Auto-sync: 2026-06-29 02:31:09 (part 3)

results/paper_analysis.json

@@ -1,6 +1,6 @@
 {
   "best_clean_key": "residual_k4_consensus_noopbonus003",
-  "generated_utc": "2026-06-29T06:15:24+00:00",
+  "generated_utc": "2026-06-29T06:28:26+00:00",
   "mechanism_gap": {
     "best_clean_vs_direct_same_ckpt": 0.06956521739130428,
     "best_clean_vs_h16": 0.05507246376811592,
@@ -349,6 +349,314 @@
       "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_margin015_noopbonus003": {
+      "ci95_success": 0.027439420289855042,
+      "label": "K4 mean-by-type tangent consensus, margin 0.15, no-op bonus 0.03",
+      "mean_action_mse_to_best": 0.3972578248076573,
+      "mean_progress": 0.5666977553927132,
+      "mean_success": 0.3507246376811594,
+      "num_completed": 3,
+      "per_task_success": {
+        "LiftPegUpright-v1": {
+          "mean_num_groups": 102.0,
+          "mean_success": 0.2599364286510761,
+          "std_success": 0.054567202237336666
+        },
+        "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": 0.751397595193677,
+          "std_success": 0.07098946144071037
+        },
+        "StackCube-v1": {
+          "mean_num_groups": 93.33333333333333,
+          "mean_success": 0.19740999740999743,
+          "std_success": 0.07397934256534576
+        }
+      },
+      "seed_action_mse_to_best": {
+        "0": 0.3823875710076612,
+        "1": 0.3907143801985228,
+        "2": 0.4186715232167879
+      },
+      "seed_progress": {
+        "0": 0.5523235452620555,
+        "1": 0.5681803968581168,
+        "2": 0.5795893240579671
+      },
+      "seed_success": {
+        "0": 0.3443478260869565,
+        "1": 0.3443478260869565,
+        "2": 0.3634782608695652
+      },
+      "selected_candidate_type_counts": {
+        "retrieval_residual_policy_residual": 1598,
+        "retrieval_residual_residual_no_op": 90,
+        "retrieval_residual_residual_wrong_gripper": 37
+      },
+      "selected_residual_scale_counts": {
+        "0.4": 1725
+      },
+      "selected_type_outcomes": {
+        "retrieval_residual_policy_residual": {
+          "count": 1598.0,
+          "mean_progress": 0.5609213063446945,
+          "success_rate": 0.344180225281602
+        },
+        "retrieval_residual_residual_no_op": {
+          "count": 90.0,
+          "mean_progress": 0.6413994556499852,
+          "success_rate": 0.4222222222222222
+        },
+        "retrieval_residual_residual_wrong_gripper": {
+          "count": 37.0,
+          "mean_progress": 0.63447106770567,
+          "success_rate": 0.4594594594594595
+        }
+      },
+      "source": "results/h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_k4s040_safe_margin0p15_mean_by_type_noopbonus0p03_summary.json",
+      "std_success": 0.011044961671453697
+    },
+    "residual_k4_consensus_margin015_srcscorebonus002": {
+      "ci95_success": 0.02628111194117414,
+      "label": "K4 mean-by-type tangent consensus, margin 0.15, source-score bonus 0.02",
+      "mean_action_mse_to_best": 0.39648515512518906,
+      "mean_progress": 0.5663530205259472,
+      "mean_success": 0.34956521739130436,
+      "num_completed": 3,
+      "per_task_success": {
+        "LiftPegUpright-v1": {
+          "mean_num_groups": 102.0,
+          "mean_success": 0.2599364286510761,
+          "std_success": 0.054567202237336666
+        },
+        "PickCube-v1": {
+          "mean_num_groups": 196.66666666666666,
+          "mean_success": 0.31930509104422145,
+          "std_success": 0.027120739980041123
+        },
+        "PullCube-v1": {
