Auto-sync: 2026-06-28 13:57:28 (part 2)

results/h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_k2grid_broad_safe_ray_margin0p20_summary.json

@@ -0,0 +1,316 @@
+{
+  "run_root": "/scratch/knguy52/dovla/experiments/dovla_h16_policy_ckpt_runs",
+  "objective": "near_miss_policy_bc5",
+  "out_name": "policy_rollout_retrieval_residual_k2grid_broad_safe_ray_margin0p20.json",
+  "num_completed": 3,
+  "baseline_h4_policy_success": 0.2967,
+  "baseline_h16_rank_checkpoint_success": 0.29739130434782607,
+  "mean_success": 0.34956521739130436,
+  "std_success": 0.015061311370164159,
+  "mean_progress": 0.5672198920562453,
+  "mean_action_mse_to_best": 0.41953608097427564,
+  "gain_vs_h4": 0.05286521739130434,
+  "gain_vs_h16_rank_checkpoint": 0.05217391304347829,
+  "rows": [
+    {
+      "seed": 0,
+      "path": "/scratch/knguy52/dovla/experiments/dovla_h16_policy_ckpt_runs/near_miss_policy_bc5/seed_0/policy_rollout_retrieval_residual_k2grid_broad_safe_ray_margin0p20.json",
+      "num_groups": 575,
+      "selection_mode": "retrieval_residual",
+      "num_candidates": 128,
+      "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": 2,
+      "retrieval_metric": "raw",
+      "retrieval_type_min_success": 0.0,
+      "retrieval_residual_scale": 0.4,
+      "retrieval_residual_scales": [
+        0.2,
+        0.35,
+        0.5,
+        0.65
+      ],
+      "retrieval_residual_anchor": "expert",
+      "retrieval_residual_reduce": "none",
+      "policy_rollout_success_rate": 0.3408695652173913,
+      "policy_rollout_progress": 0.5572800162560104,
+      "oracle_success_rate": 0.8573913043478261,
+      "action_mse_to_best": 0.40957117927544145,
+      "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": 0.34625995102981777,
+          "expert_success_rate": 0.8865979381443299,
+          "num_groups": 97,
+          "oracle_success_rate": 0.9278350515463918,
+          "policy_expert_regret": 1.0811832799124963,
+          "policy_oracle_regret": 1.1075987576209392,
+          "policy_rollout_progress": 0.5686136136964425,
+          "policy_rollout_success_rate": 0.2268041237113402,
+          "restore_max_error": 4.76837158203125e-07
+        },
+        "PickCube-v1": {
+          "action_mse_to_best": 0.3053258811123669,
+          "expert_success_rate": 0.9375,
+          "num_groups": 208,
+          "oracle_success_rate": 0.9471153846153846,
+          "policy_expert_regret": 0.9611840278554994,
+          "policy_oracle_regret": 0.9746068952820048,
+          "policy_rollout_progress": 0.6160826159712787,
+          "policy_rollout_success_rate": 0.34134615384615385,
+          "restore_max_error": 4.76837158203125e-07
+        },
+        "PullCube-v1": {
+          "action_mse_to_best": 0.6023399162215072,
+          "expert_success_rate": 0.19480519480519481,
+          "num_groups": 77,
+          "oracle_success_rate": 0.36363636363636365,
+          "policy_expert_regret": 0.1975214997844698,
+          "policy_oracle_regret": 0.3693149077264568,
+          "policy_rollout_progress": 0.34371752446432685,
+          "policy_rollout_success_rate": 0.24675324675324675,
+          "restore_max_error": 4.76837158203125e-07
+        },
+        "PushCube-v1": {
+          "action_mse_to_best": 0.3695450182211526,
+          "expert_success_rate": 0.8617021276595744,
+          "num_groups": 94,
+          "oracle_success_rate": 0.9893617021276596,
+          "policy_expert_regret": 0.5719942746961371,
+          "policy_oracle_regret": 0.6495450359392674,
+          "policy_rollout_progress": 0.6908804959756263,
+          "policy_rollout_success_rate": 0.6595744680851063,
+          "restore_max_error": 4.76837158203125e-07
+        },
+        "StackCube-v1": {
+          "action_mse_to_best": 0.5786972151542402,
+          "expert_success_rate": 0.6767676767676768,
+          "num_groups": 99,
+          "oracle_success_rate": 0.8585858585858586,
+          "policy_expert_regret": 0.966132034105484,
+          "policy_oracle_regret": 1.1368058207962248,
+          "policy_rollout_progress": 0.46188170395114203,
+          "policy_rollout_success_rate": 0.2222222222222222,
+          "restore_max_error": 3.948807716369629e-07
+        }
+      }
+    },
+    {
+      "seed": 1,
+      "path": "/scratch/knguy52/dovla/experiments/dovla_h16_policy_ckpt_runs/near_miss_policy_bc5/seed_1/policy_rollout_retrieval_residual_k2grid_broad_safe_ray_margin0p20.json",
+      "num_groups": 575,
+      "selection_mode": "retrieval_residual",
+      "num_candidates": 128,
+      "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": 2,
+      "retrieval_metric": "raw",
+      "retrieval_type_min_success": 0.0,
+      "retrieval_residual_scale": 0.4,
+      "retrieval_residual_scales": [
+        0.2,
+        0.35,
+        0.5,
+        0.65
+      ],
+      "retrieval_residual_anchor": "expert",
+      "retrieval_residual_reduce": "none",
+      "policy_rollout_success_rate": 0.3408695652173913,
+      "policy_rollout_progress": 0.5645663425974224,
+      "oracle_success_rate": 0.8695652173913043,
+      "action_mse_to_best": 0.4078142614306315,
+      "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": 0.3326879640447751,
+          "expert_success_rate": 0.8584070796460177,
+          "num_groups": 113,
+          "oracle_success_rate": 0.9380530973451328,
+          "policy_expert_regret": 0.9886335991652666,
+          "policy_oracle_regret": 1.0678357542616077,
+          "policy_rollout_progress": 0.6342759715244833,
+          "policy_rollout_success_rate": 0.23008849557522124,
