Auto-sync: 2026-06-29 12:50:13 (part 3)

results/paper_analysis.json

@@ -1,6 +1,6 @@
 {
   "best_clean_key": "residual_k4_composemasked_grid035040045_noopbonus003",
-  "generated_utc": "2026-06-29T16:37:59+00:00",
+  "generated_utc": "2026-06-29T16:47:33+00:00",
   "mechanism_gap": {
     "best_clean_vs_direct_same_ckpt": 0.07246376811594196,
     "best_clean_vs_h16": 0.0579710144927536,
@@ -1154,14 +1154,210 @@
       "std_success": 0.010190374394925787
     },
     "residual_k4_composemasked_l2comp002_grid035040045_noopbonus003": {
+      "ci95_success": 0.026399229893358696,
       "label": "K4 composed type-consensus tangents, masked, composite L2 penalty 0.02",
-      "missing": true,
-      "source": "results/h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_k4_composemasked_l2comp002_grid035040045_safe_margin0p20_noopbonus0p03_summary.json"
+      "mean_action_mse_to_best": 0.4078853273853336,
+      "mean_progress": 0.5703304209737426,
+      "mean_success": 0.3553623188405797,
+      "num_completed": 3,
+      "per_task_success": {
+        "LiftPegUpright-v1": {
+          "mean_num_groups": 102.0,
+          "mean_success": 0.26340865087329823,
+          "std_success": 0.06057881385012102
+        },
+        "PickCube-v1": {
+          "mean_num_groups": 196.66666666666666,
+          "mean_success": 0.3278169093386485,
+          "std_success": 0.02533117351097172
+        },
+        "PullCube-v1": {
+          "mean_num_groups": 81.0,
+          "mean_success": 0.20035568719779248,
+          "std_success": 0.02638291361020334
+        },
+        "PushCube-v1": {
+          "mean_num_groups": 102.0,
+          "mean_success": 0.757403601199683,
+          "std_success": 0.07781968814460603
+        },
+        "StackCube-v1": {
+          "mean_num_groups": 93.33333333333333,
+          "mean_success": 0.20044030044030045,
+          "std_success": 0.08896982358351117
+        }
+      },
+      "seed_action_mse_to_best": {
+        "0": 0.40025916028443886,
+        "1": 0.3954629489504125,
+        "2": 0.42793387292114937
+      },
+      "seed_progress": {
+        "0": 0.5580391604931134,
+        "1": 0.5702176707529504,
+        "2": 0.5827344316751639
+      },
+      "seed_success": {
+        "0": 0.35304347826086957,
+        "1": 0.34608695652173915,
+        "2": 0.36695652173913046
+      },
+      "selected_candidate_type_counts": {
+        "retrieval_residual_policy_residual": 1611,
+        "retrieval_residual_residual_near_miss": 8,
+        "retrieval_residual_residual_near_miss+residual_no_op": 9,
+        "retrieval_residual_residual_near_miss+residual_wrong_gripper": 4,
+        "retrieval_residual_residual_no_op": 59,
+        "retrieval_residual_residual_no_op+residual_wrong_gripper": 14,
+        "retrieval_residual_residual_wrong_gripper": 20
+      },
+      "selected_residual_scale_counts": {
+        "0.35": 1628,
+        "0.4": 10,
+        "0.45": 87
+      },
+      "selected_type_outcomes": {
+        "retrieval_residual_policy_residual": {
+          "count": 1611.0,
+          "mean_progress": 0.5615387783291779,
+          "success_rate": 0.345747982619491
+        },
+        "retrieval_residual_residual_near_miss": {
+          "count": 8.0,
+          "mean_progress": 0.8527381531894207,
+          "success_rate": 0.625
+        },
+        "retrieval_residual_residual_near_miss+residual_no_op": {
+          "count": 9.0,
+          "mean_progress": 0.41239654852284324,
+          "success_rate": 0.3333333333333333
+        },
+        "retrieval_residual_residual_near_miss+residual_wrong_gripper": {
+          "count": 4.0,
+          "mean_progress": 0.898208275437355,
+          "success_rate": 0.75
+        },
+        "retrieval_residual_residual_no_op": {
+          "count": 59.0,
+          "mean_progress": 0.7251843288225137,
+          "success_rate": 0.5084745762711864
+        },
+        "retrieval_residual_residual_no_op+residual_wrong_gripper": {
+          "count": 14.0,
+          "mean_progress": 0.6851060901369367,
+          "success_rate": 0.42857142857142855
+        },
+        "retrieval_residual_residual_wrong_gripper": {
+          "count": 20.0,
+          "mean_progress": 0.6338668182492256,
+          "success_rate": 0.45
+        }
+      },
