Auto-sync: 2026-06-28 22:04:20 (part 3)

results/h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_k4_fieldsoftmax_grid_safe_margin0p00_noopbonus0p03_summary.json

@@ -0,0 +1,337 @@
+{
+  "run_root": "/scratch/knguy52/dovla/experiments/dovla_h16_policy_ckpt_runs",
+  "objective": "near_miss_policy_bc5",
+  "out_name": "policy_rollout_retrieval_residual_k4_fieldsoftmax_grid_safe_margin0p00_noopbonus0p03.json",
+  "num_completed": 3,
+  "baseline_h4_policy_success": 0.2967,
+  "baseline_h16_rank_checkpoint_success": 0.29739130434782607,
+  "mean_success": 0.34840579710144925,
+  "std_success": 0.007028611972743255,
+  "mean_progress": 0.5657099829109359,
+  "mean_action_mse_to_best": 0.41668070839151095,
+  "gain_vs_h4": 0.051705797101449236,
+  "gain_vs_h16_rank_checkpoint": 0.051014492753623186,
+  "rows": [
+    {
+      "seed": 0,
+      "path": "/scratch/knguy52/dovla/experiments/dovla_h16_policy_ckpt_runs/near_miss_policy_bc5/seed_0/policy_rollout_retrieval_residual_k4_fieldsoftmax_grid_safe_margin0p00_noopbonus0p03.json",
+      "num_groups": 575,
+      "selection_mode": "retrieval_residual",
+      "num_candidates": 4,
+      "candidate_sigma": 0.0,
+      "selection_margin": 0.0,
+      "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": 1.0,
+      "retrieval_residual_scales": [
+        0.35,
+        0.4,
+        0.45
+      ],
+      "retrieval_residual_anchor": "expert",
+      "retrieval_residual_reduce": "field_softmax",
+      "candidate_type_bonuses": {
+        "residual_no_op": 0.03
+      },
+      "selected_residual_scale_counts": {
+        "0.35": 57,
+        "0.4": 3,
+        "0.45": 65
+      },
+      "policy_rollout_success_rate": 0.3408695652173913,
+      "policy_rollout_progress": 0.5521027563876756,
+      "oracle_success_rate": 0.8573913043478261,
+      "action_mse_to_best": 0.39842233470762556,
+      "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": 0.331810000371764,
+          "expert_success_rate": 0.8865979381443299,
+          "num_groups": 97,
+          "oracle_success_rate": 0.9278350515463918,
+          "policy_expert_regret": 1.1049151976698453,
+          "policy_oracle_regret": 1.1313306753782881,
+          "policy_rollout_progress": 0.5655002526401245,
+          "policy_rollout_success_rate": 0.20618556701030927,
+          "restore_max_error": 2.384185791015625e-07
+        },
+        "PickCube-v1": {
+          "action_mse_to_best": 0.3170182326844392,
+          "expert_success_rate": 0.9375,
+          "num_groups": 208,
+          "oracle_success_rate": 0.9471153846153846,
+          "policy_expert_regret": 0.980829830120246,
+          "policy_oracle_regret": 0.9942526975467515,
+          "policy_rollout_progress": 0.6060521983219167,
+          "policy_rollout_success_rate": 0.3317307692307692,
+          "restore_max_error": 2.384185791015625e-07
+        },
+        "PullCube-v1": {
+          "action_mse_to_best": 0.5987812528362522,
+          "expert_success_rate": 0.19480519480519481,
+          "num_groups": 77,
+          "oracle_success_rate": 0.36363636363636365,
+          "policy_expert_regret": 0.19700949599772885,
+          "policy_oracle_regret": 0.3710168403839426,
+          "policy_rollout_progress": 0.3334215797012212,
+          "policy_rollout_success_rate": 0.23376623376623376,
+          "restore_max_error": 2.3096799850463867e-07
+        },
+        "PushCube-v1": {
+          "action_mse_to_best": 0.3739329240423568,
+          "expert_success_rate": 0.8617021276595744,
+          "num_groups": 94,
+          "oracle_success_rate": 0.9893617021276596,
+          "policy_expert_regret": 0.5508279580068081,
+          "policy_oracle_regret": 0.6283786044792926,
+          "policy_rollout_progress": 0.7014086295632606,
+          "policy_rollout_success_rate": 0.6702127659574468,
+          "restore_max_error": 2.384185791015625e-07
+        },
+        "StackCube-v1": {
+          "action_mse_to_best": 0.502137663576639,
+          "expert_success_rate": 0.6767676767676768,
+          "num_groups": 99,
+          "oracle_success_rate": 0.8585858585858586,
+          "policy_expert_regret": 0.9566237846709261,
+          "policy_oracle_regret": 1.1158919719734577,
+          "policy_rollout_progress": 0.45394778131234526,
+          "policy_rollout_success_rate": 0.26262626262626265,
+          "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_k4_fieldsoftmax_grid_safe_margin0p00_noopbonus0p03.json",
+      "num_groups": 575,
+      "selection_mode": "retrieval_residual",
+      "num_candidates": 4,
+      "candidate_sigma": 0.0,
+      "selection_margin": 0.0,
+      "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": 1.0,
+      "retrieval_residual_scales": [
+        0.35,
+        0.4,
+        0.45
+      ],
+      "retrieval_residual_anchor": "expert",
+      "retrieval_residual_reduce": "field_softmax",
+      "candidate_type_bonuses": {
+        "residual_no_op": 0.03
+      },
+      "selected_residual_scale_counts": {
+        "0.45": 54,
+        "0.35": 82,
+        "0.4": 9
+      },
+      "policy_rollout_success_rate": 0.34956521739130436,
+      "policy_rollout_progress": 0.5716547556253879,
+      "oracle_success_rate": 0.8695652173913043,
+      "action_mse_to_best": 0.40391011639538665,
+      "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": 0.33176156484925773,
+          "expert_success_rate": 0.8584070796460177,
+          "num_groups": 113,
+          "oracle_success_rate": 0.9380530973451328,
+          "policy_expert_regret": 0.9882967606582472,
+          "policy_oracle_regret": 1.0694300268603638,