+          "restore_max_error": 4.76837158203125e-07
+        },
+        "PickCube-v1": {
+          "action_mse_to_best": 0.32456262678960746,
+          "expert_success_rate": 0.9402173913043478,
+          "num_groups": 184,
+          "oracle_success_rate": 0.9456521739130435,
+          "policy_expert_regret": 1.0765762930715699,
+          "policy_oracle_regret": 1.0766002975465,
+          "policy_rollout_progress": 0.5740603387619004,
+          "policy_rollout_success_rate": 0.28804347826086957,
+          "restore_max_error": 4.76837158203125e-07
+        },
+        "PullCube-v1": {
+          "action_mse_to_best": 0.5971522048900002,
+          "expert_success_rate": 0.25,
+          "num_groups": 76,
+          "oracle_success_rate": 0.40789473684210525,
+          "policy_expert_regret": 0.3019960091547354,
+          "policy_oracle_regret": 0.548387196527696,
+          "policy_rollout_progress": 0.27790169827149885,
+          "policy_rollout_success_rate": 0.17105263157894737,
+          "restore_max_error": 4.76837158203125e-07
+        },
+        "PushCube-v1": {
+          "action_mse_to_best": 0.36999872665818745,
+          "expert_success_rate": 0.8198198198198198,
+          "num_groups": 111,
+          "oracle_success_rate": 1.0,
+          "policy_expert_regret": 0.3303051149254447,
+          "policy_oracle_regret": 0.3801627168515781,
+          "policy_rollout_progress": 0.8180354813466201,
+          "policy_rollout_success_rate": 0.8018018018018018,
+          "restore_max_error": 4.76837158203125e-07
+        },
+        "StackCube-v1": {
+          "action_mse_to_best": 0.5574341848963401,
+          "expert_success_rate": 0.6923076923076923,
+          "num_groups": 91,
+          "oracle_success_rate": 0.8571428571428571,
+          "policy_expert_regret": 1.1312875028822449,
+          "policy_oracle_regret": 1.2905735427534186,
+          "policy_rollout_progress": 0.3890427734170641,
+          "policy_rollout_success_rate": 0.16483516483516483,
+          "restore_max_error": 3.948807716369629e-07
+        }
+      }
+    },
+    {
+      "seed": 2,
+      "path": "/scratch/knguy52/dovla/experiments/dovla_h16_policy_ckpt_runs/near_miss_policy_bc5/seed_2/policy_rollout_retrieval_residual_k2grid_broad_safe_ray_margin0p20.json",
+      "num_groups": 575,
+      "selection_mode": "retrieval_residual",
+      "num_candidates": 128,
+      "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": 2,
+      "retrieval_metric": "raw",
+      "retrieval_type_min_success": 0.0,
+      "retrieval_residual_scale": 0.4,
+      "retrieval_residual_scales": [
+        0.2,
+        0.35,
+        0.5,
+        0.65
+      ],
+      "retrieval_residual_anchor": "expert",
+      "retrieval_residual_reduce": "none",
+      "policy_rollout_success_rate": 0.36695652173913046,
+      "policy_rollout_progress": 0.5798133173153031,
+      "oracle_success_rate": 0.8765217391304347,
+      "action_mse_to_best": 0.44122280221675403,
+      "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": 0.35413535257976037,
+          "expert_success_rate": 0.8229166666666666,
+          "num_groups": 96,
+          "oracle_success_rate": 0.9270833333333334,
+          "policy_expert_regret": 0.8470402403424183,
+          "policy_oracle_regret": 0.9533149556567272,
+          "policy_rollout_progress": 0.6469602342694998,
+          "policy_rollout_success_rate": 0.3333333333333333,
+          "restore_max_error": 3.5762786865234375e-07
+        },
+        "PickCube-v1": {
+          "action_mse_to_best": 0.3460169666907703,
+          "expert_success_rate": 0.9444444444444444,
+          "num_groups": 198,
+          "oracle_success_rate": 0.9595959595959596,
+          "policy_expert_regret": 0.9843737633259158,
+          "policy_oracle_regret": 0.9914572138546242,
+          "policy_rollout_progress": 0.6300086113671312,
+          "policy_rollout_success_rate": 0.3333333333333333,
+          "restore_max_error": 4.76837158203125e-07
+        },
+        "PullCube-v1": {
+          "action_mse_to_best": 0.6635176117221514,
+          "expert_success_rate": 0.24444444444444444,
+          "num_groups": 90,
+          "oracle_success_rate": 0.4666666666666667,
+          "policy_expert_regret": 0.2572224053933597,
+          "policy_oracle_regret": 0.5031192180252119,
+          "policy_rollout_progress": 0.3282055651907083,
+          "policy_rollout_success_rate": 0.2222222222222222,
+          "restore_max_error": 4.0978193283081055e-07
+        },
+        "PushCube-v1": {
+          "action_mse_to_best": 0.41323581219899774,
+          "expert_success_rate": 0.8514851485148515,
+          "num_groups": 101,
+          "oracle_success_rate": 1.0,
+          "policy_expert_regret": 0.39989533297496266,
+          "policy_oracle_regret": 0.43690602510872456,
+          "policy_rollout_progress": 0.7908167471685031,
+          "policy_rollout_success_rate": 0.7722772277227723,
+          "restore_max_error": 4.76837158203125e-07
+        },
+        "StackCube-v1": {
+          "action_mse_to_best": 0.5526817326123515,
+          "expert_success_rate": 0.7666666666666667,
+          "num_groups": 90,
+          "oracle_success_rate": 0.9111111111111111,
+          "policy_expert_regret": 1.1537007934517332,
+          "policy_oracle_regret": 1.3027729585766792,
+          "policy_rollout_progress": 0.4125753064950307,
+          "policy_rollout_success_rate": 0.16666666666666666,
+          "restore_max_error": 4.76837158203125e-07
+        }
+      }
+    }
+  ]
+}