+      "source": "results/h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_k4_composemasked_l2comp002_grid035040045_safe_margin0p20_noopbonus0p03_summary.json",
+      "std_success": 0.010626262481057023
     },
     "residual_k4_composemasked_l2comp005_grid035040045_noopbonus003": {
+      "ci95_success": 0.02405601694176965,
       "label": "K4 composed type-consensus tangents, masked, composite L2 penalty 0.05",
-      "missing": true,
-      "source": "results/h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_k4_composemasked_l2comp005_grid035040045_safe_margin0p20_noopbonus0p03_summary.json"
+      "mean_action_mse_to_best": 0.4051644595277806,
+      "mean_progress": 0.5694483268211469,
+      "mean_success": 0.3547826086956522,
+      "num_completed": 3,
+      "per_task_success": {
+        "LiftPegUpright-v1": {
+          "mean_num_groups": 102.0,
+          "mean_success": 0.26340865087329823,
+          "std_success": 0.06057881385012102
+        },
+        "PickCube-v1": {
+          "mean_num_groups": 196.66666666666666,
+          "mean_success": 0.3278169093386485,
+          "std_success": 0.02533117351097172
+        },
+        "PullCube-v1": {
+          "mean_num_groups": 81.0,
+          "mean_success": 0.20035568719779248,
+          "std_success": 0.02638291361020334
+        },
+        "PushCube-v1": {
+          "mean_num_groups": 102.0,
+          "mean_success": 0.7541032711666796,
+          "std_success": 0.07644139692615382
+        },
+        "StackCube-v1": {
+          "mean_num_groups": 93.33333333333333,
+          "mean_success": 0.20044030044030045,
+          "std_success": 0.08896982358351117
+        }
+      },
+      "seed_action_mse_to_best": {
+        "0": 0.39534642129002706,
+        "1": 0.3933826872949367,
+        "2": 0.4267642699983781
+      },
+      "seed_progress": {
+        "0": 0.5575822278395619,
+        "1": 0.5696634550133477,
+        "2": 0.581099297610531
+      },
+      "seed_success": {
+        "0": 0.35304347826086957,
+        "1": 0.34608695652173915,
+        "2": 0.3652173913043478
+      },
+      "selected_candidate_type_counts": {
+        "retrieval_residual_policy_residual": 1613,
+        "retrieval_residual_residual_near_miss": 9,
+        "retrieval_residual_residual_near_miss+residual_no_op": 8,
+        "retrieval_residual_residual_near_miss+residual_wrong_gripper": 5,
+        "retrieval_residual_residual_no_op": 60,
+        "retrieval_residual_residual_no_op+residual_wrong_gripper": 8,
+        "retrieval_residual_residual_wrong_gripper": 22
+      },
+      "selected_residual_scale_counts": {
+        "0.35": 1628,
+        "0.4": 11,
+        "0.45": 86
+      },
+      "selected_type_outcomes": {
+        "retrieval_residual_policy_residual": {
+          "count": 1613.0,
+          "mean_progress": 0.5610057968103742,
+          "success_rate": 0.3453192808431494
+        },
+        "retrieval_residual_residual_near_miss": {
+          "count": 9.0,
+          "mean_progress": 0.8691005806128184,
+          "success_rate": 0.6666666666666666
+        },
+        "retrieval_residual_residual_near_miss+residual_no_op": {
+          "count": 8.0,
+          "mean_progress": 0.33894611708819866,
+          "success_rate": 0.25
+        },
+        "retrieval_residual_residual_near_miss+residual_wrong_gripper": {
+          "count": 5.0,
+          "mean_progress": 0.918566620349884,
+          "success_rate": 0.8
+        },
+        "retrieval_residual_residual_no_op": {
+          "count": 60.0,
+          "mean_progress": 0.7145914144503573,
+          "success_rate": 0.5
+        },
+        "retrieval_residual_residual_no_op+residual_wrong_gripper": {
+          "count": 8.0,
+          "mean_progress": 0.7917308211326599,
+          "success_rate": 0.5
+        },
+        "retrieval_residual_residual_wrong_gripper": {
+          "count": 22.0,
+          "mean_progress": 0.5936534005132589,
+          "success_rate": 0.4090909090909091
+        }
+      },
+      "source": "results/h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_k4_composemasked_l2comp005_grid035040045_safe_margin0p20_noopbonus0p03_summary.json",
+      "std_success": 0.00968306845709568
     },
     "residual_k4_consensus": {
       "ci95_success": 0.04490086956521744,

