic
+22. K4 mean-by-type residual retrieval + abstention margin fine sweep
+23. Source-progress viability gate diagnostics
+24. K2/K4 task-relative retrieval metric diagnostics
+25. K4 kernel-weighted residual consensus + no-op prior diagnostics
+26. K4 field-softmax residual barycenter + margin diagnostics
+27. K4 mean-by-type residual retrieval + wrong-gripper typed-prior diagnostics
+28. K2 broad tangent ray-search
+29. Residual-tangent distillation policy
+30. Residual+Gaussian hybrid, K32 sigma0.35
+31. Lattice, near-miss only
+32. Lattice, no expert
+33. Lattice, no expert + policy baseline candidate
+34. Lattice, full
+35. Oracle ceiling
 
 Suggested claim:
 
@@ -131,7 +133,10 @@ Suggested claim:
 > suggesting transferable residuals need not beat the expert anchor in their source
 > state. Continuous train-family success priors likewise tie or drop rather than
 > explain the top row. A train-neighbor consensus penalty is also negative/near-tie,
-> suggesting the current field already performs most of the useful abstention. Ungated KNN residual
+> suggesting the current field already performs most of the useful abstention. Repair-tangent
+> transport is negative, showing that simply reversing train failures into
+> near-miss-to-expert correction vectors is not the missing deployment proposal.
+> Ungated KNN residual
 > retrieval, field-gradient ascent, broader non-expert BC targets, field-teacher/tangent distillation, z-score/task-relative retrieval,
 > train-family reliability priors, policy-relative anchoring, residual+Gaussian hybrids,
 > source-progress/source-advantage viability gates, no-op-only family masking, off-peak abstention margins, overly strong train-outcome priors, tangent consensus, kernel-weighted tangent interpolation, field-softmax tangent barycenters, tangent ray-search, wrong-gripper typed priors, and same-state policy-baseline fallback fail to improve the main rows.