results/h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_k2grid_broad_safe_ray_margin0p20_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_k2grid_broad_safe_ray_margin0p20.json`
+Completed seeds: 3
+Baseline h=4 policy success: 29.67%
+Baseline h=16 rank-checkpoint success: 29.74%
+
+Mean success: 34.96% +/- 1.51%
+Gain vs h=16 rank checkpoint: +5.22%
+Mean progress: 56.72%
+Mean action MSE to best: 0.420
+
+| 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 | 128 | no | 2 | raw | expert | none | 0.00 | 0.40 | 0.20,0.35,0.50,0.65 | 0.200 | 0.00 | 0 | 0.00 | 34.09% | 55.73% | 85.74% | 0.410 |
+| 1 | retrieval_residual | 128 | no | 2 | raw | expert | none | 0.00 | 0.40 | 0.20,0.35,0.50,0.65 | 0.200 | 0.00 | 0 | 0.00 | 34.09% | 56.46% | 86.96% | 0.408 |
+| 2 | retrieval_residual | 128 | no | 2 | raw | expert | none | 0.00 | 0.40 | 0.20,0.35,0.50,0.65 | 0.200 | 0.00 | 0 | 0.00 | 36.70% | 57.98% | 87.65% | 0.441 |

results/h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_k2grid_tight_safe_ray_margin0p20_summary.json