results/paper_analysis.md

@@ -1,6 +1,6 @@
 # Paper Analysis
 
-Generated: `2026-06-29T16:37:59+00:00`
+Generated: `2026-06-29T16:47:33+00:00`
 
 ## Main Seed Statistics
 
@@ -48,8 +48,8 @@ Generated: `2026-06-29T16:37:59+00:00`
 | residual_k4_composemasked_grid035040045 | K4 composed type-consensus tangents, masked, scales 0.35/0.40/0.45 | 3 | 35.30% +/- 1.22 | +/- 3.02 | 56.91% | 0.410 | +5.57 pp |
 | residual_k4_composemasked_grid035040045_noopbonus003 | K4 composed type-consensus tangents, masked, scales 0.35/0.40/0.45, no-op bonus 0.03 | 3 | 35.54% +/- 1.02 | +/- 2.53 | 57.02% | 0.411 | +5.80 pp |
 | residual_k4_composemasked_compbonus_grid035040045_noopbonus003 | K4 composed type-consensus tangents, masked, component no-op bonus 0.03 | 3 | 35.36% +/- 1.16 | +/- 2.88 | 56.98% | 0.413 | +5.62 pp |
-| residual_k4_composemasked_l2comp002_grid035040045_noopbonus003 | K4 composed type-consensus tangents, masked, composite L2 penalty 0.02 | 0 | missing | missing | missing | missing | missing |
-| residual_k4_composemasked_l2comp005_grid035040045_noopbonus003 | K4 composed type-consensus tangents, masked, composite L2 penalty 0.05 | 0 | missing | missing | missing | missing | missing |
+| residual_k4_composemasked_l2comp002_grid035040045_noopbonus003 | K4 composed type-consensus tangents, masked, composite L2 penalty 0.02 | 3 | 35.54% +/- 1.06 | +/- 2.64 | 57.03% | 0.408 | +5.80 pp |
+| residual_k4_composemasked_l2comp005_grid035040045_noopbonus003 | K4 composed type-consensus tangents, masked, composite L2 penalty 0.05 | 3 | 35.48% +/- 0.97 | +/- 2.41 | 56.94% | 0.405 | +5.74 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 |