results/paper_story_memo.md

@@ -33,6 +33,7 @@ when queried on proposal geometry that matches those local counterfactuals.
 | Source-advantage priors/gates are too brittle | source-advantage bonuses 0.02/0.05 reach 35.13%; no-op+advantage bonus reaches 35.30%; positive-advantage gates reach 35.13% with or without no-op prior | Negative diagnostic: useful transferable tangents need not beat the expert anchor in their own source state |
 | Continuous train-family success priors do not add the gain | scale-grid family-success bonuses 0.02/0.03/0.05 reach 35.25%; no-op+family-success 0.02 ties the best at 35.42% | Negative/tie diagnostic: train terminal success is not the right confidence signal for transferred tangents |
 | Train-neighbor consensus confidence does not improve the top row | consensus-only 0.05 reaches 35.19%; no-op+consensus penalties 0.02/0.05/0.10 reach 35.36% | Negative/near-tie diagnostic: residual dispersion is a plausible confidence signal, but the field+margin already abstains better |
+| Repair-tangent transport is not the missing clean proposal | reversing residual direction to build near-miss/failure-to-expert tangents reaches only 34.14-34.43%, below the 35.42% scale-grid no-op row | Negative diagnostic: the failure-to-expert vector hypothesis is cleaner than a new prior, but does not explain the gap |
 | Kernel-weighted tangent interpolation does not beat equal consensus | K4 kernel-weighted residual consensus reaches 34.96%; with no-op prior and scales 0.35/0.40/0.45 it reaches 35.13%/35.19%/35.19%, below the 35.25% mean-consensus plateau | Negative/near-tie diagnostic |
 | Field-conditioned tangent barycenters identify good sparse corrections but do not close the proposal gap | K4 field-softmax transport reaches 34.96%; with no-op prior and margins 0.10/0.05/0.00 it reaches 35.19%/35.07%/34.84%. Selected aggregate residuals are high-value (up to 60.00% success), but selecting more of them degrades the global row | Negative/near-tie diagnostic |
 | 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 scale-grid mean-consensus row at 35.42% | Near-tie/negative diagnostic |
@@ -80,25 +81,26 @@ clean proposal result, the intended main rows are:
 21. K4 mean-by-type tangent consensus + train-source advantage prior/gate: 35.13% at bonuses 0.02/0.05; 35.30% with no-op+advantage; 35.13% with positive-advantage gates
 22. K4 mean-by-type tangent consensus + train-family success bonus: 35.25% alone; 35.42% with no-op bonus 0.03
 23. K4 mean-by-type tangent consensus + train-neighbor consensus penalty: 35.19% alone; 35.36% with no-op bonus 0.03
-24. K4 mean-by-type tangent consensus, no-op-only residuals: 35.19% with either no-op bonus 0.03 or source-score bonus 0.02
-25. K4 mean-by-type abstention margin sweep: 35.07% / 35.25% / 34.84% for typed no-op margins 0.15 / 0.20 / 0.25; 34.96% / 35.25% / 34.84% for source-score margins
-26. Source-progress viability gates: 35.19% / 34.96% / 34.72% for thresholds 0.25 / 0.50 / 0.75
-27. K4 kernel-weighted tangent consensus / + no-op prior: 34.96% / 35.19%
-28. K4 field-softmax tangent transport / best margin sweep: 34.96% / 35.19%
-29. Wrong-gripper prior / no-op+wrong-gripper prior: 35.19% / 35.25%
-30. K2 broad tangent ray-search: 34.96%
-31. K1/K2 tight tangent ray-search: 34.84% / 34.84%
-32. K4 tight tangent ray-search: 34.55%
-33. Residual-tangent distillation policy: 28.87%
-34. Z-score residual retrieval: 32.23-32.81%
-35. Task-relative residual retrieval metric: 34.26-34.43%
-36. Train-family reliability prior: 33.28-33.33%
-37. Residual+Gaussian hybrid K32/K64: 31.30% / 30.90%
-38. Lattice, near-miss only: 55.94%
-39. Lattice, no expert: 56.99%
-40. Lattice, no expert + policy baseline candidate: 40.70%
-41. Lattice, full: 69.33%
-42. Oracle ceiling: 86.78%
+24. K4 repair-tangent transport: 34.14-34.43%
+25. K4 mean-by-type tangent consensus, no-op-only residuals: 35.19% with either no-op bonus 0.03 or source-score bonus 0.02
+26. K4 mean-by-type abstention margin sweep: 35.07% / 35.25% / 34.84% for typed no-op margins 0.15 / 0.20 / 0.25; 34.96% / 35.25% / 34.84% for source-score margins
+27. Source-progress viability gates: 35.19% / 34.96% / 34.72% for thresholds 0.25 / 0.50 / 0.75
+28. K4 kernel-weighted tangent consensus / + no-op prior: 34.96% / 35.19%
+29. K4 field-softmax tangent transport / best margin sweep: 34.96% / 35.19%
+30. Wrong-gripper prior / no-op+wrong-gripper prior: 35.19% / 35.25%
+31. K2 broad tangent ray-search: 34.96%
+32. K1/K2 tight tangent ray-search: 34.84% / 34.84%
+33. K4 tight tangent ray-search: 34.55%
+34. Residual-tangent distillation policy: 28.87%
+35. Z-score residual retrieval: 32.23-32.81%
+36. Task-relative residual retrieval metric: 34.26-34.43%
+37. Train-family reliability prior: 33.28-33.33%
+38. Residual+Gaussian hybrid K32/K64: 31.30% / 30.90%
+39. Lattice, near-miss only: 55.94%
+40. Lattice, no expert: 56.99%
+41. Lattice, no expert + policy baseline candidate: 40.70%
+42. Lattice, full: 69.33%
+43. Oracle ceiling: 86.78%
 
 ## Novelty Framing
 
@@ -126,12 +128,13 @@ test-time search. The cleaner novelty is:
 
 ## Job Status
 
-Last checked: `2026-06-29 11:36 UTC`. The K4 mean-by-type scale-grid sweep
+Last checked: `2026-06-29 12:03 UTC`. The K4 mean-by-type scale-grid sweep
 completed and produced a new clean best, 35.42%, while upper/wide,
 minimum-energy, source-advantage, and train-family success-prior follow-ups
 completed without improving it. Consensus-confidence follow-ups also completed
-below the best row. The paper table/paired analysis continue to use the
-scale-grid no-op row as `best_clean_key`.
+below the best row. Repair-tangent follow-ups completed below the best row too,
+with a best repair result of 34.43%. The paper table/paired analysis continue
+to use the scale-grid no-op row as `best_clean_key`.
 
