Auto-sync: 2026-06-28 20:44:54 (part 3)

results/h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_k4s040_safe_margin0p20_mean_by_type_noop0p025_wg0p02_summary.json

@@ -0,0 +1,322 @@
+{
+  "run_root": "/scratch/knguy52/dovla/experiments/dovla_h16_policy_ckpt_runs",
+  "objective": "near_miss_policy_bc5",
+  "out_name": "policy_rollout_retrieval_residual_k4s040_safe_margin0p20_mean_by_type_noop0p025_wg0p02.json",
+  "num_completed": 3,
+  "baseline_h4_policy_success": 0.2967,
+  "baseline_h16_rank_checkpoint_success": 0.29739130434782607,
+  "mean_success": 0.35246376811594204,
+  "std_success": 0.014164396200429698,
+  "mean_progress": 0.5669270546512079,
+  "mean_action_mse_to_best": 0.39555692845307616,
+  "gain_vs_h4": 0.055763768115942025,
+  "gain_vs_h16_rank_checkpoint": 0.055072463768115976,
+  "rows": [
+    {
+      "seed": 0,
+      "path": "/scratch/knguy52/dovla/experiments/dovla_h16_policy_ckpt_runs/near_miss_policy_bc5/seed_0/policy_rollout_retrieval_residual_k4s040_safe_margin0p20_mean_by_type_noop0p025_wg0p02.json",
+      "num_groups": 575,
+      "selection_mode": "retrieval_residual",
+      "num_candidates": 6,
+      "candidate_sigma": 0.0,
+      "selection_margin": 0.2,
+      "prepend_policy_candidate": false,
+      "field_optim_steps": 0,
+      "field_optim_step_size": 0.0,
+      "field_optim_trust_radius": 0.0,
+      "field_optim_l2_penalty": 0.0,
+      "retrieval_neighbors": 4,
+      "retrieval_metric": "raw",
+      "retrieval_type_min_success": 0.0,
+      "retrieval_residual_scale": 0.4,
+      "retrieval_residual_scales": [],
+      "retrieval_residual_anchor": "expert",
+      "retrieval_residual_reduce": "mean_by_type",
+      "candidate_type_bonuses": {
+        "residual_no_op": 0.025,
+        "residual_wrong_gripper": 0.02
+      },
+      "selected_residual_scale_counts": {
+        "0.4": 575
+      },
+      "policy_rollout_success_rate": 0.34608695652173915,
+      "policy_rollout_progress": 0.551369690887792,
+      "oracle_success_rate": 0.8573913043478261,
+      "action_mse_to_best": 0.38197267810611624,
+      "best_policy_val": {
+        "bc_loss": 0.13721593966086706,
+        "field_effect_loss": 0.009290305380192068,
+        "field_potential_loss": 0.2666468388504452,
+        "field_preference_loss": 0.5130573478009965,
+        "lattice_edges": 3833.3333333333335,
+        "progress_mae": 0.1933159919248687,
+        "rank_acc": 0.8265031774838766,
+        "rank_loss": 0.5130523675017886,
+        "regret_mae": 0.3756548762321472,
+        "success_accuracy": 0.8773836526605818,
+        "total_loss": 1.5581054819954767
+      },
+      "per_task": {
+        "LiftPegUpright-v1": {
+          "action_mse_to_best": 0.31969772460726426,
+          "expert_success_rate": 0.8865979381443299,
+          "num_groups": 97,
+          "oracle_success_rate": 0.9278350515463918,
+          "policy_expert_regret": 1.0783334703175063,
+          "policy_oracle_regret": 1.1047489480259491,
+          "policy_rollout_progress": 0.5714634232914325,
+          "policy_rollout_success_rate": 0.2268041237113402,
+          "restore_max_error": 2.384185791015625e-07
+        },
+        "PickCube-v1": {
+          "action_mse_to_best": 0.3026410622450595,
+          "expert_success_rate": 0.9375,
+          "num_groups": 208,
+          "oracle_success_rate": 0.9471153846153846,
+          "policy_expert_regret": 0.961157998499962,
+          "policy_oracle_regret": 0.9745808659264674,
+          "policy_rollout_progress": 0.6161086453268161,
+          "policy_rollout_success_rate": 0.34134615384615385,
+          "restore_max_error": 2.384185791015625e-07
+        },
+        "PullCube-v1": {
+          "action_mse_to_best": 0.6028722151146306,
+          "expert_success_rate": 0.19480519480519481,
+          "num_groups": 77,
+          "oracle_success_rate": 0.36363636363636365,
+          "policy_expert_regret": 0.24377496006547825,
+          "policy_oracle_regret": 0.41733099815931607,
+          "policy_rollout_progress": 0.312481374268166,
+          "policy_rollout_success_rate": 0.2077922077922078,
+          "restore_max_error": 2.3096799850463867e-07
+        },
+        "PushCube-v1": {
+          "action_mse_to_best": 0.36465501579198434,
+          "expert_success_rate": 0.8617021276595744,
+          "num_groups": 94,
+          "oracle_success_rate": 0.9893617021276596,
+          "policy_expert_regret": 0.5508309629686335,
+          "policy_oracle_regret": 0.647163599729538,
+          "policy_rollout_progress": 0.6932619321853557,
+          "policy_rollout_success_rate": 0.6595744680851063,
+          "restore_max_error": 2.384185791015625e-07