results/paper_core_results.md

@@ -56,6 +56,7 @@ lattice is `+27.25 pp`, and the remaining clean-to-same-state proposal gap is
 | K4 composed type-consensus residual retrieval, masked | No | No | 35.30% | +5.57 pp | Clean anti-goal composite masking removes the confound; tangent composition alone near-ties but does not beat mean-by-type scale-grid transport |
 | K4 composed type-consensus residual retrieval, masked + no-op prior 0.03 | No | No | 35.54% | +5.80 pp | Current best clean diagnostic; masked local tangent composition adds one success over the previous scale-grid no-op row while staying sparse |
 | K4 composed type-consensus residual retrieval, masked + component-wise no-op prior | No | No | 35.36% | +5.62 pp | Propagating the no-op prior into every composite over-selects composed tangents and drops below the exact-prior row; useful negative calibration |
+| K4 composed type-consensus residual retrieval, masked + composite trust penalty | No | No | 35.48-35.54% | +5.74-5.80 pp | A small composite-only L2 penalty ties the best while lowering action MSE and slightly improving progress; stronger penalty drops one seed, so trust-radius regularization is explanatory but not a new success gain |
 | 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 |
@@ -104,27 +105,28 @@ Suggested main-table rows:
 14. K4 mean-by-type residual retrieval + scale-grid no-op prior 0.03
 15. K4 masked composed type-consensus residual retrieval + no-op prior 0.03
 16. K4 masked composed type-consensus residual retrieval + component-wise no-op prior diagnostic
-17. K4 mean-by-type residual retrieval + upper/wide tangent-length diagnostics
-18. K4 mean-by-type residual retrieval + minimum-energy action penalty diagnostics
-19. K4 mean-by-type residual retrieval + source-progress/source-score/source-advantage prior diagnostics
-20. K4 mean-by-type residual retrieval + train-family success bonus diagnostics
-21. K4 mean-by-type residual retrieval + train-neighbor consensus-confidence diagnostics
-22. K4 repair-tangent residual transport diagnostics
-23. K4 mean-by-type residual retrieval + no-op-only family diagnostic
-24. K4 mean-by-type residual retrieval + abstention margin fine sweep
-25. Source-progress viability gate diagnostics
-26. K2/K4 task-relative retrieval metric diagnostics
-27. K4 kernel-weighted residual consensus + no-op prior diagnostics
-28. K4 field-softmax residual barycenter + margin diagnostics
-29. K4 mean-by-type residual retrieval + wrong-gripper typed-prior diagnostics
-30. K2 broad tangent ray-search
-31. Residual-tangent distillation policy
-32. Residual+Gaussian hybrid, K32 sigma0.35
-33. Lattice, near-miss only
-34. Lattice, no expert
-35. Lattice, no expert + policy baseline candidate
-36. Lattice, full
-37. Oracle ceiling
+17. K4 masked composed type-consensus residual retrieval + composite trust penalty diagnostic
+18. K4 mean-by-type residual retrieval + upper/wide tangent-length diagnostics
+19. K4 mean-by-type residual retrieval + minimum-energy action penalty diagnostics
+20. K4 mean-by-type residual retrieval + source-progress/source-score/source-advantage prior diagnostics
+21. K4 mean-by-type residual retrieval + train-family success bonus diagnostics
+22. K4 mean-by-type residual retrieval + train-neighbor consensus-confidence diagnostics
+23. K4 repair-tangent residual transport diagnostics
+24. K4 mean-by-type residual retrieval + no-op-only family diagnostic
+25. K4 mean-by-type residual retrieval + abstention margin fine sweep
+26. Source-progress viability gate diagnostics
+27. K2/K4 task-relative retrieval metric diagnostics
+28. K4 kernel-weighted residual consensus + no-op prior diagnostics
+29. K4 field-softmax residual barycenter + margin diagnostics
+30. K4 mean-by-type residual retrieval + wrong-gripper typed-prior diagnostics
+31. K2 broad tangent ray-search
+32. Residual-tangent distillation policy
+33. Residual+Gaussian hybrid, K32 sigma0.35
+34. Lattice, near-miss only
+35. Lattice, no expert
+36. Lattice, no expert + policy baseline candidate
+37. Lattice, full
+38. Oracle ceiling
 
 Suggested claim:
 