+          "mean_num_groups": 81.0,
+          "mean_success": 0.20542391331865018,
+          "std_success": 0.007170913229566424
+        },
+        "PushCube-v1": {
+          "mean_num_groups": 102.0,
+          "mean_success": 0.7445511658698934,
+          "std_success": 0.07505800752379664
+        },
+        "StackCube-v1": {
+          "mean_num_groups": 93.33333333333333,
+          "mean_success": 0.19740999740999743,
+          "std_success": 0.07397934256534576
+        }
+      },
+      "seed_action_mse_to_best": {
+        "0": 0.3825673672265333,
+        "1": 0.38947549535042564,
+        "2": 0.4174126027986083
+      },
+      "seed_progress": {
+        "0": 0.5518181501450422,
+        "1": 0.5673385079528974,
+        "2": 0.579902403479902
+      },
+      "seed_success": {
+        "0": 0.3443478260869565,
+        "1": 0.3426086956521739,
+        "2": 0.3617391304347826
+      },
+      "selected_candidate_type_counts": {
+        "retrieval_residual_policy_residual": 1604,
+        "retrieval_residual_residual_no_op": 78,
+        "retrieval_residual_residual_wrong_gripper": 43
+      },
+      "selected_residual_scale_counts": {
+        "0.4": 1725
+      },
+      "selected_type_outcomes": {
+        "retrieval_residual_policy_residual": {
+          "count": 1604.0,
+          "mean_progress": 0.5570878207019525,
+          "success_rate": 0.33915211970074816
+        },
+        "retrieval_residual_residual_no_op": {
+          "count": 78.0,
+          "mean_progress": 0.7125367106726537,
+          "success_rate": 0.48717948717948717
+        },
+        "retrieval_residual_residual_wrong_gripper": {
+          "count": 43.0,
+          "mean_progress": 0.6467961062525593,
+          "success_rate": 0.4883720930232558
+        }
+      },
+      "source": "results/h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_k4s040_safe_margin0p15_mean_by_type_srcscorebonus0p02_summary.json",
+      "std_success": 0.010578717443996917
+    },
+    "residual_k4_consensus_margin025_noopbonus003": {
+      "ci95_success": 0.034922898550724606,
+      "label": "K4 mean-by-type tangent consensus, margin 0.25, no-op bonus 0.03",
+      "mean_action_mse_to_best": 0.39512178331424574,
+      "mean_progress": 0.5640931326198588,
+      "mean_success": 0.3484057971014493,
+      "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": 0.31741255926038536,
+          "std_success": 0.03084062759150011
+        },
+        "PullCube-v1": {
+          "mean_num_groups": 81.0,
+          "mean_success": 0.19226601858180806,
+          "std_success": 0.029812623771358626
+        },
+        "PushCube-v1": {
+          "mean_num_groups": 102.0,
+          "mean_success": 0.7445511658698934,
+          "std_success": 0.07505800752379664
+        },
+        "StackCube-v1": {
+          "mean_num_groups": 93.33333333333333,
+          "mean_success": 0.2011137011137011,
+          "std_success": 0.07105663963231741
+        }
+      },
+      "seed_action_mse_to_best": {
+        "0": 0.3818597935172527,
+        "1": 0.3877675680119706,
+        "2": 0.415737988413514
+      },
+      "seed_progress": {
+        "0": 0.5509842345887876,
+        "1": 0.5622477454109036,
+        "2": 0.5790474178598853
+      },
+      "seed_success": {
+        "0": 0.34608695652173915,
+        "1": 0.3356521739130435,
+        "2": 0.3634782608695652
+      },
+      "selected_candidate_type_counts": {
+        "retrieval_residual_policy_residual": 1681,
+        "retrieval_residual_residual_no_op": 29,