@@ -0,0 +1,313 @@
+{
+  "run_root": "/scratch/knguy52/dovla/experiments/dovla_h16_policy_ckpt_runs",
+  "objective": "near_miss_policy_bc5",
+  "out_name": "policy_rollout_retrieval_residual_k2grid_tight_safe_ray_margin0p20.json",
+  "num_completed": 3,
+  "baseline_h4_policy_success": 0.2967,
+  "baseline_h16_rank_checkpoint_success": 0.29739130434782607,
+  "mean_success": 0.34840579710144925,
+  "std_success": 0.013053136520808913,
+  "mean_progress": 0.5661374297490176,
+  "mean_action_mse_to_best": 0.40390177049552617,
+  "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_k2grid_tight_safe_ray_margin0p20.json",
+      "num_groups": 575,
+      "selection_mode": "retrieval_residual",
+      "num_candidates": 96,
+      "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": 2,
+      "retrieval_metric": "raw",
+      "retrieval_type_min_success": 0.0,
+      "retrieval_residual_scale": 0.4,
+      "retrieval_residual_scales": [
+        0.3,
+        0.4,
+        0.5
+      ],
+      "retrieval_residual_anchor": "expert",
+      "retrieval_residual_reduce": "none",
+      "policy_rollout_success_rate": 0.3408695652173913,
+      "policy_rollout_progress": 0.5560484701245213,
+      "oracle_success_rate": 0.8573913043478261,
+      "action_mse_to_best": 0.38957618820359524,
+      "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": 0.31969772514973566,
+          "expert_success_rate": 0.8865979381443299,
+          "num_groups": 97,
+          "oracle_success_rate": 0.9278350515463918,
+          "policy_expert_regret": 1.0792337135248578,
+          "policy_oracle_regret": 1.1056491912333006,
+          "policy_rollout_progress": 0.570563180084081,
+          "policy_rollout_success_rate": 0.2268041237113402,
+          "restore_max_error": 4.76837158203125e-07
+        },
+        "PickCube-v1": {
+          "action_mse_to_best": 0.3040950526841558,
+          "expert_success_rate": 0.9375,
+          "num_groups": 208,
+          "oracle_success_rate": 0.9471153846153846,
+          "policy_expert_regret": 0.9611001174825315,
+          "policy_oracle_regret": 0.974522984909037,
+          "policy_rollout_progress": 0.6161665263442466,
+          "policy_rollout_success_rate": 0.34134615384615385,
+          "restore_max_error": 4.76837158203125e-07
+        },
+        "PullCube-v1": {
+          "action_mse_to_best": 0.6019391601348852,
+          "expert_success_rate": 0.19480519480519481,
+          "num_groups": 77,
+          "oracle_success_rate": 0.36363636363636365,
+          "policy_expert_regret": 0.19750071654713094,
+          "policy_oracle_regret": 0.36829960986355004,
+          "policy_rollout_progress": 0.345251500591418,
+          "policy_rollout_success_rate": 0.24675324675324675,
+          "restore_max_error": 4.76837158203125e-07
+        },
+        "PushCube-v1": {
+          "action_mse_to_best": 0.366488233882379,
+          "expert_success_rate": 0.8617021276595744,
+          "num_groups": 94,
+          "oracle_success_rate": 0.9893617021276596,
+          "policy_expert_regret": 0.5719942746961371,
+          "policy_oracle_regret": 0.6495450359392674,
+          "policy_rollout_progress": 0.6908804959756263,
+          "policy_rollout_success_rate": 0.6595744680851063,
+          "restore_max_error": 4.76837158203125e-07
+        },
+        "StackCube-v1": {
+          "action_mse_to_best": 0.4943903899384719,
+          "expert_success_rate": 0.6767676767676768,
+          "num_groups": 99,
+          "oracle_success_rate": 0.8585858585858586,
+          "policy_expert_regret": 0.9730729231930743,
+          "policy_oracle_regret": 1.147238310539361,
+          "policy_rollout_progress": 0.45144921420800566,
+          "policy_rollout_success_rate": 0.2222222222222222,
+          "restore_max_error": 3.948807716369629e-07
+        }
+      }
+    },
+    {
+      "seed": 1,
+      "path": "/scratch/knguy52/dovla/experiments/dovla_h16_policy_ckpt_runs/near_miss_policy_bc5/seed_1/policy_rollout_retrieval_residual_k2grid_tight_safe_ray_margin0p20.json",
+      "num_groups": 575,
+      "selection_mode": "retrieval_residual",
+      "num_candidates": 96,
+      "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": 2,
+      "retrieval_metric": "raw",
+      "retrieval_type_min_success": 0.0,
+      "retrieval_residual_scale": 0.4,
+      "retrieval_residual_scales": [
+        0.3,
+        0.4,
+        0.5
+      ],
+      "retrieval_residual_anchor": "expert",
+      "retrieval_residual_reduce": "none",
+      "policy_rollout_success_rate": 0.3408695652173913,
+      "policy_rollout_progress": 0.5655716087507165,
+      "oracle_success_rate": 0.8695652173913043,
+      "action_mse_to_best": 0.39600895531475544,
+      "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": 0.3261066819829619,
+          "expert_success_rate": 0.8584070796460177,
+          "num_groups": 113,
+          "oracle_success_rate": 0.9380530973451328,
+          "policy_expert_regret": 0.987694250535121,
+          "policy_oracle_regret": 1.066896405631462,
+          "policy_rollout_progress": 0.6352153201546289,
+          "policy_rollout_success_rate": 0.23008849557522124,
+          "restore_max_error": 4.76837158203125e-07
+        },
+        "PickCube-v1": {
+          "action_mse_to_best": 0.32072854108865495,
+          "expert_success_rate": 0.9402173913043478,
+          "num_groups": 184,
+          "oracle_success_rate": 0.9456521739130435,
+          "policy_expert_regret": 1.0790405539876742,
+          "policy_oracle_regret": 1.0790645584626042,
+          "policy_rollout_progress": 0.5715960778457962,
+          "policy_rollout_success_rate": 0.28804347826086957,
+          "restore_max_error": 4.76837158203125e-07
+        },
+        "PullCube-v1": {
+          "action_mse_to_best": 0.5901441709383538,
+          "expert_success_rate": 0.25,
+          "num_groups": 76,
+          "oracle_success_rate": 0.40789473684210525,
+          "policy_expert_regret": 0.30155902473550095,
+          "policy_oracle_regret": 0.5481829527569445,
+          "policy_rollout_progress": 0.27816353404992505,
+          "policy_rollout_success_rate": 0.17105263157894737,
+          "restore_max_error": 4.76837158203125e-07
+        },
+        "PushCube-v1": {
+          "action_mse_to_best": 0.3623764889801408,
+          "expert_success_rate": 0.8198198198198198,
+          "num_groups": 111,
+          "oracle_success_rate": 1.0,
+          "policy_expert_regret": 0.33009791642695935,
+          "policy_oracle_regret": 0.36357842345495484,
+          "policy_rollout_progress": 0.8256107657342344,
+          "policy_rollout_success_rate": 0.8108108108108109,
+          "restore_max_error": 4.76837158203125e-07
+        },
+        "StackCube-v1": {
+          "action_mse_to_best": 0.5139153342141392,
+          "expert_success_rate": 0.6923076923076923,
+          "num_groups": 91,
+          "oracle_success_rate": 0.8571428571428571,
+          "policy_expert_regret": 1.128705463730372,
+          "policy_oracle_regret": 1.3008532178598446,
+          "policy_rollout_progress": 0.38975210929964926,
+          "policy_rollout_success_rate": 0.15384615384615385,
+          "restore_max_error": 3.948807716369629e-07
+        }
+      }
+    },
+    {
+      "seed": 2,
+      "path": "/scratch/knguy52/dovla/experiments/dovla_h16_policy_ckpt_runs/near_miss_policy_bc5/seed_2/policy_rollout_retrieval_residual_k2grid_tight_safe_ray_margin0p20.json",
+      "num_groups": 575,
+      "selection_mode": "retrieval_residual",
+      "num_candidates": 96,
+      "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": 2,
+      "retrieval_metric": "raw",
+      "retrieval_type_min_success": 0.0,
+      "retrieval_residual_scale": 0.4,
+      "retrieval_residual_scales": [
+        0.3,
+        0.4,
+        0.5
+      ],
+      "retrieval_residual_anchor": "expert",
+      "retrieval_residual_reduce": "none",
+      "policy_rollout_success_rate": 0.3634782608695652,
+      "policy_rollout_progress": 0.5767922103718149,
+      "oracle_success_rate": 0.8765217391304347,
+      "action_mse_to_best": 0.4261201679682278,
+      "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": 0.354290255427865,
+          "expert_success_rate": 0.8229166666666666,
+          "num_groups": 96,
+          "oracle_success_rate": 0.9270833333333334,
+          "policy_expert_regret": 0.8470402403424183,
+          "policy_oracle_regret": 0.9533149556567272,
+          "policy_rollout_progress": 0.6469602342694998,
+          "policy_rollout_success_rate": 0.3333333333333333,
+          "restore_max_error": 3.5762786865234375e-07
+        },
+        "PickCube-v1": {
+          "action_mse_to_best": 0.3393588808336944,
+          "expert_success_rate": 0.9444444444444444,
+          "num_groups": 198,
+          "oracle_success_rate": 0.9595959595959596,
+          "policy_expert_regret": 0.9770182932284895,
+          "policy_oracle_regret": 0.9841017437571979,
+          "policy_rollout_progress": 0.6323135764140523,
+          "policy_rollout_success_rate": 0.3383838383838384,
+          "restore_max_error": 4.76837158203125e-07
+        },
+        "PullCube-v1": {
+          "action_mse_to_best": 0.65280382550425,
+          "expert_success_rate": 0.24444444444444444,
+          "num_groups": 90,
+          "oracle_success_rate": 0.4666666666666667,
+          "policy_expert_regret": 0.2978827285338816,
+          "policy_oracle_regret": 0.5235301449342943,
+          "policy_rollout_progress": 0.31890559706975585,
+          "policy_rollout_success_rate": 0.2111111111111111,
+          "restore_max_error": 4.0978193283081055e-07
+        },
+        "PushCube-v1": {
+          "action_mse_to_best": 0.39388676090995867,
+          "expert_success_rate": 0.8514851485148515,
+          "num_groups": 101,
+          "oracle_success_rate": 1.0,
+          "policy_expert_regret": 0.39989533297496266,
+          "policy_oracle_regret": 0.43690602510872456,
+          "policy_rollout_progress": 0.7908167471685031,
+          "policy_rollout_success_rate": 0.7722772277227723,
+          "restore_max_error": 4.76837158203125e-07
+        },
+        "StackCube-v1": {
+          "action_mse_to_best": 0.5031029612032903,
+          "expert_success_rate": 0.7666666666666667,
+          "num_groups": 90,
+          "oracle_success_rate": 0.9111111111111111,
+          "policy_expert_regret": 1.1780803667174444,
+          "policy_oracle_regret": 1.3400676523645718,
+          "policy_rollout_progress": 0.3975028349293603,
+          "policy_rollout_success_rate": 0.14444444444444443,
+          "restore_max_error": 4.76837158203125e-07
+        }
+      }
+    }
+  ]
+}