@@ -142,8 +144,9 @@ Suggested claim:
 > suggesting the current field already performs most of the useful abstention.
 > The clean composition row improves only after anti-goal composite masking, and
 > component-wise propagation of the no-op prior into every composite is negative.
-> This supports a controlled local tangent-chart story rather than an
-> uncontrolled proposal pile-up. Repair-tangent transport is negative, showing that simply reversing train failures into
+> A small composite-only trust penalty ties success while lowering action MSE,
+> supporting a controlled local tangent-chart story rather than an uncontrolled
+> proposal pile-up. 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,

results/paper_story_memo.md

@@ -31,6 +31,7 @@ when queried on proposal geometry that matches those local counterfactuals.
 | Field-gated tangent length calibration improves the clean bridge | K4 mean-by-type scale grid 0.35/0.40/0.45 with no-op bonus 0.03 reaches 35.42%; the source-score version reaches 35.30%. Upper 0.40/0.45/0.50 nearly ties at 35.36%, while wide 0.35/0.45/0.55 drops to 35.13% | Previous clean best; local scale calibration, not a larger-step effect |
 | Masked tangent composition gives a small clean lift | K4 composed type-consensus transport is 35.30% after anti-goal composite masking, and reaches 35.54% with the typed no-op prior; raw selected types contain no random-negative or wrong-direction composites | Current best clean result; controlled local tangent-chart composition, not unmasked proposal accumulation |
 | Component-wise composite priors do not add the gain | propagating the no-op bonus into composite types reaches 35.36%, below the exact-prior masked composition row at 35.54% | Negative/near-tie diagnostic: sparse exact priors are better than broadly rewarding every no-op-containing composite |
+| Composite trust-radius penalty explains but does not improve the success top line | composite-only L2 penalty 0.02 ties 35.54% while lowering action MSE from 0.4106 to 0.4079 and slightly raising progress; penalty 0.05 drops to 35.48% | Tie/negative diagnostic: composed tangents need a local trust radius, but the field+margin already finds the success optimum |
 | Minimum-energy residual regularization does not add the gain | action L2 penalty 0.05 ties the previous 35.42% scale-grid row, while 0.10/0.20 reach 35.36% | Negative/tie diagnostic: the clean bridge is not explained by shortest-action bias |
 | 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 previous scale-grid row at 35.42% | Negative/tie diagnostic: train terminal success is not the right confidence signal for transferred tangents |
@@ -78,33 +79,34 @@ clean proposal result, the intended main rows are:
 16. K4 mean-by-type tangent consensus + scale-grid typed no-op prior: 35.42%
 17. K4 masked composed type-consensus tangent transport: 35.30%; with typed no-op prior: 35.54%
 18. K4 masked composed type-consensus tangent transport + component-wise no-op prior: 35.36%
-19. K4 mean-by-type tangent consensus + upper/wide scale diagnostics: 35.36% for 0.40/0.45/0.50; 35.13% for 0.35/0.45/0.55
-20. K4 mean-by-type tangent consensus + action L2 penalty: 35.42% at 0.05; 35.36% at 0.10/0.20
-21. K4 mean-by-type tangent consensus + train-source progress prior: 35.25% at bonus 0.03; 35.13% at bonus 0.05