+        },
+        "StackCube-v1": {
+          "action_mse_to_best": 0.45429835988754275,
+          "expert_success_rate": 0.6767676767676768,
+          "num_groups": 99,
+          "oracle_success_rate": 0.8585858585858586,
+          "policy_expert_regret": 0.9403071221378114,
+          "policy_oracle_regret": 1.10289698026397,
+          "policy_rollout_progress": 0.4467407528198127,
+          "policy_rollout_success_rate": 0.2828282828282828,
+          "restore_max_error": 2.384185791015625e-07
+        }
+      }
+    },
+    {
+      "seed": 1,
+      "path": "/scratch/knguy52/dovla/experiments/dovla_h16_policy_ckpt_runs/near_miss_policy_bc5/seed_1/policy_rollout_retrieval_residual_k4s040_safe_margin0p20_mean_by_type_noop0p025_wg0p02.json",
+      "num_groups": 575,
+      "selection_mode": "retrieval_residual",
+      "num_candidates": 6,
+      "candidate_sigma": 0.0,
+      "selection_margin": 0.2,
+      "prepend_policy_candidate": false,
+      "field_optim_steps": 0,
+      "field_optim_step_size": 0.0,
+      "field_optim_trust_radius": 0.0,
+      "field_optim_l2_penalty": 0.0,
+      "retrieval_neighbors": 4,
+      "retrieval_metric": "raw",
+      "retrieval_type_min_success": 0.0,
+      "retrieval_residual_scale": 0.4,
+      "retrieval_residual_scales": [],
+      "retrieval_residual_anchor": "expert",
+      "retrieval_residual_reduce": "mean_by_type",
+      "candidate_type_bonuses": {
+        "residual_no_op": 0.025,
+        "residual_wrong_gripper": 0.02
+      },
+      "selected_residual_scale_counts": {
+        "0.4": 575
+      },
+      "policy_rollout_success_rate": 0.3426086956521739,
+      "policy_rollout_progress": 0.5665413653461829,
+      "oracle_success_rate": 0.8695652173913043,
+      "action_mse_to_best": 0.3883226982811871,
+      "best_policy_val": {
+        "bc_loss": 0.1315989200439718,
+        "field_effect_loss": 0.010110370814800262,
+        "field_potential_loss": 0.31546536915832096,
+        "field_preference_loss": 0.4960637440284093,
+        "lattice_edges": 3833.3333333333335,
+        "progress_mae": 0.19023226698239645,
+        "rank_acc": 0.8255469501018524,
+        "rank_loss": 0.4960531195004781,
+        "regret_mae": 0.42156532737943864,
+        "success_accuracy": 0.8807473679383596,
+        "total_loss": 1.561984618504842
+      },
+      "per_task": {
+        "LiftPegUpright-v1": {
+          "action_mse_to_best": 0.3255710795406351,
+          "expert_success_rate": 0.8584070796460177,
+          "num_groups": 113,
+          "oracle_success_rate": 0.9380530973451328,
+          "policy_expert_regret": 0.9873926745051831,
+          "policy_oracle_regret": 1.0665947045900126,
+          "policy_rollout_progress": 0.6355170211960784,
+          "policy_rollout_success_rate": 0.23008849557522124,
+          "restore_max_error": 2.384185791015625e-07
+        },
+        "PickCube-v1": {
+          "action_mse_to_best": 0.3200199794550629,
+          "expert_success_rate": 0.9402173913043478,
+          "num_groups": 184,
+          "oracle_success_rate": 0.9456521739130435,
+          "policy_expert_regret": 1.0707762707429735,
+          "policy_oracle_regret": 1.0708002752179038,
+          "policy_rollout_progress": 0.5744255784818012,
+          "policy_rollout_success_rate": 0.29347826086956524,
+          "restore_max_error": 2.384185791015625e-07
+        },
+        "PullCube-v1": {
+          "action_mse_to_best": 0.5715771112591028,
+          "expert_success_rate": 0.25,
+          "num_groups": 76,
+          "oracle_success_rate": 0.40789473684210525,
+          "policy_expert_regret": 0.30013844498286124,
+          "policy_oracle_regret": 0.5234715515061429,
+          "policy_rollout_progress": 0.28913191000097677,
+          "policy_rollout_success_rate": 0.18421052631578946,
+          "restore_max_error": 2.384185791015625e-07
+        },
+        "PushCube-v1": {
+          "action_mse_to_best": 0.35765871029716356,
+          "expert_success_rate": 0.8198198198198198,
+          "num_groups": 111,
+          "oracle_success_rate": 1.0,
+          "policy_expert_regret": 0.32971009934270706,
+          "policy_oracle_regret": 0.38007361851296984,
+          "policy_rollout_progress": 0.8181245796852283,
+          "policy_rollout_success_rate": 0.8018018018018018,
+          "restore_max_error": 2.3795291781425476e-07
+        },
+        "StackCube-v1": {
+          "action_mse_to_best": 0.48870731873826667,
+          "expert_success_rate": 0.6923076923076923,
+          "num_groups": 91,
+          "oracle_success_rate": 0.8571428571428571,
+          "policy_expert_regret": 1.1411885159028756,