+          "policy_rollout_progress": 0.6326816989257272,
+          "policy_rollout_success_rate": 0.23008849557522124,
+          "restore_max_error": 2.384185791015625e-07
+        },
+        "PickCube-v1": {
+          "action_mse_to_best": 0.3457128898969487,
+          "expert_success_rate": 0.9402173913043478,
+          "num_groups": 184,
+          "oracle_success_rate": 0.9456521739130435,
+          "policy_expert_regret": 1.0666085756991222,
+          "policy_oracle_regret": 1.0666325801740522,
+          "policy_rollout_progress": 0.5710404113582943,
+          "policy_rollout_success_rate": 0.29347826086956524,
+          "restore_max_error": 2.384185791015625e-07
+        },
+        "PullCube-v1": {
+          "action_mse_to_best": 0.5504576860878029,
+          "expert_success_rate": 0.25,
+          "num_groups": 76,
+          "oracle_success_rate": 0.40789473684210525,
+          "policy_expert_regret": 0.2950336691237202,
+          "policy_oracle_regret": 0.5187720929174439,
+          "policy_rollout_progress": 0.2991931551920348,
+          "policy_rollout_success_rate": 0.18421052631578946,
+          "restore_max_error": 2.384185791015625e-07
+        },
+        "PushCube-v1": {
+          "action_mse_to_best": 0.37546562134951084,
+          "expert_success_rate": 0.8198198198198198,
+          "num_groups": 111,
+          "oracle_success_rate": 1.0,
+          "policy_expert_regret": 0.31306307756149016,
+          "policy_oracle_regret": 0.33023064469431973,
+          "policy_rollout_progress": 0.8409405264768515,
+          "policy_rollout_success_rate": 0.8288288288288288,
+          "restore_max_error": 2.3795291781425476e-07
+        },
+        "StackCube-v1": {
+          "action_mse_to_best": 0.5234793433612519,
+          "expert_success_rate": 0.6923076923076923,
+          "num_groups": 91,
+          "oracle_success_rate": 0.8571428571428571,
+          "policy_expert_regret": 1.1237573417333455,
+          "policy_oracle_regret": 1.2834191888898283,
+          "policy_rollout_progress": 0.3961971272806545,
+          "policy_rollout_success_rate": 0.16483516483516483,
+          "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_k4_fieldsoftmax_grid_safe_margin0p00_noopbonus0p03.json",
+      "num_groups": 575,
+      "selection_mode": "retrieval_residual",
+      "num_candidates": 4,
+      "candidate_sigma": 0.0,
+      "selection_margin": 0.0,
+      "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": 1.0,
+      "retrieval_residual_scales": [
+        0.35,
+        0.4,
+        0.45
+      ],
+      "retrieval_residual_anchor": "expert",
+      "retrieval_residual_reduce": "field_softmax",
+      "candidate_type_bonuses": {
+        "residual_no_op": 0.03
+      },
+      "selected_residual_scale_counts": {
+        "0.35": 70,
+        "0.4": 13,
+        "0.45": 115
+      },
+      "policy_rollout_success_rate": 0.35478260869565215,
+      "policy_rollout_progress": 0.5733724367197441,
+      "oracle_success_rate": 0.8765217391304347,
+      "action_mse_to_best": 0.44770967407152057,
+      "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": 0.36598759308011114,
+          "expert_success_rate": 0.8229166666666666,
+          "num_groups": 96,
+          "oracle_success_rate": 0.9270833333333334,
+          "policy_expert_regret": 0.8476141745535036,
+          "policy_oracle_regret": 0.9538838644511998,
+          "policy_rollout_progress": 0.6463913254750272,
+          "policy_rollout_success_rate": 0.3333333333333333,
+          "restore_max_error": 1.955777406692505e-07
+        },
+        "PickCube-v1": {
+          "action_mse_to_best": 0.3816109399565242,
+          "expert_success_rate": 0.9444444444444444,
+          "num_groups": 198,
+          "oracle_success_rate": 0.9595959595959596,
+          "policy_expert_regret": 1.016553858643153,
+          "policy_oracle_regret": 1.0307835541721997,
+          "policy_rollout_progress": 0.6180411353425095,
+          "policy_rollout_success_rate": 0.31313131313131315,
+          "restore_max_error": 2.384185791015625e-07
+        },
+        "PullCube-v1": {
+          "action_mse_to_best": 0.647157601101531,
+          "expert_success_rate": 0.24444444444444444,
+          "num_groups": 90,
+          "oracle_success_rate": 0.4666666666666667,
+          "policy_expert_regret": 0.3434156913103329,
+          "policy_oracle_regret": 0.6106782714111938,
+          "policy_rollout_progress": 0.2862149072604047,
+          "policy_rollout_success_rate": 0.17777777777777778,
+          "restore_max_error": 2.384185791015625e-07
+        },
+        "PushCube-v1": {
+          "action_mse_to_best": 0.4132169201822564,
+          "expert_success_rate": 0.8514851485148515,
+          "num_groups": 101,
+          "oracle_success_rate": 1.0,
+          "policy_expert_regret": 0.36129841503530447,
+          "policy_oracle_regret": 0.398308122217065,
+          "policy_rollout_progress": 0.8096126698621429,
+          "policy_rollout_success_rate": 0.7920792079207921,
+          "restore_max_error": 2.384185791015625e-07
+        },
+        "StackCube-v1": {
+          "action_mse_to_best": 0.5195577167388465,
+          "expert_success_rate": 0.7666666666666667,
+          "num_groups": 90,
+          "oracle_success_rate": 0.9111111111111111,
+          "policy_expert_regret": 1.1750922918319702,
+          "policy_oracle_regret": 1.311360280878014,
+          "policy_rollout_progress": 0.41925797512133917,
+          "policy_rollout_success_rate": 0.15555555555555556,
+          "restore_max_error": 2.2351741790771484e-07
+        }
+      }
+    }
+  ]
+}