 - `14858328`-`14858333`: completed train-split `field_selected_noexpert_bc5`;
   direct rollout is 26.84%, field-guided best is 27.65%.
@@ -256,6 +259,14 @@ scale-grid no-op row as `best_clean_key`.
   consensus penalties `0.02`, `0.05`, and `0.10` all reach 35.36%, one success
   below the 35.42% best. Summary jobs `14903385`/`14903387`/`14903389`/
   `14903391` and rebuild job `14903392` completed.
+- `14904575`: completed CPU smoke for repair-tangent residual direction
+  (`anchor_minus_candidate`). The smoke wrote valid metadata and selected the
+  expected repair direction under the new CLI/Slurm path.
+- `14904737`/`14904740`/`14904742`/`14904744`: completed repair-tangent GPU
+  arrays. Near-miss-only repair grids reach 34.14-34.38%, and the safe-family
+  repair row reaches 34.43%. Summary jobs `14904738`/`14904741`/`14904743`/
+  `14904745` completed, local paper builders updated the artifacts, and the
+  queued rebuild job `14904803` was canceled after local rebuilds finished.
 - `14894281`: completed the Apptainer unit smoke for the train-source
   progress-viability gate, including the variable residual-count padding check
   (`source_progress_lengths == [3, 3]`).
@@ -353,14 +364,15 @@ scale-grid no-op row as `best_clean_key`.
   claim. The no-op-only family ablation reaches 35.19%, so wrong-gripper
   residuals are a marginal helper rather than the core mechanism. The margin
   fine sweep confirms `0.20` is a local abstention optimum for both typed and
-  measured train-outcome priors. The completed K2/ray-search rows are near-ties
-  that support the
-  sparse-intervention story.
+  measured train-outcome priors. The repair-tangent rows reach only
+  34.14-34.43%, so the missing clean proposal is not simply a transported
+  failure-to-expert correction vector. 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 and task-relative retrieval metrics, source-progress/source-advantage viability gates,
   repaired train-family reliability priors, Gaussian hybrids,
-  field optimization, field-teacher/tangent distillation, policy-relative anchoring, tangent consensus,
+  field optimization, field-teacher/tangent distillation, repair-tangent transport, policy-relative anchoring, tangent consensus,
   kernel-weighted tangent interpolation, field-softmax tangent barycenters,
   wrong-gripper typed priors, and same-state policy-baseline fallback as negative
   or near-tie diagnostics that sharpen the story around local counterfactual