+          "policy_oracle_regret": 1.3008503630593582,
+          "policy_rollout_progress": 0.38975496410013555,
+          "policy_rollout_success_rate": 0.15384615384615385,
+          "restore_max_error": 1.9371509552001953e-07
+        }
+      }
+    },
+    {
+      "seed": 2,
+      "path": "/scratch/knguy52/dovla/experiments/dovla_h16_policy_ckpt_runs/near_miss_policy_bc5/seed_2/policy_rollout_retrieval_residual_k4s040_safe_margin0p20_mean_by_type_noop0p025_wg0p02.json",
+      "num_groups": 575,
+      "selection_mode": "retrieval_residual",
+      "num_candidates": 6,
+      "candidate_sigma": 0.0,
+      "selection_margin": 0.2,
+      "prepend_policy_candidate": false,
+      "field_optim_steps": 0,
+      "field_optim_step_size": 0.0,
+      "field_optim_trust_radius": 0.0,
+      "field_optim_l2_penalty": 0.0,
+      "retrieval_neighbors": 4,
+      "retrieval_metric": "raw",
+      "retrieval_type_min_success": 0.0,
+      "retrieval_residual_scale": 0.4,
+      "retrieval_residual_scales": [],
+      "retrieval_residual_anchor": "expert",
+      "retrieval_residual_reduce": "mean_by_type",
+      "candidate_type_bonuses": {
+        "residual_no_op": 0.025,
+        "residual_wrong_gripper": 0.02
+      },
+      "selected_residual_scale_counts": {
+        "0.4": 575
+      },
+      "policy_rollout_success_rate": 0.36869565217391304,
+      "policy_rollout_progress": 0.5828701077196488,
+      "oracle_success_rate": 0.8765217391304347,
+      "action_mse_to_best": 0.41637540897192515,
+      "best_policy_val": {
+        "bc_loss": 0.11367896075050037,
+        "field_effect_loss": 0.009670218582161598,
+        "field_potential_loss": 0.2641640139950646,
+        "field_preference_loss": 0.5130490180518892,
+        "lattice_edges": 3833.3333333333335,
+        "progress_mae": 0.2021729110015763,
+        "rank_acc": 0.8333857821093665,
+        "rank_loss": 0.5130119257503085,
+        "regret_mae": 0.3958987047274907,
+        "success_accuracy": 0.8680730561415354,
+        "total_loss": 1.4394984311527677
+      },
+      "per_task": {
+        "LiftPegUpright-v1": {
+          "action_mse_to_best": 0.35425721921395353,
+          "expert_success_rate": 0.8229166666666666,
+          "num_groups": 96,
+          "oracle_success_rate": 0.9270833333333334,
+          "policy_expert_regret": 0.8476849501021206,
+          "policy_oracle_regret": 0.9539546399998168,
+          "policy_rollout_progress": 0.6463205499264101,
+          "policy_rollout_success_rate": 0.3333333333333333,
+          "restore_max_error": 1.955777406692505e-07
+        },
+        "PickCube-v1": {
+          "action_mse_to_best": 0.335480088330429,
+          "expert_success_rate": 0.9444444444444444,
+          "num_groups": 198,
+          "oracle_success_rate": 0.9595959595959596,
+          "policy_expert_regret": 0.9807514474924767,
+          "policy_oracle_regret": 0.9878348980211851,
+          "policy_rollout_progress": 0.6357372662409989,
+          "policy_rollout_success_rate": 0.3383838383838384,
+          "restore_max_error": 2.384185791015625e-07
+        },
+        "PullCube-v1": {
+          "action_mse_to_best": 0.6275676680935753,
+          "expert_success_rate": 0.24444444444444444,
+          "num_groups": 90,
+          "oracle_success_rate": 0.4666666666666667,
+          "policy_expert_regret": 0.3140720418619821,
+          "policy_oracle_regret": 0.5427349449100802,
+          "policy_rollout_progress": 0.3188257521984269,
+          "policy_rollout_success_rate": 0.2111111111111111,
+          "restore_max_error": 2.384185791015625e-07
+        },
+        "PushCube-v1": {
+          "action_mse_to_best": 0.3851195577495169,
+          "expert_success_rate": 0.8514851485148515,
+          "num_groups": 101,
+          "oracle_success_rate": 1.0,
+          "policy_expert_regret": 0.3620445999768701,
+          "policy_oracle_regret": 0.41694685138098087,
+          "policy_rollout_progress": 0.8008749307972369,
+          "policy_rollout_success_rate": 0.7821782178217822,
+          "restore_max_error": 2.384185791015625e-07
+        },
+        "StackCube-v1": {
+          "action_mse_to_best": 0.4844882684863276,
+          "expert_success_rate": 0.7666666666666667,
+          "num_groups": 90,
+          "oracle_success_rate": 0.9111111111111111,
+          "policy_expert_regret": 1.1607355892658233,
+          "policy_oracle_regret": 1.2970718792743152,
+          "policy_rollout_progress": 0.4182763857973946,
+          "policy_rollout_success_rate": 0.16666666666666666,
+          "restore_max_error": 2.2351741790771484e-07
+        }
+      }
+    }
+  ]
+}