results/h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_k4_fieldsoftmax_grid_safe_margin0p00_noopbonus0p03_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_k4_fieldsoftmax_grid_safe_margin0p00_noopbonus0p03.json`
+Completed seeds: 3
+Baseline h=4 policy success: 29.67%
+Baseline h=16 rank-checkpoint success: 29.74%
+
+Mean success: 34.84% +/- 0.70%
+Gain vs h=16 rank checkpoint: +5.10%
+Mean progress: 56.57%
+Mean action MSE to best: 0.417
+
+| 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 | 4 | no | 4 | raw | expert | field_softmax | 0.00 | 1.00 | 0.35,0.40,0.45 | 0.000 | 0.00 | 0 | 0.00 | 34.09% | 55.21% | 85.74% | 0.398 |
+| 1 | retrieval_residual | 4 | no | 4 | raw | expert | field_softmax | 0.00 | 1.00 | 0.35,0.40,0.45 | 0.000 | 0.00 | 0 | 0.00 | 34.96% | 57.17% | 86.96% | 0.404 |
+| 2 | retrieval_residual | 4 | no | 4 | raw | expert | field_softmax | 0.00 | 1.00 | 0.35,0.40,0.45 | 0.000 | 0.00 | 0 | 0.00 | 35.48% | 57.34% | 87.65% | 0.448 |