+        "retrieval_residual_residual_wrong_gripper": 15
+      },
+      "selected_residual_scale_counts": {
+        "0.4": 1725
+      },
+      "selected_type_outcomes": {
+        "retrieval_residual_policy_residual": {
+          "count": 1681.0,
+          "mean_progress": 0.5609548186565506,
+          "success_rate": 0.3438429506246282
+        },
+        "retrieval_residual_residual_no_op": {
+          "count": 29.0,
+          "mean_progress": 0.7565103556061613,
+          "success_rate": 0.5862068965517241
+        },
+        "retrieval_residual_residual_wrong_gripper": {
+          "count": 15.0,
+          "mean_progress": 0.5437868863344193,
+          "success_rate": 0.4
+        }
+      },
+      "source": "results/h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_k4s040_safe_margin0p25_mean_by_type_noopbonus0p03_summary.json",
+      "std_success": 0.014057223945486528
+    },
+    "residual_k4_consensus_margin025_srcscorebonus002": {
+      "ci95_success": 0.034922898550724606,
+      "label": "K4 mean-by-type tangent consensus, margin 0.25, source-score bonus 0.02",
+      "mean_action_mse_to_best": 0.39499614254531007,
+      "mean_progress": 0.5641164365458499,
+      "mean_success": 0.3484057971014493,
+      "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": 0.31741255926038536,
+          "std_success": 0.03084062759150011
+        },
+        "PullCube-v1": {
+          "mean_num_groups": 81.0,
+          "mean_success": 0.19226601858180806,
+          "std_success": 0.029812623771358626
+        },
+        "PushCube-v1": {
+          "mean_num_groups": 102.0,
+          "mean_success": 0.7445511658698934,
+          "std_success": 0.07505800752379664
+        },
+        "StackCube-v1": {
+          "mean_num_groups": 93.33333333333333,
+          "mean_success": 0.2011137011137011,
+          "std_success": 0.07105663963231741
+        }
+      },
+      "seed_action_mse_to_best": {
+        "0": 0.3809117384157751,
+        "1": 0.38808461398048244,
+        "2": 0.41599207523967263
+      },
+      "seed_progress": {
+        "0": 0.5509842345887876,
+        "1": 0.5623445302021244,
+        "2": 0.5790205448466377
+      },
+      "seed_success": {
+        "0": 0.34608695652173915,
+        "1": 0.3356521739130435,
+        "2": 0.3634782608695652
+      },
+      "selected_candidate_type_counts": {
+        "retrieval_residual_policy_residual": 1684,
+        "retrieval_residual_residual_no_op": 24,
+        "retrieval_residual_residual_wrong_gripper": 17
+      },
+      "selected_residual_scale_counts": {
+        "0.4": 1725
+      },
+      "selected_type_outcomes": {
+        "retrieval_residual_policy_residual": {
+          "count": 1684.0,
+          "mean_progress": 0.560429067915488,
+          "success_rate": 0.34323040380047504
+        },
+        "retrieval_residual_residual_no_op": {
+          "count": 24.0,
+          "mean_progress": 0.7992291407038769,
+          "success_rate": 0.625
+        },
+        "retrieval_residual_residual_wrong_gripper": {
+          "count": 17.0,
+          "mean_progress": 0.5974590173538994,
+          "success_rate": 0.47058823529411764
+        }
+      },
+      "source": "results/h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_k4s040_safe_margin0p25_mean_by_type_srcscorebonus0p02_summary.json",
+      "std_success": 0.014057223945486528
+    },
     "residual_k4_consensus_noop0025_wg002": {
       "ci95_success": 0.03518915067855206,
       "label": "K4 mean-by-type tangent consensus, no-op 0.025 + wrong-gripper 0.02",