results/h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_k4s040_safe_margin0p20_mean_by_type_noop0p025_wg0p02_summary.md

@@ -0,0 +1,19 @@
+# h=16 Best-Policy Checkpoint Rollout
+
+Run root: `/scratch/knguy52/dovla/experiments/dovla_h16_policy_ckpt_runs`
+Objective: `near_miss_policy_bc5`
+Result file: `policy_rollout_retrieval_residual_k4s040_safe_margin0p20_mean_by_type_noop0p025_wg0p02.json`
+Completed seeds: 3
+Baseline h=4 policy success: 29.67%
+Baseline h=16 rank-checkpoint success: 29.74%
+
+Mean success: 35.25% +/- 1.42%
+Gain vs h=16 rank checkpoint: +5.51%
+Mean progress: 56.69%
+Mean action MSE to best: 0.396
+
+| seed | mode | k | policy cand | retrieval K | retrieval metric | residual anchor | residual reduce | min type success | residual scale | residual scales | margin | sigma | opt steps | trust | success | progress | oracle | action MSE |
+|---:|---|---:|---|---:|---|---|---|---:|---:|---|---:|---:|---:|---:|---:|---:|---:|---:|
+| 0 | retrieval_residual | 6 | no | 4 | raw | expert | mean_by_type | 0.00 | 0.40 | none | 0.200 | 0.00 | 0 | 0.00 | 34.61% | 55.14% | 85.74% | 0.382 |
+| 1 | retrieval_residual | 6 | no | 4 | raw | expert | mean_by_type | 0.00 | 0.40 | none | 0.200 | 0.00 | 0 | 0.00 | 34.26% | 56.65% | 86.96% | 0.388 |
+| 2 | retrieval_residual | 6 | no | 4 | raw | expert | mean_by_type | 0.00 | 0.40 | none | 0.200 | 0.00 | 0 | 0.00 | 36.87% | 58.29% | 87.65% | 0.416 |