results/h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_k2grid_tight_safe_ray_margin0p20_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_k2grid_tight_safe_ray_margin0p20.json`
+Completed seeds: 3
+Baseline h=4 policy success: 29.67%
+Baseline h=16 rank-checkpoint success: 29.74%
+
+Mean success: 34.84% +/- 1.31%
+Gain vs h=16 rank checkpoint: +5.10%
+Mean progress: 56.61%
+Mean action MSE to best: 0.404
+
+| 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 | 96 | no | 2 | raw | expert | none | 0.00 | 0.40 | 0.30,0.40,0.50 | 0.200 | 0.00 | 0 | 0.00 | 34.09% | 55.60% | 85.74% | 0.390 |
+| 1 | retrieval_residual | 96 | no | 2 | raw | expert | none | 0.00 | 0.40 | 0.30,0.40,0.50 | 0.200 | 0.00 | 0 | 0.00 | 34.09% | 56.56% | 86.96% | 0.396 |
+| 2 | retrieval_residual | 96 | no | 2 | raw | expert | none | 0.00 | 0.40 | 0.30,0.40,0.50 | 0.200 | 0.00 | 0 | 0.00 | 36.35% | 57.68% | 87.65% | 0.426 |

results/h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_k4grid_tight_safe_ray_margin0p20_summary.json