results/paper_analysis.md

@@ -1,6 +1,6 @@
 # Paper Analysis
 
-Generated: `2026-06-29T06:15:24+00:00`
+Generated: `2026-06-29T06:28:26+00:00`
 
 ## Main Seed Statistics
 
@@ -24,6 +24,10 @@ Generated: `2026-06-29T06:15:24+00:00`
 | residual_k8_fieldsoftmax_grid_noopbonus003 | K8 field-softmax tangent transport, scales 0.35/0.40/0.45, no-op bonus 0.03 | 3 | 34.84% +/- 1.35 | +/- 3.36 | 56.55% | 0.397 | +5.10 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_noopbonus003_srcprog025 | K4 mean-by-type tangent consensus, no-op bonus 0.03, source progress >= 0.25 | 3 | 35.19% +/- 1.32 | +/- 3.27 | 56.69% | 0.396 | +5.45 pp |
+| residual_k4_consensus_margin015_noopbonus003 | K4 mean-by-type tangent consensus, margin 0.15, no-op bonus 0.03 | 3 | 35.07% +/- 1.10 | +/- 2.74 | 56.67% | 0.397 | +5.33 pp |
+| residual_k4_consensus_margin025_noopbonus003 | K4 mean-by-type tangent consensus, margin 0.25, no-op bonus 0.03 | 3 | 34.84% +/- 1.41 | +/- 3.49 | 56.41% | 0.395 | +5.10 pp |
+| residual_k4_consensus_margin015_srcscorebonus002 | K4 mean-by-type tangent consensus, margin 0.15, source-score bonus 0.02 | 3 | 34.96% +/- 1.06 | +/- 2.63 | 56.64% | 0.396 | +5.22 pp |
+| residual_k4_consensus_margin025_srcscorebonus002 | K4 mean-by-type tangent consensus, margin 0.25, source-score bonus 0.02 | 3 | 34.84% +/- 1.41 | +/- 3.49 | 56.41% | 0.395 | +5.10 pp |
 | residual_k4_consensus_nooponly_noopbonus003 | K4 mean-by-type tangent consensus, no-op-only residuals, no-op bonus 0.03 | 3 | 35.19% +/- 1.18 | +/- 2.94 | 56.57% | 0.394 | +5.45 pp |
 | residual_k4_consensus_nooponly_srcscorebonus002 | K4 mean-by-type tangent consensus, no-op-only residuals, source-score bonus 0.02 | 3 | 35.19% +/- 1.18 | +/- 2.94 | 56.55% | 0.394 | +5.45 pp |
 | residual_k4_consensus_noopbonus003_srcprog050 | K4 mean-by-type tangent consensus, no-op bonus 0.03, source progress >= 0.50 | 3 | 34.96% +/- 1.55 | +/- 3.84 | 56.53% | 0.396 | +5.22 pp |

results/paper_core_results.md

@@ -46,6 +46,7 @@ and the remaining clean-to-same-state proposal gap is `+21.74 pp`.
 | K4 mean-by-type residual retrieval + source-score prior 0.015/0.020 | No | No | 35.25% | +5.51 pp | Full train reward score, including terminal success, also replaces the typed prior without improving it |
 | 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 35.25% 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 |
 | K4 mean-by-type residual retrieval + no-op prior + source-progress gate | No | No | 34.72-35.19% | +4.99-5.45 pp | Train-source progress viability is a near-tie/negative gate; soft threshold 0.25 is closest, while stricter thresholds over-abstain below the no-op plateau |
 | K4 task-relative mean-by-type residual retrieval + no-op prior 0.03 | No | No | 34.43% | +4.70 pp | Task-relative target/reference pose retrieval underperforms the raw-metric no-op plateau |
 | K4 kernel-weighted residual consensus + no-op prior 0.03 | No | No | 35.13-35.19% | +5.39-5.45 pp | Distance-weighted tangent interpolation is plausible but does not beat equal mean-consensus no-op plateau |
@@ -89,19 +90,20 @@ Suggested main-table rows:
 13. K4 mean-by-type residual retrieval + no-op prior plateau, canonical 0.03
 14. K4 mean-by-type residual retrieval + source-progress/source-score prior diagnostics
 15. K4 mean-by-type residual retrieval + no-op-only family diagnostic
-16. Source-progress viability gate diagnostics
-17. K2/K4 task-relative retrieval metric diagnostics
-18. K4 kernel-weighted residual consensus + no-op prior diagnostics
-19. K4 field-softmax residual barycenter + margin diagnostics
-20. K4 mean-by-type residual retrieval + wrong-gripper typed-prior diagnostics
-21. K2 broad tangent ray-search
-22. Residual-tangent distillation policy
-23. Residual+Gaussian hybrid, K32 sigma0.35
-24. Lattice, near-miss only
-25. Lattice, no expert
-26. Lattice, no expert + policy baseline candidate
-27. Lattice, full
-28. Oracle ceiling
+16. K4 mean-by-type residual retrieval + abstention margin fine sweep
+17. Source-progress viability gate diagnostics
+18. K2/K4 task-relative retrieval metric diagnostics
+19. K4 kernel-weighted residual consensus + no-op prior diagnostics
+20. K4 field-softmax residual barycenter + margin diagnostics
+21. K4 mean-by-type residual retrieval + wrong-gripper typed-prior diagnostics
+22. K2 broad tangent ray-search
+23. Residual-tangent distillation policy
+24. Residual+Gaussian hybrid, K32 sigma0.35
+25. Lattice, near-miss only
+26. Lattice, no expert
+27. Lattice, no expert + policy baseline candidate
+28. Lattice, full
+29. Oracle ceiling
 