results/paper_analysis.json

@@ -1,6 +1,6 @@
 {
   "best_clean_key": "residual_k4_consensus_noopbonus003",
-  "generated_utc": "2026-06-29T00:39:30+00:00",
+  "generated_utc": "2026-06-29T00:41:07+00:00",
   "mechanism_gap": {
     "best_clean_vs_direct_same_ckpt": 0.06956521739130428,
     "best_clean_vs_h16": 0.05507246376811592,

results/paper_analysis.md

@@ -1,6 +1,6 @@
 # Paper Analysis
 
-Generated: `2026-06-29T00:39:30+00:00`
+Generated: `2026-06-29T00:41:07+00:00`
 
 ## Main Seed Statistics
 

results/paper_core_results.md

@@ -40,6 +40,7 @@ and the remaining clean-to-same-state proposal gap is `+21.74 pp`.
 | K2 train-state residual retrieval, safe residuals + advantage margin 0.20 | No | No | 35.01% | +5.28 pp | Previous best clean diagnostic; abstention makes a small train-neighborhood useful |
 | K4 train-state residual retrieval, safe residuals + mean-by-type tangent consensus | No | No | 34.96% | +5.22 pp | Near-tie clean diagnostic; consensus alone does not beat raw K2 residuals |
 | K4 mean-by-type residual retrieval + no-op prior 0.03 | No | No | 35.25% | +5.51 pp | Current best clean diagnostic; 0.025-0.035 forms a small plateau that nudges high-value no-op residuals without changing the core proposal family |
+| K4 mean-by-type residual retrieval + wrong-gripper typed prior | No | No | 35.19-35.25% | +5.45-5.51 pp | Wrong-gripper-only is lower and two-family priors only tie the no-op plateau; useful negative/tie diagnostic |
 | K1 train-state residual ray-search, tight scales | No | No | 34.84% | +5.10 pp | Scale-grid ray-search is a near-tie but does not beat the typed-prior clean row |
 | K2 train-state residual ray-search, tight scales | No | No | 34.84% | +5.10 pp | More scale choices along the same local rays do not improve the clean row |
 | K2 train-state residual ray-search, broad scales | No | No | 34.96% | +5.22 pp | Best ray-search row, still below the typed-prior clean row |
@@ -77,14 +78,15 @@ Suggested main-table rows:
 12. K2 train-state residual retrieval, typed safe families + advantage margin 0.20
 13. K4 train-state residual retrieval, mean-by-type tangent consensus
 14. K4 mean-by-type residual retrieval + no-op prior plateau, canonical 0.03
-15. K2 broad tangent ray-search
-16. Residual-tangent distillation policy
-17. Residual+Gaussian hybrid, K32 sigma0.35
-18. Lattice, near-miss only
-19. Lattice, no expert
-20. Lattice, no expert + policy baseline candidate
-21. Lattice, full
-22. Oracle ceiling
+15. K4 mean-by-type residual retrieval + wrong-gripper typed-prior diagnostics
+16. K2 broad tangent ray-search
+17. Residual-tangent distillation policy
+18. Residual+Gaussian hybrid, K32 sigma0.35
+19. Lattice, near-miss only
+20. Lattice, no expert
+21. Lattice, no expert + policy baseline candidate
+22. Lattice, full
+23. Oracle ceiling
 