-22. K4 mean-by-type tangent consensus + train-source reward-score prior: 35.25% at bonuses 0.015/0.020; 35.30% with scale grid; 35.19% at 0.025
-23. 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
-24. K4 mean-by-type tangent consensus + train-family success bonus: 35.25% alone; 35.42% with no-op bonus 0.03
-25. K4 mean-by-type tangent consensus + train-neighbor consensus penalty: 35.19% alone; 35.36% with no-op bonus 0.03
-26. K4 repair-tangent transport: 34.14-34.43%
-27. 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
-28. 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
-29. Source-progress viability gates: 35.19% / 34.96% / 34.72% for thresholds 0.25 / 0.50 / 0.75
-30. K4 kernel-weighted tangent consensus / + no-op prior: 34.96% / 35.19%
-31. K4 field-softmax tangent transport / best margin sweep: 34.96% / 35.19%
-32. Wrong-gripper prior / no-op+wrong-gripper prior: 35.19% / 35.25%
-33. K2 broad tangent ray-search: 34.96%
-34. K1/K2 tight tangent ray-search: 34.84% / 34.84%
-35. K4 tight tangent ray-search: 34.55%
-36. Residual-tangent distillation policy: 28.87%
-37. Z-score residual retrieval: 32.23-32.81%
-38. Task-relative residual retrieval metric: 34.26-34.43%
-39. Train-family reliability prior: 33.28-33.33%
-40. Residual+Gaussian hybrid K32/K64: 31.30% / 30.90%
-41. Lattice, near-miss only: 55.94%
-42. Lattice, no expert: 56.99%
-43. Lattice, no expert + policy baseline candidate: 40.70%
-44. Lattice, full: 69.33%
-45. Oracle ceiling: 86.78%
+19. K4 masked composed type-consensus tangent transport + composite L2 trust penalty: 35.54% at 0.02; 35.48% at 0.05
+20. K4 mean-by-type tangent consensus + upper/wide scale diagnostics: 35.36% for 0.40/0.45/0.50; 35.13% for 0.35/0.45/0.55
+21. K4 mean-by-type tangent consensus + action L2 penalty: 35.42% at 0.05; 35.36% at 0.10/0.20
+22. K4 mean-by-type tangent consensus + train-source progress prior: 35.25% at bonus 0.03; 35.13% at bonus 0.05
+23. K4 mean-by-type tangent consensus + train-source reward-score prior: 35.25% at bonuses 0.015/0.020; 35.30% with scale grid; 35.19% at 0.025
+24. 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
+25. K4 mean-by-type tangent consensus + train-family success bonus: 35.25% alone; 35.42% with no-op bonus 0.03
+26. K4 mean-by-type tangent consensus + train-neighbor consensus penalty: 35.19% alone; 35.36% with no-op bonus 0.03
+27. K4 repair-tangent transport: 34.14-34.43%
+28. 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
+29. 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
+30. Source-progress viability gates: 35.19% / 34.96% / 34.72% for thresholds 0.25 / 0.50 / 0.75
+31. K4 kernel-weighted tangent consensus / + no-op prior: 34.96% / 35.19%
+32. K4 field-softmax tangent transport / best margin sweep: 34.96% / 35.19%
+33. Wrong-gripper prior / no-op+wrong-gripper prior: 35.19% / 35.25%
+34. K2 broad tangent ray-search: 34.96%
+35. K1/K2 tight tangent ray-search: 34.84% / 34.84%
+36. K4 tight tangent ray-search: 34.55%
+37. Residual-tangent distillation policy: 28.87%
+38. Z-score residual retrieval: 32.23-32.81%
+39. Task-relative residual retrieval metric: 34.26-34.43%
+40. Train-family reliability prior: 33.28-33.33%
+41. Residual+Gaussian hybrid K32/K64: 31.30% / 30.90%
+42. Lattice, near-miss only: 55.94%
+43. Lattice, no expert: 56.99%
+44. Lattice, no expert + policy baseline candidate: 40.70%
+45. Lattice, full: 69.33%
+46. Oracle ceiling: 86.78%
 