 Suggested claim:
 
@@ -111,7 +113,7 @@ Suggested claim:
 > progress/reward-score 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/task-relative retrieval,
 > train-family reliability priors, policy-relative anchoring, residual+Gaussian hybrids,
-> source-progress viability gates, no-op-only family masking, 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.
+> source-progress 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.
 > The large effect appears only when the field is queried on
 > same-state intervention proposals, and the mechanism is isolated to local near-miss
 > counterfactual geometry.

results/paper_story_memo.md

@@ -41,6 +41,7 @@ when queried on proposal geometry that matches those local counterfactuals.
 | Train-source progress viability is too blunt a residual gate | source-progress thresholds 0.25/0.50/0.75 reach 35.19%/34.96%/34.72%, below the unfiltered no-op plateau at 35.25% | Negative/near-tie diagnostic |
 | Continuous train-source progress prior can replace the typed no-op prior but not improve it | source-progress bonus 0.03 ties the 35.25% best exactly; bonus 0.05 drops to 35.13% | Cleaner tie diagnostic |
 | Full train-source reward-score prior also ties but does not improve the clean best | source-score bonuses 0.015/0.020 tie 35.25%; 0.025 drops to 35.19% | Cleaner tie diagnostic |
+| Advantage margin 0.20 is a local optimum for K4 tangent consensus | no-op prior margins 0.15/0.20/0.25 reach 35.07%/35.25%/34.84%; source-score prior margins reach 34.96%/35.25%/34.84% | Abstention plateau sharpened |
 | 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 |
 | Residual-tangent distillation does not solve clean proposal generation | aligned allmap tangent student reaches 28.87% despite low pseudo-target BC loss | Negative diagnostic |
 | Policy-relative residual anchoring does not improve the bridge | policy-anchor safe residual transport ties 33.74% rather than improving expert-anchor residuals | Negative diagnostic |
@@ -69,14 +70,15 @@ clean proposal result, the intended main rows are:
 16. K4 mean-by-type tangent consensus + train-source progress prior: 35.25% at bonus 0.03; 35.13% at bonus 0.05
 17. K4 mean-by-type tangent consensus + train-source reward-score prior: 35.25% at bonuses 0.015/0.020; 35.19% at 0.025
 18. 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
-19. Source-progress viability gates: 35.19% / 34.96% / 34.72% for thresholds 0.25 / 0.50 / 0.75
-20. K4 kernel-weighted tangent consensus / + no-op prior: 34.96% / 35.19%
-21. K4 field-softmax tangent transport / best margin sweep: 34.96% / 35.19%
-22. Wrong-gripper prior / no-op+wrong-gripper prior: 35.19% / 35.25%
-23. K2 broad tangent ray-search: 34.96%
-24. K1/K2 tight tangent ray-search: 34.84% / 34.84%
-25. K4 tight tangent ray-search: 34.55%
-26. Residual-tangent distillation policy: 28.87%
+19. 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
+20. Source-progress viability gates: 35.19% / 34.96% / 34.72% for thresholds 0.25 / 0.50 / 0.75
+21. K4 kernel-weighted tangent consensus / + no-op prior: 34.96% / 35.19%
+22. K4 field-softmax tangent transport / best margin sweep: 34.96% / 35.19%
+23. Wrong-gripper prior / no-op+wrong-gripper prior: 35.19% / 35.25%
+24. K2 broad tangent ray-search: 34.96%
+25. K1/K2 tight tangent ray-search: 34.84% / 34.84%
+26. K4 tight tangent ray-search: 34.55%
+27. Residual-tangent distillation policy: 28.87%
 26. Z-score residual retrieval: 32.23-32.81%
 27. Task-relative residual retrieval metric: 34.26-34.43%
 28. Train-family reliability prior: 33.28-33.33%
@@ -113,9 +115,9 @@ test-time search. The cleaner novelty is:
 
 ## Job Status
 
-Last checked: `2026-06-29 06:13 UTC`. The no-op-only residual-family ablation
-completed after CPU smokes and GPU arrays, and the paper table/paired analysis
-now include both full safe-family and no-op-only train-outcome prior rows.
+Last checked: `2026-06-29 06:29 UTC`. The K4 abstention-margin fine sweep
+completed, and the paper table/paired analysis now include margin `0.15`,
+`0.20`, and `0.25` rows for both typed no-op and source-score priors.
 