 Suggested claim:
 
@@ -93,7 +95,7 @@ Suggested claim:
 > abstention and a small typed no-op prior plateau gives the strongest clean gain so far, while ungated KNN residual
 > retrieval, field-gradient ascent, broader non-expert BC targets, field-teacher/tangent distillation, z-score retrieval,
 > train-family reliability priors, policy-relative anchoring, residual+Gaussian hybrids,
-> tangent consensus, tangent ray-search, and same-state policy-baseline fallback fail to improve the main rows.
+> tangent consensus, 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

@@ -30,6 +30,7 @@ when queried on proposal geometry that matches those local counterfactuals.
 | Tangent consensus is close but needs sparse typing | K4 mean-by-type residual consensus reaches 34.96%; a small no-op residual prior plateau at 0.025-0.035 raises it to 35.25% | Current best clean result |
 | Tangent ray-search does not beat the typed-prior clean row | K1/K2 tight scale-grid ray search reach 34.84%; K2 broad reaches 34.96%; K4 tight reaches 34.55%, all below the no-op-prior row at 35.25% | Near-tie/negative diagnostic |
 | Typed no-op residual prior improves the clean bridge | CPU smoke `14883591` passed; bonuses 0.025/0.03/0.035 tie at 35.25%, while 0.01/0.02/0.05/0.08 are slightly lower | Current best clean diagnostic |
+| Wrong-gripper typed prior does not add a new clean bridge | wrong-gripper-only reaches 35.19%; no-op+wrong-gripper 0.02 ties 35.25%; no-op+wrong-gripper 0.04 drops to 35.13% | Negative/tie diagnostic |
 | The proposal gap is now quantified | `paper_analysis.md` reports best clean +5.51 pp over canonical h16, same-state no-expert +27.25 pp, leaving a +21.74 pp clean-to-same-state gap | Core paper tension |
 | Policy fallback is not the same-state mechanism | adding a policy baseline candidate to the no-expert same-state lattice drops 56.99% to 40.70% even with margin 0.00 | Negative diagnostic |
 | Z-score retrieval metric does not help | z-score rows reach 32.23-32.81%, below raw retrieval | Negative diagnostic |
@@ -58,18 +59,19 @@ clean proposal result, the intended main rows are:
 13. K2 train-state residual retrieval, typed safe families + advantage margin: 35.01%
 14. K4 mean-by-type tangent consensus: 34.96%
 15. K4 mean-by-type tangent consensus + typed no-op prior 0.025-0.035: 35.25%
-16. K2 broad tangent ray-search: 34.96%
-17. K1/K2 tight tangent ray-search: 34.84% / 34.84%
-18. K4 tight tangent ray-search: 34.55%
-19. Residual-tangent distillation policy: 28.87%
-20. Z-score residual retrieval: 32.23-32.81%
-21. Train-family reliability prior: 33.28-33.33%
-22. Residual+Gaussian hybrid K32/K64: 31.30% / 30.90%
-23. Lattice, near-miss only: 55.94%
-24. Lattice, no expert: 56.99%
-25. Lattice, no expert + policy baseline candidate: 40.70%
-26. Lattice, full: 69.33%
-27. Oracle ceiling: 86.78%
+16. Wrong-gripper prior / no-op+wrong-gripper prior: 35.19% / 35.25%
+17. K2 broad tangent ray-search: 34.96%
+18. K1/K2 tight tangent ray-search: 34.84% / 34.84%
+19. K4 tight tangent ray-search: 34.55%
+20. Residual-tangent distillation policy: 28.87%
+21. Z-score residual retrieval: 32.23-32.81%
+22. Train-family reliability prior: 33.28-33.33%
+23. Residual+Gaussian hybrid K32/K64: 31.30% / 30.90%
+24. Lattice, near-miss only: 55.94%
+25. Lattice, no expert: 56.99%
+26. Lattice, no expert + policy baseline candidate: 40.70%
+27. Lattice, full: 69.33%
+28. Oracle ceiling: 86.78%
 