results/paper_analysis.json

@@ -1,6 +1,6 @@
 {
   "best_clean_key": "residual_k4_consensus_noopbonus003",
-  "generated_utc": "2026-06-29T01:49:02+00:00",
+  "generated_utc": "2026-06-29T02:05:44+00:00",
   "mechanism_gap": {
     "best_clean_vs_direct_same_ckpt": 0.06956521739130428,
     "best_clean_vs_h16": 0.05507246376811592,
@@ -1269,19 +1269,223 @@
       "std_success": 0.015939393721585544
     },
     "residual_k4_fieldsoftmax_grid_margin000_noopbonus003": {
+      "ci95_success": 0.01746144927536228,
       "label": "K4 field-softmax tangent transport, margin 0.00, no-op bonus 0.03",
-      "missing": true,
-      "source": "results/h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_k4_fieldsoftmax_grid_safe_margin0p00_noopbonus0p03_summary.json"
+      "mean_action_mse_to_best": 0.41668070839151095,
+      "mean_progress": 0.5657099829109359,
+      "mean_success": 0.3484057971014492,
+      "num_completed": 3,
+      "per_task_success": {
+        "LiftPegUpright-v1": {
+          "mean_num_groups": 102.0,
+          "mean_success": 0.2565357986396213,
+          "std_success": 0.06757391138426685
+        },
+        "PickCube-v1": {
+          "mean_num_groups": 196.66666666666666,
+          "mean_success": 0.3127801144105492,
+          "std_success": 0.01912867231114315
+        },
+        "PullCube-v1": {
+          "mean_num_groups": 81.0,
+          "mean_success": 0.198584845953267,
+          "std_success": 0.030637274679270955
+        },
+        "PushCube-v1": {
+          "mean_num_groups": 102.0,
+          "mean_success": 0.7637069342356892,
+          "std_success": 0.08302712379548827
+        },
+        "StackCube-v1": {
+          "mean_num_groups": 93.33333333333333,
+          "mean_success": 0.19433899433899435,
+          "std_success": 0.05932024144167464
+        }
+      },
+      "seed_action_mse_to_best": {
+        "0": 0.39842233470762556,
+        "1": 0.40391011639538665,
+        "2": 0.44770967407152057
+      },
+      "seed_progress": {
+        "0": 0.5521027563876756,
+        "1": 0.5716547556253879,
+        "2": 0.5733724367197441
+      },
+      "seed_success": {
+        "0": 0.3408695652173913,
+        "1": 0.34956521739130436,
+        "2": 0.35478260869565215
+      },
+      "selected_candidate_type_counts": {
+        "retrieval_residual_field_softmax": 468,
+        "retrieval_residual_policy_residual": 1257
+      },
+      "selected_residual_scale_counts": {
+        "0.35": 209,
+        "0.4": 25,
+        "0.45": 234
+      },
+      "selected_type_outcomes": {
+        "retrieval_residual_field_softmax": {
+          "count": 468.0,
+          "mean_progress": 0.6394197220265324,
+          "success_rate": 0.4551282051282051
+        },
+        "retrieval_residual_policy_residual": {
+          "count": 1257.0,
+          "mean_progress": 0.5382667387533391,
+          "success_rate": 0.30867143993635643
+        }
+      },
+      "source": "results/h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_k4_fieldsoftmax_grid_safe_margin0p00_noopbonus0p03_summary.json",
+      "std_success": 0.007028611972743255
     },
     "residual_k4_fieldsoftmax_grid_margin005_noopbonus003": {
+      "ci95_success": 0.027439420289855042,
       "label": "K4 field-softmax tangent transport, margin 0.05, no-op bonus 0.03",
-      "missing": true,
-      "source": "results/h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_k4_fieldsoftmax_grid_safe_margin0p05_noopbonus0p03_summary.json"
+      "mean_action_mse_to_best": 0.40925245372929436,
+      "mean_progress": 0.5672646419938815,
+      "mean_success": 0.3507246376811594,
+      "num_completed": 3,
+      "per_task_success": {
+        "LiftPegUpright-v1": {
+          "mean_num_groups": 102.0,
+          "mean_success": 0.25997222475645976,
+          "std_success": 0.06389511734286676
+        },
+        "PickCube-v1": {
+          "mean_num_groups": 196.66666666666666,
+          "mean_success": 0.3126991768296116,
+          "std_success": 0.021898452407335747
+        },
+        "PullCube-v1": {
+          "mean_num_groups": 81.0,
+          "mean_success": 0.20838839523050048,
+          "std_success": 0.032937378023892676
+        },
+        "PushCube-v1": {
+          "mean_num_groups": 102.0,
+          "mean_success": 0.7607039312326863,
+          "std_success": 0.07958563709480587
+        },
+        "StackCube-v1": {
+          "mean_num_groups": 93.33333333333333,
+          "mean_success": 0.19774669774669773,
+          "std_success": 0.06525071919571808
+        }
+      },
+      "seed_action_mse_to_best": {
+        "0": 0.3920214500521188,
+        "1": 0.3982289221664162,
+        "2": 0.43750698896934803
+      },
+      "seed_progress": {
+        "0": 0.5518923319673733,
+        "1": 0.5683723092014374,
+        "2": 0.581529284812834
+      },
+      "seed_success": {
+        "0": 0.3443478260869565,
+        "1": 0.3443478260869565,
+        "2": 0.3634782608695652
+      },
+      "selected_candidate_type_counts": {
+        "retrieval_residual_field_softmax": 267,
+        "retrieval_residual_policy_residual": 1458
+      },
+      "selected_residual_scale_counts": {
+        "0.35": 79,
+        "0.4": 21,
+        "0.45": 167
+      },
+      "selected_type_outcomes": {
+        "retrieval_residual_field_softmax": {
+          "count": 267.0,
+          "mean_progress": 0.6703486572951078,
+          "success_rate": 0.5056179775280899
+        },
+        "retrieval_residual_policy_residual": {
+          "count": 1458.0,
+          "mean_progress": 0.5483871165580603,
+          "success_rate": 0.3223593964334705
+        }
+      },
+      "source": "results/h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_k4_fieldsoftmax_grid_safe_margin0p05_noopbonus0p03_summary.json",
+      "std_success": 0.011044961671453697
     },
     "residual_k4_fieldsoftmax_grid_margin010_noopbonus003": {
+      "ci95_success": 0.030652840142141275,
       "label": "K4 field-softmax tangent transport, margin 0.10, no-op bonus 0.03",
-      "missing": true,
-      "source": "results/h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_k4_fieldsoftmax_grid_safe_margin0p10_noopbonus0p03_summary.json"
+      "mean_action_mse_to_best": 0.4014113438483058,
+      "mean_progress": 0.5673825791396276,
+      "mean_success": 0.35188405797101446,
+      "num_completed": 3,
+      "per_task_success": {
+        "LiftPegUpright-v1": {
+          "mean_num_groups": 102.0,
+          "mean_success": 0.25997222475645976,
+          "std_success": 0.06389511734286676
+        },
+        "PickCube-v1": {
+          "mean_num_groups": 196.66666666666666,
+          "mean_success": 0.3174597141988446,
+          "std_success": 0.02906191304415352
+        },
+        "PullCube-v1": {
+          "mean_num_groups": 81.0,
+          "mean_success": 0.2046846915267968,
+          "std_success": 0.02913515326101664
+        },
+        "PushCube-v1": {
+          "mean_num_groups": 102.0,
+          "mean_success": 0.7577009282296832,
+          "std_success": 0.0763436450533849
+        },
+        "StackCube-v1": {
+          "mean_num_groups": 93.33333333333333,
+          "mean_success": 0.2011137011137011,
+          "std_success": 0.07105663963231741
+        }
+      },
+      "seed_action_mse_to_best": {
+        "0": 0.38557657074831103,
+        "1": 0.3929036483346768,
+        "2": 0.4257538124619295
+      },
+      "seed_progress": {
+        "0": 0.5548038772388321,
+        "1": 0.5665490727742081,
+        "2": 0.5807947874058427
+      },
+      "seed_success": {
+        "0": 0.34956521739130436,
+        "1": 0.3408695652173913,
+        "2": 0.3652173913043478
+      },
+      "selected_candidate_type_counts": {
+        "retrieval_residual_field_softmax": 128,
+        "retrieval_residual_policy_residual": 1597
+      },
+      "selected_residual_scale_counts": {
+        "0.35": 32,
+        "0.4": 8,
+        "0.45": 88
+      },
+      "selected_type_outcomes": {
+        "retrieval_residual_field_softmax": {
+          "count": 128.0,
+          "mean_progress": 0.7108588826667983,
+          "success_rate": 0.5625
+        },
+        "retrieval_residual_policy_residual": {
+          "count": 1597.0,
+          "mean_progress": 0.5558829129834111,
+          "success_rate": 0.33500313087038197
+        }
+      },
+      "source": "results/h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_k4_fieldsoftmax_grid_safe_margin0p10_noopbonus0p03_summary.json",
+      "std_success": 0.01233843284277841
     },
     "residual_k4_fieldsoftmax_grid_noopbonus003": {
       "ci95_success": 0.038402227815547745,