results/paper_table_status.json

@@ -1220,6 +1220,82 @@
       "best_config": null,
       "gain_vs_h16_policy": 0.05623188405797103
     },
+    {
+      "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",
+      "path": "h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_repair_nearmiss_k4_grid025035050_margin0p20_summary.json",
+      "clean_deployment": "yes",
+      "same_state_proposals": "no",
+      "expert_proposal": "no",
+      "story_role": "deployment-clean corrective tangent transport from train near-misses back toward expert actions",
+      "fallback_success": null,
+      "pending_job": "14904737/14904738",
+      "path_exists": true,
+      "status": "complete",
+      "success": 0.34318840579710147,
+      "std_success": 0.013508614722007019,
+      "completed_seeds": null,
+      "num_completed": 3,
+      "best_config": null,
+      "gain_vs_h16_policy": 0.0457971014492754
+    },
+    {
+      "key": "retrieval_repair_nearmiss_k4_grid035050075_margin020",
+      "label": "K4 near-miss-to-expert repair tangent, scales 0.35/0.50/0.75, margin 0.20",
+      "path": "h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_repair_nearmiss_k4_grid035050075_margin0p20_summary.json",
+      "clean_deployment": "yes",
+      "same_state_proposals": "no",
+      "expert_proposal": "no",
+      "story_role": "repair-tangent scale diagnostic for near-miss counterfactual geometry",
+      "fallback_success": null,
+      "pending_job": "14904740/14904741",
+      "path_exists": true,
+      "status": "complete",
+      "success": 0.343768115942029,
+      "std_success": 0.014994222790002973,
+      "completed_seeds": null,
+      "num_completed": 3,
+      "best_config": null,
+      "gain_vs_h16_policy": 0.04637681159420293
+    },
+    {
+      "key": "retrieval_repair_nearmiss_k4_grid025035050_margin010",
+      "label": "K4 near-miss-to-expert repair tangent, scales 0.25/0.35/0.50, margin 0.10",
+      "path": "h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_repair_nearmiss_k4_grid025035050_margin0p10_summary.json",
+      "clean_deployment": "yes",
+      "same_state_proposals": "no",
+      "expert_proposal": "no",
+      "story_role": "repair-tangent abstention diagnostic for near-miss counterfactual geometry",
+      "fallback_success": null,
+      "pending_job": "14904742/14904743",
+      "path_exists": true,
+      "status": "complete",
+      "success": 0.34144927536231884,
+      "std_success": 0.014791131388026771,
+      "completed_seeds": null,
+      "num_completed": 3,
+      "best_config": null,
+      "gain_vs_h16_policy": 0.04405797101449277
+    },
+    {
+      "key": "retrieval_repair_safe_k4_grid025035050_margin020",
+      "label": "K4 safe-family-to-expert repair tangent, scales 0.25/0.35/0.50, margin 0.20",
+      "path": "h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_repair_safe_k4_grid025035050_margin0p20_summary.json",
+      "clean_deployment": "yes",
+      "same_state_proposals": "no",
+      "expert_proposal": "no",
+      "story_role": "repair-tangent family diagnostic including near-miss, no-op, and wrong-gripper corrections",
+      "fallback_success": null,
+      "pending_job": "14904744/14904745",
+      "path_exists": true,
+      "status": "complete",
+      "success": 0.3443478260869565,
+      "std_success": 0.012541047914657353,
+      "completed_seeds": null,
+      "num_completed": 3,
+      "best_config": null,
+      "gain_vs_h16_policy": 0.04695652173913045
+    },
     {
       "key": "retrieval_residual_k4_mean_grid035040045_noopbonus003_l2penalty005",
       "label": "K4 mean-by-type residual retrieval, scales 0.35/0.40/0.45, margin 0.20, no-op bonus 0.03, action L2 penalty 0.05",

results/paper_table_status.md

@@ -67,6 +67,10 @@ Baseline h=16 policy: 29.74%
 | retrieval_residual_k4_mean_grid035040045_noopbonus003_consensus002 | K4 mean-by-type residual retrieval, scales 0.35/0.40/0.45, margin 0.20, no-op bonus 0.03, consensus penalty 0.02 | complete | 35.36% | +5.62 pp | yes | no | no | train-neighbor tangent-consensus confidence on the current best typed prior |
 | retrieval_residual_k4_mean_grid035040045_noopbonus003_consensus005 | K4 mean-by-type residual retrieval, scales 0.35/0.40/0.45, margin 0.20, no-op bonus 0.03, consensus penalty 0.05 | complete | 35.36% | +5.62 pp | yes | no | no | train-neighbor tangent-consensus confidence on the current best typed prior |
 | 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_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 |
+| retrieval_repair_safe_k4_grid025035050_margin020 | K4 safe-family-to-expert repair tangent, scales 0.25/0.35/0.50, margin 0.20 | complete | 34.43% | +4.70 pp | yes | no | no | repair-tangent family diagnostic including near-miss, no-op, and wrong-gripper corrections |
 | retrieval_residual_k4_mean_grid035040045_noopbonus003_l2penalty005 | K4 mean-by-type residual retrieval, scales 0.35/0.40/0.45, margin 0.20, no-op bonus 0.03, action L2 penalty 0.05 | complete | 35.42% | +5.68 pp | yes | no | no | minimum-energy tangent diagnostic for sparse mean-consensus residual transport |
 | retrieval_residual_k4_mean_grid035040045_noopbonus003_l2penalty010 | K4 mean-by-type residual retrieval, scales 0.35/0.40/0.45, margin 0.20, no-op bonus 0.03, action L2 penalty 0.10 | complete | 35.36% | +5.62 pp | yes | no | no | minimum-energy tangent diagnostic for sparse mean-consensus residual transport |
 | retrieval_residual_k4_mean_grid035040045_noopbonus003_l2penalty020 | K4 mean-by-type residual retrieval, scales 0.35/0.40/0.45, margin 0.20, no-op bonus 0.03, action L2 penalty 0.20 | complete | 35.36% | +5.62 pp | yes | no | no | minimum-energy tangent diagnostic for sparse mean-consensus residual transport |