 ## Novelty Framing
 
@@ -132,13 +134,14 @@ test-time search. The cleaner novelty is:
 
 ## Job Status
 
-Last checked: `2026-06-29 16:27 UTC`. The K4 masked composed type-consensus
+Last checked: `2026-06-29 16:46 UTC`. The K4 masked composed type-consensus
 sweep completed and produced a new clean best, 35.54%, while the pure masked
 composition row reached 35.30%. Raw selected candidate types show no
 random-negative or wrong-direction composite leak. The paper table and paired
 analysis now use the masked composed no-op row as `best_clean_key`. The
 component-wise composite prior follow-up completed at 35.36%, below the exact
-no-op prior row.
+no-op prior row. Composite-only L2 trust penalties completed: 0.02 ties the
+best success with lower action MSE, while 0.05 drops to 35.48%.
 
 - `14858328`-`14858333`: completed train-split `field_selected_noexpert_bc5`;
   direct rollout is 26.84%, field-guided best is 27.65%.
@@ -288,6 +291,16 @@ no-op prior row.
   summary. It reaches 35.36% with seeds 34.78%/34.61%/36.70%, below the 35.54%
   exact-prior row. Local paper builders updated artifacts; redundant rebuild
   job `14912563` was canceled while still pending.
+- `14913943`: completed CPU smoke for composite-only L2 trust penalties on
+  masked composed type-consensus transport. Metadata records
+  `retrieval_residual_composite_l2_penalty_scale=0.02`, no anti-goal selected
+  types, and 8/8 smoke groups selected `policy_residual`.
+- `14913944`/`14913951`: completed composite L2 penalty GPU arrays. Penalty
+  `0.02` ties the 35.54% exact-prior row with seeds
+  35.30%/34.61%/36.70%, slightly higher progress and lower action MSE. Penalty
+  `0.05` reaches 35.48%. Local summaries and builders updated artifacts;
+  redundant summary/rebuild jobs `14913955`/`14913956`/`14913958` were canceled
+  while pending.
 - `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]`).
@@ -378,7 +391,10 @@ no-op prior row.
   no-op row remains 35.42%; pure masked composition is 35.30%, and
   component-wise propagation of the no-op prior reaches 35.36%, so composition
   helps only when the sparse exact typed prior is preserved and anti-goal
-  composite parts are masked. The fixed-scale no-op plateau remains 35.25%; continuous
+  composite parts are masked. A composite-only L2 trust penalty ties the best
+  success at 0.02 while lowering action MSE, but 0.05 drops below the best, so
+  this is explanatory regularization rather than a new success gain. The
+  fixed-scale no-op plateau remains 35.25%; continuous
   train-source progress/reward-score priors tie that fixed-scale row, and
   scale-grid source-score reaches 35.30% but not the new best. Source-advantage
   priors/gates reach at most 35.30%, so local utility lift over the source
@@ -399,7 +415,7 @@ no-op prior row.
   repaired train-family reliability priors, Gaussian hybrids,
   field optimization, field-teacher/tangent distillation, repair-tangent transport, policy-relative anchoring, tangent consensus,
   kernel-weighted tangent interpolation, field-softmax tangent barycenters,
-  unmasked, prior-free, or component-prior tangent composition, wrong-gripper
-  typed priors, and same-state policy-baseline fallback as negative or near-tie
-  diagnostics that sharpen the story around local counterfactual proposal
-  geometry.
+  unmasked, prior-free, component-prior, or over-regularized tangent
+  composition, wrong-gripper typed priors, and same-state policy-baseline
+  fallback as negative or near-tie diagnostics that sharpen the story around
+  local counterfactual proposal geometry.

results/paper_table_status.json

@@ -1325,14 +1325,14 @@
       "story_role": "trust-radius penalty on composed local tangent candidates",
       "fallback_success": null,
       "pending_job": "14913944/14913955",
-      "path_exists": false,
-      "status": "pending",
-      "success": null,
-      "std_success": null,
+      "path_exists": true,
+      "status": "complete",
+      "success": 0.3553623188405797,
+      "std_success": 0.010626262481057023,
       "completed_seeds": null,
-      "num_completed": null,
+      "num_completed": 3,
       "best_config": null,
-      "gain_vs_h16_policy": null
+      "gain_vs_h16_policy": 0.05797101449275366
     },
     {
       "key": "retrieval_residual_k4_composemasked_l2comp005_grid035040045_noopbonus003",
@@ -1344,14 +1344,14 @@
       "story_role": "stronger trust-radius penalty on composed local tangent candidates",
       "fallback_success": null,
       "pending_job": "14913951/14913956",
-      "path_exists": false,
-      "status": "pending",
-      "success": null,
-      "std_success": null,
+      "path_exists": true,
+      "status": "complete",
+      "success": 0.3547826086956522,
+      "std_success": 0.00968306845709568,
       "completed_seeds": null,
-      "num_completed": null,
+      "num_completed": 3,
       "best_config": null,
-      "gain_vs_h16_policy": null
+      "gain_vs_h16_policy": 0.05739130434782613
     },
     {
       "key": "retrieval_repair_nearmiss_k4_grid025035050_margin020",