results/paper_analysis.md

@@ -1,6 +1,6 @@
 # Paper Analysis
 
-Generated: `2026-06-29T01:49:02+00:00`
+Generated: `2026-06-29T02:05:44+00:00`
 
 ## Main Seed Statistics
 
@@ -17,9 +17,9 @@ Generated: `2026-06-29T01:49:02+00:00`
 | residual_k4_kernel_consensus_s045_noopbonus003 | K4 kernel-weighted tangent consensus, scale 0.45, no-op bonus 0.03 | 3 | 35.19% +/- 1.02 | +/- 2.53 | 56.71% | 0.397 | +5.45 pp |
 | residual_k4_fieldsoftmax_grid | K4 field-softmax tangent transport, scales 0.35/0.40/0.45 | 3 | 34.96% +/- 1.59 | +/- 3.96 | 56.52% | 0.397 | +5.22 pp |
 | residual_k4_fieldsoftmax_grid_noopbonus003 | K4 field-softmax tangent transport, scales 0.35/0.40/0.45, no-op bonus 0.03 | 3 | 34.96% +/- 1.55 | +/- 3.84 | 56.55% | 0.397 | +5.22 pp |
-| residual_k4_fieldsoftmax_grid_margin010_noopbonus003 | K4 field-softmax tangent transport, margin 0.10, no-op bonus 0.03 | 0 | missing | missing | missing | missing | missing |
-| residual_k4_fieldsoftmax_grid_margin005_noopbonus003 | K4 field-softmax tangent transport, margin 0.05, no-op bonus 0.03 | 0 | missing | missing | missing | missing | missing |
-| residual_k4_fieldsoftmax_grid_margin000_noopbonus003 | K4 field-softmax tangent transport, margin 0.00, no-op bonus 0.03 | 0 | missing | missing | missing | missing | missing |
+| residual_k4_fieldsoftmax_grid_margin010_noopbonus003 | K4 field-softmax tangent transport, margin 0.10, no-op bonus 0.03 | 3 | 35.19% +/- 1.23 | +/- 3.07 | 56.74% | 0.401 | +5.45 pp |
+| residual_k4_fieldsoftmax_grid_margin005_noopbonus003 | K4 field-softmax tangent transport, margin 0.05, no-op bonus 0.03 | 3 | 35.07% +/- 1.10 | +/- 2.74 | 56.73% | 0.409 | +5.33 pp |
+| residual_k4_fieldsoftmax_grid_margin000_noopbonus003 | K4 field-softmax tangent transport, margin 0.00, no-op bonus 0.03 | 3 | 34.84% +/- 0.70 | +/- 1.75 | 56.57% | 0.417 | +5.10 pp |
 | 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_noopbonus001 | K4 mean-by-type tangent consensus, no-op bonus 0.01 | 3 | 35.19% +/- 1.32 | +/- 3.27 | 56.63% | 0.395 | +5.45 pp |
@@ -101,6 +101,12 @@ These rows are measured from raw rollout rows. In residual retrieval, `policy_re
 | residual_k4_fieldsoftmax_grid | retrieval_residual_field_softmax | 37 | 59.46% | 72.64% |
 | residual_k4_fieldsoftmax_grid_noopbonus003 | retrieval_residual_policy_residual | 1685 | 34.36% | 56.11% |
 | residual_k4_fieldsoftmax_grid_noopbonus003 | retrieval_residual_field_softmax | 40 | 60.00% | 74.94% |
+| residual_k4_fieldsoftmax_grid_margin010_noopbonus003 | retrieval_residual_policy_residual | 1597 | 33.50% | 55.59% |
+| residual_k4_fieldsoftmax_grid_margin010_noopbonus003 | retrieval_residual_field_softmax | 128 | 56.25% | 71.09% |
+| residual_k4_fieldsoftmax_grid_margin005_noopbonus003 | retrieval_residual_policy_residual | 1458 | 32.24% | 54.84% |
+| residual_k4_fieldsoftmax_grid_margin005_noopbonus003 | retrieval_residual_field_softmax | 267 | 50.56% | 67.03% |
+| residual_k4_fieldsoftmax_grid_margin000_noopbonus003 | retrieval_residual_policy_residual | 1257 | 30.87% | 53.83% |
+| residual_k4_fieldsoftmax_grid_margin000_noopbonus003 | retrieval_residual_field_softmax | 468 | 45.51% | 63.94% |
 | residual_k8_fieldsoftmax_grid_noopbonus003 | retrieval_residual_policy_residual | 1688 | 34.30% | 56.09% |
 | residual_k8_fieldsoftmax_grid_noopbonus003 | retrieval_residual_field_softmax | 37 | 59.46% | 77.72% |
 | residual_k4_consensus_noopbonus003 | retrieval_residual_policy_residual | 1649 | 34.57% | 56.12% |