 ## Novelty Framing
 
@@ -97,7 +99,7 @@ test-time search. The cleaner novelty is:
 
 ## Job Status
 
-Last checked: `2026-06-28 22:10 UTC`. The counterfactual tangent ray-search batch
+Last checked: `2026-06-29 00:41 UTC`. The counterfactual tangent ray-search batch
 completed, and the typed no-op residual-prior clean sweep completed after a passing
 CPU smoke.
 
@@ -166,6 +168,15 @@ CPU smoke.
   35.19%, 35.25%, and 35.25%; `0.025`/`0.03`/`0.035` form a small best plateau.
   Summary jobs `14884376`/`14884378`/`14884380`/`14884382` and rebuild job
   `14884383` completed.
+- `14890019`: completed CPU smoke for two-family candidate-type bonuses
+  (`residual_no_op=0.03`, `residual_wrong_gripper=0.02`), validating multi-type
+  parsing before GPU rollout.
+- `14890071`/`14890073`/`14890075`/`14890077`: completed wrong-gripper and
+  no-op+wrong-gripper typed-prior GPU sweeps. Results are 35.19% for
+  wrong-gripper-only, 35.25% for no-op 0.03 + wrong-gripper 0.02, 35.13% for
+  no-op 0.03 + wrong-gripper 0.04, and 35.25% for no-op 0.025 + wrong-gripper
+  0.02. Summary jobs `14890072`/`14890074`/`14890076`/`14890078` and rebuild
+  job `14890079` 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
@@ -194,5 +205,6 @@ CPU smoke.
   selection histograms when writing reviewer-facing tables.
 - Treat z-score retrieval, repaired train-family reliability priors, Gaussian hybrids,
   field optimization, field-teacher/tangent distillation, policy-relative anchoring, tangent consensus,
+  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.

scripts/build_paper_analysis.py

@@ -59,6 +59,38 @@ METHODS = [
             "k4s040_safe_margin0p20_mean_by_type_summary.json"
         ),
     ),
+    MethodSpec(
+        key="residual_k4_kernel_consensus",
+        label="K4 kernel-weighted tangent consensus",
+        summary_path=(
+            "h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_"
+            "k4s040_safe_margin0p20_kernel_mean_by_type_summary.json"
+        ),
+    ),
+    MethodSpec(
+        key="residual_k4_kernel_consensus_noopbonus003",
+        label="K4 kernel-weighted tangent consensus, no-op bonus 0.03",
+        summary_path=(
+            "h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_"
+            "k4s040_safe_margin0p20_kernel_mean_by_type_noopbonus0p03_summary.json"
+        ),
+    ),
+    MethodSpec(
+        key="residual_k4_kernel_consensus_s035_noopbonus003",
+        label="K4 kernel-weighted tangent consensus, scale 0.35, no-op bonus 0.03",
+        summary_path=(
+            "h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_"
+            "k4s035_safe_margin0p20_kernel_mean_by_type_noopbonus0p03_summary.json"
+        ),
+    ),
+    MethodSpec(
+        key="residual_k4_kernel_consensus_s045_noopbonus003",
+        label="K4 kernel-weighted tangent consensus, scale 0.45, no-op bonus 0.03",
+        summary_path=(
+            "h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_"
+            "k4s045_safe_margin0p20_kernel_mean_by_type_noopbonus0p03_summary.json"
+        ),
+    ),
     MethodSpec(
         key="residual_k4_consensus_noopbonus003",
         label="K4 mean-by-type tangent consensus, no-op bonus 0.03",
@@ -521,6 +553,10 @@ def _render_markdown(report: dict[str, Any]) -> str:
     for key in [
         "best_clean_residual_k2",
         "residual_k4_consensus",
+        "residual_k4_kernel_consensus",
+        "residual_k4_kernel_consensus_noopbonus003",
+        "residual_k4_kernel_consensus_s035_noopbonus003",
+        "residual_k4_kernel_consensus_s045_noopbonus003",
         "residual_k4_consensus_noopbonus003",
         "residual_k4_consensus_noopbonus001",
         "residual_k4_consensus_noopbonus002",