@@ -0,0 +1,313 @@
+{
+  "run_root": "/scratch/knguy52/dovla/experiments/dovla_h16_policy_ckpt_runs",
+  "objective": "near_miss_policy_bc5",
+  "out_name": "policy_rollout_retrieval_residual_k4grid_tight_safe_ray_margin0p20.json",
+  "num_completed": 3,
+  "baseline_h4_policy_success": 0.2967,
+  "baseline_h16_rank_checkpoint_success": 0.29739130434782607,
+  "mean_success": 0.34550724637681157,
+  "std_success": 0.013282828101321259,
+  "mean_progress": 0.5659294649053387,
+  "mean_action_mse_to_best": 0.4081513763904787,
+  "gain_vs_h4": 0.04880724637681155,
+  "gain_vs_h16_rank_checkpoint": 0.0481159420289855,
+  "rows": [
+    {
+      "seed": 0,
+      "path": "/scratch/knguy52/dovla/experiments/dovla_h16_policy_ckpt_runs/near_miss_policy_bc5/seed_0/policy_rollout_retrieval_residual_k4grid_tight_safe_ray_margin0p20.json",
+      "num_groups": 575,
+      "selection_mode": "retrieval_residual",
+      "num_candidates": 192,
+      "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": [
+        0.3,
+        0.4,
+        0.5
+      ],
+      "retrieval_residual_anchor": "expert",
+      "retrieval_residual_reduce": "none",
+      "policy_rollout_success_rate": 0.3339130434782609,
+      "policy_rollout_progress": 0.5525535609281581,
+      "oracle_success_rate": 0.8573913043478261,
+      "action_mse_to_best": 0.39564024614980037,
+      "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": 0.32080327893712934,
+          "expert_success_rate": 0.8865979381443299,
+          "num_groups": 97,
+          "oracle_success_rate": 0.9278350515463918,
+          "policy_expert_regret": 1.0792337135248578,
+          "policy_oracle_regret": 1.1056491912333006,
+          "policy_rollout_progress": 0.570563180084081,
+          "policy_rollout_success_rate": 0.2268041237113402,
+          "restore_max_error": 4.76837158203125e-07
+        },
+        "PickCube-v1": {
+          "action_mse_to_best": 0.3056291518374704,
+          "expert_success_rate": 0.9375,
+          "num_groups": 208,
+          "oracle_success_rate": 0.9471153846153846,
+          "policy_expert_regret": 0.9843698578277746,
+          "policy_oracle_regret": 0.9977927252542801,
+          "policy_rollout_progress": 0.6073198629220804,
+          "policy_rollout_success_rate": 0.3269230769230769,
+          "restore_max_error": 4.76837158203125e-07
+        },
+        "PullCube-v1": {
+          "action_mse_to_best": 0.6134872330086572,
+          "expert_success_rate": 0.19480519480519481,
+          "num_groups": 77,
+          "oracle_success_rate": 0.36363636363636365,
+          "policy_expert_regret": 0.19646397312836988,
+          "policy_oracle_regret": 0.3894094792376091,
+          "policy_rollout_progress": 0.33712864420437194,
+          "policy_rollout_success_rate": 0.23376623376623376,
+          "restore_max_error": 4.76837158203125e-07
+        },
+        "PushCube-v1": {
+          "action_mse_to_best": 0.366488233882379,
+          "expert_success_rate": 0.8617021276595744,
+          "num_groups": 94,
+          "oracle_success_rate": 0.9893617021276596,
+          "policy_expert_regret": 0.5719942746961371,
+          "policy_oracle_regret": 0.6495450359392674,
+          "policy_rollout_progress": 0.6908804959756263,
+          "policy_rollout_success_rate": 0.6595744680851063,
+          "restore_max_error": 4.76837158203125e-07
+        },
+        "StackCube-v1": {
+          "action_mse_to_best": 0.5163227168730262,
+          "expert_success_rate": 0.6767676767676768,
+          "num_groups": 99,
+          "oracle_success_rate": 0.8585858585858586,
+          "policy_expert_regret": 0.968466931822324,
+          "policy_oracle_regret": 1.142632319168611,
+          "policy_rollout_progress": 0.45605520557875584,
+          "policy_rollout_success_rate": 0.2222222222222222,
+          "restore_max_error": 3.948807716369629e-07
+        }
+      }
+    },
+    {
+      "seed": 1,
+      "path": "/scratch/knguy52/dovla/experiments/dovla_h16_policy_ckpt_runs/near_miss_policy_bc5/seed_1/policy_rollout_retrieval_residual_k4grid_tight_safe_ray_margin0p20.json",
+      "num_groups": 575,
+      "selection_mode": "retrieval_residual",
+      "num_candidates": 192,
+      "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": [
+        0.3,
+        0.4,
+        0.5
+      ],
+      "retrieval_residual_anchor": "expert",
+      "retrieval_residual_reduce": "none",
+      "policy_rollout_success_rate": 0.3426086956521739,
+      "policy_rollout_progress": 0.5679292374865517,
+      "oracle_success_rate": 0.8695652173913043,
+      "action_mse_to_best": 0.4010225588310024,
+      "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": 0.32670529729859754,
+          "expert_success_rate": 0.8584070796460177,
+          "num_groups": 113,
+          "oracle_success_rate": 0.9380530973451328,
+          "policy_expert_regret": 0.987694250535121,
+          "policy_oracle_regret": 1.066896405631462,
+          "policy_rollout_progress": 0.6352153201546289,
+          "policy_rollout_success_rate": 0.23008849557522124,
+          "restore_max_error": 4.76837158203125e-07
+        },
+        "PickCube-v1": {
+          "action_mse_to_best": 0.3205799974460641,
+          "expert_success_rate": 0.9402173913043478,
+          "num_groups": 184,
+          "oracle_success_rate": 0.9456521739130435,
+          "policy_expert_regret": 1.0633463955474207,
+          "policy_oracle_regret": 1.0708833005147702,
+          "policy_rollout_progress": 0.5743425531849347,
+          "policy_rollout_success_rate": 0.29347826086956524,
+          "restore_max_error": 4.76837158203125e-07
+        },
+        "PullCube-v1": {
+          "action_mse_to_best": 0.5882207980673564,
+          "expert_success_rate": 0.25,
+          "num_groups": 76,
+          "oracle_success_rate": 0.40789473684210525,
+          "policy_expert_regret": 0.3012820670479222,
+          "policy_oracle_regret": 0.5461749000180709,
+          "policy_rollout_progress": 0.28139841311464187,
+          "policy_rollout_success_rate": 0.17105263157894737,
+          "restore_max_error": 4.76837158203125e-07
+        },
+        "PushCube-v1": {
+          "action_mse_to_best": 0.3662383394265497,
+          "expert_success_rate": 0.8198198198198198,
+          "num_groups": 111,
+          "oracle_success_rate": 1.0,
+          "policy_expert_regret": 0.33009791642695935,
+          "policy_oracle_regret": 0.36357842345495484,
+          "policy_rollout_progress": 0.8256107657342344,
+          "policy_rollout_success_rate": 0.8108108108108109,
+          "restore_max_error": 4.76837158203125e-07
+        },
+        "StackCube-v1": {
+          "action_mse_to_best": 0.5420474381708018,
+          "expert_success_rate": 0.6923076923076923,
+          "num_groups": 91,
+          "oracle_success_rate": 0.8571428571428571,
+          "policy_expert_regret": 1.1220633290953688,
+          "policy_oracle_regret": 1.2942110832248415,
+          "policy_rollout_progress": 0.39639424393465233,
+          "policy_rollout_success_rate": 0.15384615384615385,
+          "restore_max_error": 3.948807716369629e-07
+        }
+      }
+    },
+    {
+      "seed": 2,
+      "path": "/scratch/knguy52/dovla/experiments/dovla_h16_policy_ckpt_runs/near_miss_policy_bc5/seed_2/policy_rollout_retrieval_residual_k4grid_tight_safe_ray_margin0p20.json",
+      "num_groups": 575,
+      "selection_mode": "retrieval_residual",
+      "num_candidates": 192,
+      "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": [
+        0.3,
+        0.4,
+        0.5
+      ],
+      "retrieval_residual_anchor": "expert",
+      "retrieval_residual_reduce": "none",
+      "policy_rollout_success_rate": 0.36,
+      "policy_rollout_progress": 0.5773055963013062,
+      "oracle_success_rate": 0.8765217391304347,
+      "action_mse_to_best": 0.4277913241906334,
+      "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": 0.354290255427865,
+          "expert_success_rate": 0.8229166666666666,
+          "num_groups": 96,
+          "oracle_success_rate": 0.9270833333333334,
+          "policy_expert_regret": 0.8470402403424183,
+          "policy_oracle_regret": 0.9533149556567272,
+          "policy_rollout_progress": 0.6469602342694998,
+          "policy_rollout_success_rate": 0.3333333333333333,
+          "restore_max_error": 3.5762786865234375e-07
+        },
+        "PickCube-v1": {
+          "action_mse_to_best": 0.34063189538816613,
+          "expert_success_rate": 0.9444444444444444,
+          "num_groups": 198,
+          "oracle_success_rate": 0.9595959595959596,
+          "policy_expert_regret": 0.9844739756481063,
+          "policy_oracle_regret": 0.9915574261768147,
+          "policy_rollout_progress": 0.6299189090296023,
+          "policy_rollout_success_rate": 0.3333333333333333,
+          "restore_max_error": 4.76837158203125e-07
+        },
+        "PullCube-v1": {
+          "action_mse_to_best": 0.649535529812177,
+          "expert_success_rate": 0.24444444444444444,
+          "num_groups": 90,
+          "oracle_success_rate": 0.4666666666666667,
+          "policy_expert_regret": 0.3138571102654977,
+          "policy_oracle_regret": 0.5394038093862643,
+          "policy_rollout_progress": 0.32220782965974243,
+          "policy_rollout_success_rate": 0.2111111111111111,
+          "restore_max_error": 4.0978193283081055e-07
+        },
+        "PushCube-v1": {
+          "action_mse_to_best": 0.3968160474802008,
+          "expert_success_rate": 0.8514851485148515,
+          "num_groups": 101,
+          "oracle_success_rate": 1.0,
+          "policy_expert_regret": 0.4179702181332182,
+          "policy_oracle_regret": 0.45498091026698007,
+          "policy_rollout_progress": 0.7826428521092575,
+          "policy_rollout_success_rate": 0.7623762376237624,
+          "restore_max_error": 4.76837158203125e-07
+        },
+        "StackCube-v1": {
+          "action_mse_to_best": 0.5109601458120677,
+          "expert_success_rate": 0.7666666666666667,
+          "num_groups": 90,
+          "oracle_success_rate": 0.9111111111111111,
+          "policy_expert_regret": 1.1636614387234052,
+          "policy_oracle_regret": 1.3256487243705326,
+          "policy_rollout_progress": 0.4119217629233996,
+          "policy_rollout_success_rate": 0.14444444444444443,
+          "restore_max_error": 4.76837158203125e-07
+        }
+      }
+    }
+  ]
+}