results/paper_core_results.md

@@ -41,6 +41,7 @@ and the remaining clean-to-same-state proposal gap is `+21.74 pp`.
 | 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 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 |
+| K4 field-softmax residual barycenter + no-op prior 0.03 | No | No | 34.84-35.19% | +5.10-5.45 pp | Field-conditioned aggregation finds high-value sparse corrections, but lower margins over-select them; it does not beat the equal mean-consensus no-op plateau |
 | 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 |
@@ -80,15 +81,16 @@ Suggested main-table rows:
 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. K4 kernel-weighted residual consensus + no-op prior diagnostics
-16. K4 mean-by-type residual retrieval + wrong-gripper typed-prior diagnostics
-17. K2 broad tangent ray-search
-18. Residual-tangent distillation policy
-19. Residual+Gaussian hybrid, K32 sigma0.35
-20. Lattice, near-miss only
-21. Lattice, no expert
-22. Lattice, no expert + policy baseline candidate
-23. Lattice, full
-24. Oracle ceiling
+16. K4 field-softmax residual barycenter + margin diagnostics
+17. K4 mean-by-type residual retrieval + wrong-gripper typed-prior diagnostics
+18. K2 broad tangent ray-search
+19. Residual-tangent distillation policy
+20. Residual+Gaussian hybrid, K32 sigma0.35
+21. Lattice, near-miss only
+22. Lattice, no expert
+23. Lattice, no expert + policy baseline candidate
+24. Lattice, full
+25. Oracle ceiling
 
 Suggested claim:
 
@@ -97,7 +99,7 @@ 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, kernel-weighted tangent interpolation, tangent ray-search, wrong-gripper typed priors, and same-state policy-baseline fallback fail to improve the main rows.
+> 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

@@ -29,6 +29,7 @@ when queried on proposal geometry that matches those local counterfactuals.
 | Clean residual transport behaves like sparse intervention | `paper_analysis.md` shows the best clean row abstains to zero-residual policy on 95.6% of states, while selected nonzero no-op residuals succeed at 52.83% vs 34.57% for abstention | Stronger clean-mechanism framing |
 | Tangent consensus is close but needs sparse typing | K4 mean-by-type residual consensus reaches 34.96%; a small no-op residual prior plateau at 0.025-0.035 raises it to 35.25% | Current best clean result |
 | Kernel-weighted tangent interpolation does not beat equal consensus | K4 kernel-weighted residual consensus reaches 34.96%; with no-op prior and scales 0.35/0.40/0.45 it reaches 35.13%/35.19%/35.19%, below the 35.25% mean-consensus plateau | Negative/near-tie diagnostic |
+| Field-conditioned tangent barycenters identify good sparse corrections but do not close the proposal gap | K4 field-softmax transport reaches 34.96%; with no-op prior and margins 0.10/0.05/0.00 it reaches 35.19%/35.07%/34.84%. Selected aggregate residuals are high-value (up to 60.00% success), but selecting more of them degrades the global row | Negative/near-tie diagnostic |
 | Tangent ray-search does not beat the typed-prior clean row | K1/K2 tight scale-grid ray search reach 34.84%; K2 broad reaches 34.96%; K4 tight reaches 34.55%, all below the 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 |
@@ -61,19 +62,20 @@ clean proposal result, the intended main rows are:
 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. K4 kernel-weighted tangent consensus / + no-op prior: 34.96% / 35.19%
-17. Wrong-gripper prior / no-op+wrong-gripper prior: 35.19% / 35.25%
-18. K2 broad tangent ray-search: 34.96%
-19. K1/K2 tight tangent ray-search: 34.84% / 34.84%
-20. K4 tight tangent ray-search: 34.55%
-21. Residual-tangent distillation policy: 28.87%
-22. Z-score residual retrieval: 32.23-32.81%
-23. Train-family reliability prior: 33.28-33.33%
-24. Residual+Gaussian hybrid K32/K64: 31.30% / 30.90%
-25. Lattice, near-miss only: 55.94%
-26. Lattice, no expert: 56.99%
-27. Lattice, no expert + policy baseline candidate: 40.70%
-28. Lattice, full: 69.33%
-29. Oracle ceiling: 86.78%
+17. K4 field-softmax tangent transport / best margin sweep: 34.96% / 35.19%
+18. Wrong-gripper prior / no-op+wrong-gripper prior: 35.19% / 35.25%
+19. K2 broad tangent ray-search: 34.96%
+20. K1/K2 tight tangent ray-search: 34.84% / 34.84%
+21. K4 tight tangent ray-search: 34.55%
+22. Residual-tangent distillation policy: 28.87%
+23. Z-score residual retrieval: 32.23-32.81%
+24. Train-family reliability prior: 33.28-33.33%
+25. Residual+Gaussian hybrid K32/K64: 31.30% / 30.90%
+26. Lattice, near-miss only: 55.94%
+27. Lattice, no expert: 56.99%
+28. Lattice, no expert + policy baseline candidate: 40.70%
+29. Lattice, full: 69.33%
+30. Oracle ceiling: 86.78%
 
 ## Novelty Framing
 
@@ -101,9 +103,9 @@ test-time search. The cleaner novelty is:
 
 ## Job Status
 
-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.
+Last checked: `2026-06-29 02:00 UTC`. The field-conditioned tangent-barycenter
+batch completed after passing CPU smokes, and the paper table rebuild now includes
+the field-softmax rows.
 