scripts/build_paper_table_status.py

@@ -335,6 +335,46 @@ SPECS = [
         story_role="counterfactual tangent consensus near-tie ablation",
         pending_job="14868699/14868700",
     ),
+    ResultSpec(
+        key="retrieval_residual_k4_kernel_mean",
+        label="K4 kernel-weighted residual retrieval, scale 0.40, margin 0.20",
+        path="h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_k4s040_safe_margin0p20_kernel_mean_by_type_summary.json",
+        clean_deployment="yes",
+        same_state_proposals="no",
+        expert_proposal="no",
+        story_role="local counterfactual tangent-field interpolation",
+        pending_job="14891067/14891083",
+    ),
+    ResultSpec(
+        key="retrieval_residual_k4_kernel_mean_noopbonus003",
+        label="K4 kernel-weighted residual retrieval, scale 0.40, margin 0.20, no-op residual bonus 0.03",
+        path="h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_k4s040_safe_margin0p20_kernel_mean_by_type_noopbonus0p03_summary.json",
+        clean_deployment="yes",
+        same_state_proposals="no",
+        expert_proposal="no",
+        story_role="local counterfactual tangent-field interpolation with sparse-action prior",
+        pending_job="14891072/14891085",
+    ),
+    ResultSpec(
+        key="retrieval_residual_k4_kernel_mean_s035_noopbonus003",
+        label="K4 kernel-weighted residual retrieval, scale 0.35, margin 0.20, no-op residual bonus 0.03",
+        path="h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_k4s035_safe_margin0p20_kernel_mean_by_type_noopbonus0p03_summary.json",
+        clean_deployment="yes",
+        same_state_proposals="no",
+        expert_proposal="no",
+        story_role="local counterfactual tangent-field interpolation scale check",
+        pending_job="14891076/14891087",
+    ),
+    ResultSpec(
+        key="retrieval_residual_k4_kernel_mean_s045_noopbonus003",
+        label="K4 kernel-weighted residual retrieval, scale 0.45, margin 0.20, no-op residual bonus 0.03",
+        path="h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_k4s045_safe_margin0p20_kernel_mean_by_type_noopbonus0p03_summary.json",
+        clean_deployment="yes",
+        same_state_proposals="no",
+        expert_proposal="no",
+        story_role="local counterfactual tangent-field interpolation scale check",
+        pending_job="14891082/14891088",
+    ),
     ResultSpec(
         key="retrieval_residual_k4_mean_noopbonus003",
         label="K4 mean-by-type residual retrieval, scale 0.40, margin 0.20, no-op residual bonus 0.03",

scripts/eval_maniskill_policy_rollout.py

@@ -148,9 +148,10 @@ def main(argv: list[str] | None = None) -> int:
     )
     parser.add_argument(
         "--retrieval-residual-reduce",
-        choices=("none", "mean_by_type", "median_by_type"),
+        choices=("none", "mean_by_type", "median_by_type", "kernel_mean_by_type"),
         default="none",
-        help="Optional consensus reduction over retrieved residuals with the same candidate type.",
+        help="Optional consensus reduction over retrieved residuals with the same candidate type. "
+        "'kernel_mean_by_type' weights source residuals by train-state retrieval distance.",
     )
     parser.add_argument(
         "--lattice-exclude-types",