 - `14858328`-`14858333`: completed train-split `field_selected_noexpert_bc5`;
   direct rollout is 26.84%, field-guided best is 27.65%.
@@ -262,6 +264,11 @@ now include both full safe-family and no-op-only train-outcome prior rows.
   35.25% safe-family plateau. Summary jobs `14897565`/`14897566` and rebuild
   job `14897567` were submitted; local summary/analysis rebuilds were also run
   while the CPU summary job was still pending.
+- `14897841`/`14897842`/`14897843`/`14897844`: completed K4 mean-by-type
+  abstention-margin fine sweeps around the best margin `0.20`. With no-op bonus
+  `0.03`, margins `0.15`/`0.20`/`0.25` reach 35.07%/35.25%/34.84%. With
+  source-score bonus `0.02`, they reach 34.96%/35.25%/34.84%. Summary jobs
+  `14897845`-`14897848` and rebuild job `14897849` 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
@@ -287,9 +294,11 @@ now include both full safe-family and no-op-only train-outcome prior rows.
   continuous train-source progress prior at bonus 0.03 ties this result without
   a hand typed no-op prior; train-source reward-score priors at 0.015/0.020 also
   tie it. The no-op-only family ablation reaches 35.19%, so wrong-gripper
-  residuals are a marginal helper rather than the core mechanism. This is a
-  cleaner story hook but still not a higher SOTA number. The completed
-  K2/ray-search rows are near-ties that support the sparse-intervention story.
+  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. This is a cleaner story hook but still not a
+  higher SOTA number. 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 viability gates,

results/paper_table_status.json

@@ -859,6 +859,82 @@
       "best_config": null,
       "gain_vs_h16_policy": 0.05449275362318845
     },
+    {
+      "key": "retrieval_residual_k4_mean_margin015_noopbonus003",
+      "label": "K4 mean-by-type residual retrieval, scale 0.40, margin 0.15, no-op bonus 0.03",
+      "path": "h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_k4s040_safe_margin0p15_mean_by_type_noopbonus0p03_summary.json",
+      "clean_deployment": "yes",
+      "same_state_proposals": "no",
+      "expert_proposal": "no",
+      "story_role": "advantage-abstention margin fine sweep for sparse residual transport",
+      "fallback_success": null,
+      "pending_job": "14897841/14897845",
+      "path_exists": true,
+      "status": "complete",
+      "success": 0.3507246376811594,
+      "std_success": 0.011044961671453697,
+      "completed_seeds": null,
+      "num_completed": 3,
+      "best_config": null,
+      "gain_vs_h16_policy": 0.053333333333333344
+    },
+    {
+      "key": "retrieval_residual_k4_mean_margin025_noopbonus003",
+      "label": "K4 mean-by-type residual retrieval, scale 0.40, margin 0.25, no-op bonus 0.03",
+      "path": "h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_k4s040_safe_margin0p25_mean_by_type_noopbonus0p03_summary.json",
+      "clean_deployment": "yes",
+      "same_state_proposals": "no",
+      "expert_proposal": "no",
+      "story_role": "advantage-abstention margin fine sweep for sparse residual transport",
+      "fallback_success": null,
+      "pending_job": "14897842/14897846",
+      "path_exists": true,
+      "status": "complete",
+      "success": 0.34840579710144925,
+      "std_success": 0.01405722394548653,
+      "completed_seeds": null,
+      "num_completed": 3,
+      "best_config": null,
+      "gain_vs_h16_policy": 0.051014492753623186
+    },
+    {
+      "key": "retrieval_residual_k4_mean_margin015_srcscorebonus002",
+      "label": "K4 mean-by-type residual retrieval, scale 0.40, margin 0.15, source-score bonus 0.02",
+      "path": "h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_k4s040_safe_margin0p15_mean_by_type_srcscorebonus0p02_summary.json",
+      "clean_deployment": "yes",
+      "same_state_proposals": "no",
+      "expert_proposal": "no",
+      "story_role": "advantage-abstention margin fine sweep with measured train-source prior",
+      "fallback_success": null,
+      "pending_job": "14897843/14897847",
+      "path_exists": true,
+      "status": "complete",
+      "success": 0.34956521739130436,
+      "std_success": 0.010578717443996917,
+      "completed_seeds": null,
+      "num_completed": 3,
+      "best_config": null,
+      "gain_vs_h16_policy": 0.05217391304347829
+    },
+    {
+      "key": "retrieval_residual_k4_mean_margin025_srcscorebonus002",
+      "label": "K4 mean-by-type residual retrieval, scale 0.40, margin 0.25, source-score bonus 0.02",
+      "path": "h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_k4s040_safe_margin0p25_mean_by_type_srcscorebonus0p02_summary.json",
+      "clean_deployment": "yes",
+      "same_state_proposals": "no",
+      "expert_proposal": "no",
+      "story_role": "advantage-abstention margin fine sweep with measured train-source prior",
+      "fallback_success": null,
+      "pending_job": "14897844/14897848",
+      "path_exists": true,
+      "status": "complete",
+      "success": 0.34840579710144925,
+      "std_success": 0.01405722394548653,
+      "completed_seeds": null,
+      "num_completed": 3,
+      "best_config": null,
+      "gain_vs_h16_policy": 0.051014492753623186
+    },
     {
       "key": "retrieval_residual_k4_mean_nooponly_noopbonus003",
       "label": "K4 mean-by-type residual retrieval, scale 0.40, margin 0.20, no-op-only residuals, no-op bonus 0.03",