results/h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_k4grid_tight_safe_ray_margin0p20_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_k4grid_tight_safe_ray_margin0p20.json`
+Completed seeds: 3
+Baseline h=4 policy success: 29.67%
+Baseline h=16 rank-checkpoint success: 29.74%
+
+Mean success: 34.55% +/- 1.33%
+Gain vs h=16 rank checkpoint: +4.81%
+Mean progress: 56.59%
+Mean action MSE to best: 0.408
+
+| 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 | 192 | no | 4 | raw | expert | none | 0.00 | 0.40 | 0.30,0.40,0.50 | 0.200 | 0.00 | 0 | 0.00 | 33.39% | 55.26% | 85.74% | 0.396 |
+| 1 | retrieval_residual | 192 | no | 4 | raw | expert | none | 0.00 | 0.40 | 0.30,0.40,0.50 | 0.200 | 0.00 | 0 | 0.00 | 34.26% | 56.79% | 86.96% | 0.401 |
+| 2 | retrieval_residual | 192 | no | 4 | raw | expert | none | 0.00 | 0.40 | 0.30,0.40,0.50 | 0.200 | 0.00 | 0 | 0.00 | 36.00% | 57.73% | 87.65% | 0.428 |

results/paper_analysis.json

@@ -1,5 +1,5 @@
 {
-  "generated_utc": "2026-06-28T17:49:47+00:00",
+  "generated_utc": "2026-06-28T17:53:26+00:00",
   "mechanism_gap": {
     "best_clean_vs_direct_same_ckpt": 0.06724637681159418,
     "best_clean_vs_h16": 0.05275362318840582,

results/paper_analysis.md

@@ -1,6 +1,6 @@
 # Paper Analysis
 
-Generated: `2026-06-28T17:49:47+00:00`
+Generated: `2026-06-28T17:53:26+00:00`
 
 ## Main Seed Statistics
 

results/paper_core_results.md

@@ -39,6 +39,10 @@ clean-to-same-state proposal gap is `+21.97 pp`.
 | Train-state residual retrieval, safe residuals + advantage margin 0.20 | No | No | 34.84% | +5.10 pp | Abstains unless field advantage beats policy |
 | K2 train-state residual retrieval, safe residuals + advantage margin 0.20 | No | No | 35.01% | +5.28 pp | Current best deployment-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 denoising does not beat raw K2 residuals |
+| 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 fixed K2 residual transport |
+| 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 just below fixed K2 scale-0.40 |
+| K4 train-state residual ray-search, tight scales | No | No | 34.55% | +4.81 pp | Larger neighborhood plus scale-grid dilutes the signal |
 | Policy-relative residual anchor, safe residuals | No | No | 33.74% | +4.00 pp | Policy-relative anchoring ties but does not improve expert-relative residuals |
 | Train-state residual retrieval, z-score metric | No | No | 32.23% | +2.49 pp | State normalization hurts nearest tangent retrieval here |
 | Train-state residual retrieval, z-score metric + anti-goal mask | No | No | 32.75% | +3.01 pp | Masking helps z-score but remains below raw |
@@ -71,13 +75,14 @@ Suggested main-table rows:
 11. Train-state residual retrieval, typed safe families + advantage margin 0.20
 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. Residual-tangent distillation policy
-15. Residual+Gaussian hybrid, K32 sigma0.35
-16. Lattice, near-miss only
-17. Lattice, no expert
-18. Lattice, no expert + policy baseline candidate
-19. Lattice, full
-20. Oracle ceiling
+14. K2 broad tangent ray-search
+15. Residual-tangent distillation policy
+16. Residual+Gaussian hybrid, K32 sigma0.35
+17. Lattice, near-miss only
+18. Lattice, no expert
+19. Lattice, no expert + policy baseline candidate
+20. Lattice, full
+21. Oracle ceiling
 
 Suggested claim:
 
@@ -86,7 +91,7 @@ Suggested claim:
 > abstention 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, and same-state policy-baseline fallback fail to improve the main rows.
+> tangent consensus, tangent ray-search, 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

@@ -28,6 +28,7 @@ when queried on proposal geometry that matches those local counterfactuals.
 | Counterfactual advantage abstention improves clean transport | requiring field advantage over the zero-residual policy raises typed residual transport to 34.84%, and K2 retrieval reaches 35.01% | Current best clean result |
 | Clean residual transport behaves like sparse intervention | `paper_analysis.md` shows K2 residual retrieval abstains to zero-residual policy on 93.3% of states, while selected nonzero no-op/wrong-gripper residuals succeed at ~41.7-41.9% vs 34.5% for abstention | Stronger clean-mechanism framing |
 | Tangent consensus is close but does not beat raw K2 residuals | K4 mean-by-type residual consensus reaches 34.96%, just below the 35.01% K2 raw residual row | Near-tie diagnostic |
+| Tangent ray-search does not beat the fixed clean row | K1/K2 tight scale-grid ray search reach 34.84%; K2 broad reaches 34.96%; K4 tight reaches 34.55%, all below fixed K2 scale-0.40 at 35.01% | Near-tie/negative diagnostic |
 | The proposal gap is now quantified | `paper_analysis.md` reports best clean +5.28 pp over canonical h16, same-state no-expert +27.25 pp, leaving a +21.97 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 |
@@ -55,15 +56,18 @@ clean proposal result, the intended main rows are:
 12. Train-state residual retrieval, typed safe families + advantage margin: 34.84%
 13. K2 train-state residual retrieval, typed safe families + advantage margin: 35.01%
 14. K4 mean-by-type tangent consensus: 34.96%
-15. Residual-tangent distillation policy: 28.87%
-16. Z-score residual retrieval: 32.23-32.81%
-17. Train-family reliability prior: 33.28-33.33%
-18. Residual+Gaussian hybrid K32/K64: 31.30% / 30.90%
-19. Lattice, near-miss only: 55.94%
-20. Lattice, no expert: 56.99%
-21. Lattice, no expert + policy baseline candidate: 40.70%
-22. Lattice, full: 69.33%
-23. Oracle ceiling: 86.78%
+15. K2 broad tangent ray-search: 34.96%
+16. K1/K2 tight tangent ray-search: 34.84% / 34.84%
+17. K4 tight tangent ray-search: 34.55%
+18. Residual-tangent distillation policy: 28.87%
+19. Z-score residual retrieval: 32.23-32.81%
+20. Train-family reliability prior: 33.28-33.33%
+21. Residual+Gaussian hybrid K32/K64: 31.30% / 30.90%
+22. Lattice, near-miss only: 55.94%
+23. Lattice, no expert: 56.99%
+24. Lattice, no expert + policy baseline candidate: 40.70%
+25. Lattice, full: 69.33%
+26. Oracle ceiling: 86.78%
 