 - `14858328`-`14858333`: completed train-split `field_selected_noexpert_bc5`;
   direct rollout is 26.84%, field-guided best is 27.65%.
@@ -188,6 +190,20 @@ CPU smoke.
   35.19%. Summary jobs `14891083`/`14891085`/`14891087`/`14891088` and rebuild
   job `14891089` completed. These are near-tie/negative diagnostics below the
   equal mean-consensus no-op plateau.
+- `14891870` and `14892092`: completed CPU smokes for the field-softmax residual
+  reducer, first validating the model-time aggregate path and then the final
+  candidate-bonus propagation.
+- `14891889`/`14891902`/`14891923`: completed K4/K8 field-softmax transport
+  sweeps. K4 field-softmax reaches 34.96% with or without no-op 0.03 at margin
+  `0.20`; K8 with no-op 0.03 reaches 34.84%. Summary jobs `14891934`/`14891946`/
+  `14891960` completed.
+- `14892958`/`14892975`/`14892990`: completed the K4 field-softmax no-op margin
+  sweep. Margins `0.10`, `0.05`, and `0.00` reach 35.19%, 35.07%, and 34.84%.
+  The selected field-softmax aggregates have high conditional success, but lower
+  margins over-select them and reduce the overall row, so this remains a
+  negative/near-tie diagnostic below the 35.25% mean-consensus no-op plateau.
+  Summary jobs `14893002`/`14893016`/`14893028` and rebuild job `14893069`
+  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
@@ -216,6 +232,7 @@ 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,
-  kernel-weighted tangent interpolation, 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.
+  kernel-weighted tangent interpolation, field-softmax tangent barycenters,
+  wrong-gripper typed priors, and same-state policy-baseline fallback as negative
+  or near-tie diagnostics that sharpen the story around local counterfactual
+  proposal geometry.

results/paper_table_status.json

@@ -736,14 +736,14 @@
       "story_role": "field-conditioned tangent transport abstention sweep",
       "fallback_success": null,
       "pending_job": "14892958/14893002",
-      "path_exists": false,
-      "status": "pending",
-      "success": null,
-      "std_success": null,
+      "path_exists": true,
+      "status": "complete",
+      "success": 0.3518840579710145,
+      "std_success": 0.01233843284277841,
       "completed_seeds": null,
-      "num_completed": null,
+      "num_completed": 3,
       "best_config": null,
-      "gain_vs_h16_policy": null
+      "gain_vs_h16_policy": 0.05449275362318845
     },
     {
       "key": "retrieval_residual_k4_fieldsoftmax_grid_margin005_noopbonus003",
@@ -755,14 +755,14 @@
       "story_role": "field-conditioned tangent transport abstention sweep",
       "fallback_success": null,
       "pending_job": "14892975/14893016",
-      "path_exists": false,
-      "status": "pending",
-      "success": null,
-      "std_success": null,
+      "path_exists": true,
+      "status": "complete",
+      "success": 0.3507246376811594,
+      "std_success": 0.011044961671453697,
       "completed_seeds": null,
-      "num_completed": null,
+      "num_completed": 3,
       "best_config": null,
-      "gain_vs_h16_policy": null
+      "gain_vs_h16_policy": 0.053333333333333344
     },
     {
       "key": "retrieval_residual_k4_fieldsoftmax_grid_margin000_noopbonus003",
@@ -774,14 +774,14 @@
       "story_role": "field-conditioned tangent transport no-abstention diagnostic",
       "fallback_success": null,
       "pending_job": "14892990/14893028",
-      "path_exists": false,
-      "status": "pending",
-      "success": null,
-      "std_success": null,
+      "path_exists": true,
+      "status": "complete",
+      "success": 0.34840579710144925,
+      "std_success": 0.007028611972743255,
       "completed_seeds": null,
-      "num_completed": null,
+      "num_completed": 3,
       "best_config": null,
-      "gain_vs_h16_policy": null
+      "gain_vs_h16_policy": 0.051014492753623186
     },
     {
       "key": "retrieval_residual_k8_fieldsoftmax_grid_noopbonus003",