results/paper_table_status.md

@@ -72,8 +72,8 @@ Baseline h=16 policy: 29.74%
 | retrieval_residual_k4_composemasked_grid035040045 | K4 composed type-consensus residual retrieval, masked, scales 0.35/0.40/0.45, margin 0.20 | complete | 35.30% | +5.57 pp | yes | no | no | local tangent composition with anti-goal composite masks |
 | retrieval_residual_k4_composemasked_grid035040045_noopbonus003 | K4 composed type-consensus residual retrieval, masked, scales 0.35/0.40/0.45, margin 0.20, no-op bonus 0.03 | complete | 35.54% | +5.80 pp | yes | no | no | local tangent composition with anti-goal composite masks on the current best typed prior |
 | retrieval_residual_k4_composemasked_compbonus_grid035040045_noopbonus003 | K4 composed type-consensus residual retrieval, masked, component no-op bonus 0.03 | complete | 35.36% | +5.62 pp | yes | no | no | component-wise sparse prior on the masked local tangent composition chart |
-| retrieval_residual_k4_composemasked_l2comp002_grid035040045_noopbonus003 | K4 composed type-consensus residual retrieval, masked, composite L2 penalty 0.02 | pending 14913944/14913955 | pending | pending | yes | no | no | trust-radius penalty on composed local tangent candidates |
-| retrieval_residual_k4_composemasked_l2comp005_grid035040045_noopbonus003 | K4 composed type-consensus residual retrieval, masked, composite L2 penalty 0.05 | pending 14913951/14913956 | pending | pending | yes | no | no | stronger trust-radius penalty on composed local tangent candidates |
+| retrieval_residual_k4_composemasked_l2comp002_grid035040045_noopbonus003 | K4 composed type-consensus residual retrieval, masked, composite L2 penalty 0.02 | complete | 35.54% | +5.80 pp | yes | no | no | trust-radius penalty on composed local tangent candidates |
+| retrieval_residual_k4_composemasked_l2comp005_grid035040045_noopbonus003 | K4 composed type-consensus residual retrieval, masked, composite L2 penalty 0.05 | complete | 35.48% | +5.74 pp | yes | no | no | stronger trust-radius penalty on composed local tangent candidates |
 | 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 |

scripts/build_paper_analysis.py

@@ -838,7 +838,8 @@ def _load_methods() -> dict[str, dict[str, Any]]:
 
 
 def _best_clean_key(methods: dict[str, dict[str, Any]]) -> str:
-    candidates: list[tuple[float, str]] = []
+    best_key = FALLBACK_BEST_CLEAN_KEY
+    best_success = float("-inf")
     for spec in METHODS:
         if spec.key in NON_DEPLOYMENT_KEYS:
             continue
@@ -847,10 +848,11 @@ def _best_clean_key(methods: dict[str, dict[str, Any]]) -> str:
             continue
         success = method.get("mean_success")
         if isinstance(success, (int, float)) and math.isfinite(float(success)):
-            candidates.append((float(success), spec.key))
-    if not candidates:
-        return FALLBACK_BEST_CLEAN_KEY
-    return max(candidates)[1]
+            success = float(success)
+            if success > best_success + 1.0e-12:
+                best_success = success
+                best_key = spec.key
+    return best_key
 
 
 def _paired_delta(