 ## Novelty Framing
 
@@ -91,9 +95,8 @@ test-time search. The cleaner novelty is:
 
 ## Job Status
 
-Last checked: `2026-06-28 17:40 UTC`. K1 tight ray-search job
-`14868993_[0-2]` has started running; the K2/K4 ray-search jobs are still queued,
-pending on GPU priority.
+Last checked: `2026-06-28 17:54 UTC`. The counterfactual tangent ray-search batch
+completed and table/analysis outputs were rebuilt.
 
 - `14858328`-`14858333`: completed train-split `field_selected_noexpert_bc5`;
   direct rollout is 26.84%, field-guided best is 27.65%.
@@ -144,13 +147,11 @@ pending on GPU priority.
   and `0.45` reach 34.72% and 34.84%; K4 median and K8 mean at scale `0.40`
   both reach 34.67%. The follow-up confirms K2 raw residual transport remains
   the best clean row.
-- `14868993_[0-2]`: running the K1 tight counterfactual tangent ray-search
-  rollout. Jobs `14868995`/`14868997`/`14868999` are still pending K2/K4 safe
-  residual ray-search rollouts with scale grids. Summary jobs `14868994`/
-  `14868996`/`14868998`/`14869000` and table rebuild `14869605` are
-  dependency-gated on those rollouts. Updated rebuild job `14869860` is also
-  dependency-gated and will regenerate both paper table and paired analysis
-  outputs with the current script.
+- `14868993`/`14868995`/`14868997`/`14868999`: completed counterfactual tangent
+  ray-search rollouts. Results are 34.84% for K1 tight, 34.84% for K2 tight,
+  34.96% for K2 broad, and 34.55% for K4 tight. Summary jobs `14868994`/
+  `14868996`/`14868998`/`14869000` and rebuild job `14869860` completed; the
+  paper table and paired analysis outputs now include these rows.
 - `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
@@ -171,7 +172,8 @@ pending on GPU priority.
 - Promote same-state no-expert lattice (56.99%) as the conservative mechanism
   result.
 - Use K2 typed safe residual transport with advantage abstention (35.01%) only as the current best clean
-  deployment diagnostic, not as a SOTA claim.
+  deployment diagnostic, not as a SOTA claim. The completed ray-search rows are
+  near-ties but do not replace it.
 - Use `results/paper_analysis.md` for paired seed deltas, per-task gaps, and
   selection histograms when writing reviewer-facing tables.
 - Treat z-score retrieval, repaired train-family reliability priors, Gaussian hybrids,

results/paper_table_status.json

@@ -584,14 +584,14 @@
       "story_role": "counterfactual tangent ray-search diagnostic",
       "fallback_success": null,
       "pending_job": "14868999/14869000",
-      "path_exists": false,
-      "status": "pending",
-      "success": null,
-      "std_success": null,
+      "path_exists": true,
+      "status": "complete",
+      "success": 0.34550724637681157,
+      "std_success": 0.013282828101321259,
       "completed_seeds": null,
-      "num_completed": null,
+      "num_completed": 3,
       "best_config": null,
-      "gain_vs_h16_policy": null
+      "gain_vs_h16_policy": 0.0481159420289855
     },
     {
       "key": "retrieval_residual_k4_scale040_safe_margin020_mean_by_type",

results/paper_table_status.md

@@ -33,7 +33,7 @@ Baseline h=16 policy: 29.74%
 | retrieval_residual_k1grid_tight_safe_ray_margin020 | K1 train-state residual ray search, safe residuals, scales 0.30/0.40/0.50, advantage margin 0.20 | complete | 34.84% | +5.10 pp | yes | no | no | counterfactual tangent ray-search diagnostic |
 | retrieval_residual_k2grid_tight_safe_ray_margin020 | K2 train-state residual ray search, safe residuals, scales 0.30/0.40/0.50, advantage margin 0.20 | complete | 34.84% | +5.10 pp | yes | no | no | counterfactual tangent ray-search diagnostic |
 | retrieval_residual_k2grid_broad_safe_ray_margin020 | K2 train-state residual ray search, safe residuals, scales 0.20/0.35/0.50/0.65, advantage margin 0.20 | complete | 34.96% | +5.22 pp | yes | no | no | counterfactual tangent ray-search diagnostic |
-| retrieval_residual_k4grid_tight_safe_ray_margin020 | K4 train-state residual ray search, safe residuals, scales 0.30/0.40/0.50, advantage margin 0.20 | pending 14868999/14869000 | pending | pending | yes | no | no | counterfactual tangent ray-search diagnostic |
+| retrieval_residual_k4grid_tight_safe_ray_margin020 | K4 train-state residual ray search, safe residuals, scales 0.30/0.40/0.50, advantage margin 0.20 | complete | 34.55% | +4.81 pp | yes | no | no | counterfactual tangent ray-search diagnostic |
 | retrieval_residual_k4_scale040_safe_margin020_mean_by_type | K4 train-state residual retrieval, scale 0.40, safe residuals, mean-by-type tangent consensus | complete | 34.96% | +5.22 pp | yes | no | no | counterfactual tangent consensus near-tie ablation |
 | retrieval_residual_policy_anchor_scale035_safe | Policy-relative train-state residual retrieval, scale 0.35, safe non-expert residuals | complete | 33.74% | +4.00 pp | yes | no | no | policy-relative tangent anchor diagnostic |
 | retrieval_residual_scale030_safe_types | Train-state residual retrieval, scale 0.30, policy/no-op/wrong-gripper residuals | complete | 33.51% | +3.77 pp | yes | no | no | typed tangent scale zoom sweep |