results/paper_table_status.md

@@ -41,9 +41,9 @@ Baseline h=16 policy: 29.74%
 | retrieval_residual_k4_kernel_mean_s045_noopbonus003 | K4 kernel-weighted residual retrieval, scale 0.45, margin 0.20, no-op residual bonus 0.03 | complete | 35.19% | +5.45 pp | yes | no | no | local counterfactual tangent-field interpolation scale check |
 | retrieval_residual_k4_fieldsoftmax_grid | K4 field-softmax residual transport, safe residuals, scales 0.35/0.40/0.45, margin 0.20 | complete | 34.96% | +5.22 pp | yes | no | no | field-conditioned counterfactual tangent transport |
 | retrieval_residual_k4_fieldsoftmax_grid_noopbonus003 | K4 field-softmax residual transport, safe residuals, scales 0.35/0.40/0.45, margin 0.20, no-op residual bonus 0.03 | complete | 34.96% | +5.22 pp | yes | no | no | field-conditioned tangent transport with sparse-action prior |
-| retrieval_residual_k4_fieldsoftmax_grid_margin010_noopbonus003 | K4 field-softmax residual transport, safe residuals, scales 0.35/0.40/0.45, margin 0.10, no-op residual bonus 0.03 | pending 14892958/14893002 | pending | pending | yes | no | no | field-conditioned tangent transport abstention sweep |
-| retrieval_residual_k4_fieldsoftmax_grid_margin005_noopbonus003 | K4 field-softmax residual transport, safe residuals, scales 0.35/0.40/0.45, margin 0.05, no-op residual bonus 0.03 | pending 14892975/14893016 | pending | pending | yes | no | no | field-conditioned tangent transport abstention sweep |
-| retrieval_residual_k4_fieldsoftmax_grid_margin000_noopbonus003 | K4 field-softmax residual transport, safe residuals, scales 0.35/0.40/0.45, margin 0.00, no-op residual bonus 0.03 | pending 14892990/14893028 | pending | pending | yes | no | no | field-conditioned tangent transport no-abstention diagnostic |
+| retrieval_residual_k4_fieldsoftmax_grid_margin010_noopbonus003 | K4 field-softmax residual transport, safe residuals, scales 0.35/0.40/0.45, margin 0.10, no-op residual bonus 0.03 | complete | 35.19% | +5.45 pp | yes | no | no | field-conditioned tangent transport abstention sweep |
+| retrieval_residual_k4_fieldsoftmax_grid_margin005_noopbonus003 | K4 field-softmax residual transport, safe residuals, scales 0.35/0.40/0.45, margin 0.05, no-op residual bonus 0.03 | complete | 35.07% | +5.33 pp | yes | no | no | field-conditioned tangent transport abstention sweep |
+| retrieval_residual_k4_fieldsoftmax_grid_margin000_noopbonus003 | K4 field-softmax residual transport, safe residuals, scales 0.35/0.40/0.45, margin 0.00, no-op residual bonus 0.03 | complete | 34.84% | +5.10 pp | yes | no | no | field-conditioned tangent transport no-abstention diagnostic |
 | 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_noopbonus001 | K4 mean-by-type residual retrieval, scale 0.40, margin 0.20, no-op residual bonus 0.01 | complete | 35.19% | +5.45 pp | yes | no | no | typed sparse-intervention prior fine sweep |

scripts/build_paper_table_status.py

@@ -285,6 +285,16 @@ SPECS = [
         story_role="previous best counterfactual advantage abstention",
         pending_job="14862936/14862937",
     ),
+    ResultSpec(
+        key="retrieval_residual_taskrelative_knn2_scale040_safe_margin020",
+        label="K2 task-relative residual retrieval, scale 0.40, safe residuals, advantage margin 0.20",
+        path="h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_taskrelative_knn2_scale0p40_safe_types_margin0p20_summary.json",
+        clean_deployment="yes",
+        same_state_proposals="no",
+        expert_proposal="no",
+        story_role="task-relative state metric for counterfactual tangent retrieval",
+        pending_job="14893475/14893476",
+    ),
     ResultSpec(
         key="retrieval_residual_k1grid_tight_safe_ray_margin020",
         label="K1 train-state residual ray search, safe residuals, scales 0.30/0.40/0.50, advantage margin 0.20",
@@ -445,6 +455,16 @@ SPECS = [
         story_role="current best clean typed sparse-intervention prior",
         pending_job="14883919/14883920",
     ),
+    ResultSpec(
+        key="retrieval_residual_taskrelative_k4_mean_noopbonus003",
+        label="K4 task-relative mean-by-type residual retrieval, scale 0.40, margin 0.20, no-op residual bonus 0.03",
+        path="h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_taskrelative_k4s040_safe_margin0p20_mean_by_type_noopbonus0p03_summary.json",
+        clean_deployment="yes",
+        same_state_proposals="no",
+        expert_proposal="no",
+        story_role="task-relative state metric for typed sparse-intervention prior",
+        pending_job="14893473/14893474",
+    ),
     ResultSpec(
         key="retrieval_residual_k4_mean_noopbonus001",
         label="K4 mean-by-type residual retrieval, scale 0.40, margin 0.20, no-op residual bonus 0.01",

scripts/eval_maniskill_policy_rollout.py

@@ -116,10 +116,11 @@ def main(argv: list[str] | None = None) -> int:
     )
     parser.add_argument(
         "--retrieval-metric",
-        choices=("raw", "zscore"),
+        choices=("raw", "zscore", "task_relative"),
         default="raw",
         help="State-space metric for retrieval proposals. 'raw' preserves earlier results; "
-        "'zscore' standardizes each task's train-bank features before nearest-neighbor lookup.",
+        "'zscore' standardizes each task's train-bank features before nearest-neighbor lookup; "
+        "'task_relative' retrieves by target/reference actor pose rather than full robot state.",
     )
     parser.add_argument(
         "--retrieval-type-min-success",