results/paper_table_status.md

@@ -48,6 +48,10 @@ Baseline h=16 policy: 29.74%
 | retrieval_residual_k8_fieldsoftmax_grid_noopbonus003 | K8 field-softmax residual transport, safe residuals, scales 0.35/0.40/0.45, margin 0.20, no-op residual bonus 0.03 | complete | 34.84% | +5.10 pp | yes | no | no | field-conditioned tangent transport neighborhood scaling |
 | 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_noopbonus003_srcprog025 | K4 mean-by-type residual retrieval, scale 0.40, margin 0.20, no-op bonus 0.03, source progress >= 0.25 | complete | 35.19% | +5.45 pp | yes | no | no | soft train-source viability gate for sparse residual transport |
+| retrieval_residual_k4_mean_margin015_noopbonus003 | K4 mean-by-type residual retrieval, scale 0.40, margin 0.15, no-op bonus 0.03 | complete | 35.07% | +5.33 pp | yes | no | no | advantage-abstention margin fine sweep for sparse residual transport |
+| retrieval_residual_k4_mean_margin025_noopbonus003 | K4 mean-by-type residual retrieval, scale 0.40, margin 0.25, no-op bonus 0.03 | complete | 34.84% | +5.10 pp | yes | no | no | advantage-abstention margin fine sweep for sparse residual transport |
+| retrieval_residual_k4_mean_margin015_srcscorebonus002 | K4 mean-by-type residual retrieval, scale 0.40, margin 0.15, source-score bonus 0.02 | complete | 34.96% | +5.22 pp | yes | no | no | advantage-abstention margin fine sweep with measured train-source prior |
+| retrieval_residual_k4_mean_margin025_srcscorebonus002 | K4 mean-by-type residual retrieval, scale 0.40, margin 0.25, source-score bonus 0.02 | complete | 34.84% | +5.10 pp | yes | no | no | advantage-abstention margin fine sweep with measured train-source prior |
 | retrieval_residual_k4_mean_nooponly_noopbonus003 | K4 mean-by-type residual retrieval, scale 0.40, margin 0.20, no-op-only residuals, no-op bonus 0.03 | complete | 35.19% | +5.45 pp | yes | no | no | no-op-only residual-family ablation for sparse tangent transport |
 | retrieval_residual_k4_mean_nooponly_srcscorebonus002 | K4 mean-by-type residual retrieval, scale 0.40, margin 0.20, no-op-only residuals, source-score bonus 0.02 | complete | 35.19% | +5.45 pp | yes | no | no | no-op-only residual-family ablation with measured train-source prior |
 | retrieval_residual_k4_mean_noopbonus003_srcprog050 | K4 mean-by-type residual retrieval, scale 0.40, margin 0.20, no-op bonus 0.03, source progress >= 0.50 | complete | 34.96% | +5.22 pp | yes | no | no | train-source viability gate for sparse residual transport |