Auto-sync: 2026-06-28 13:02:56 (part 3)

results/h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_k4s040_median_safe_margin0p20_summary.json

@@ -0,0 +1,298 @@
+{
+  "run_root": "/scratch/knguy52/dovla/experiments/dovla_h16_policy_ckpt_runs",
+  "objective": "near_miss_policy_bc5",
+  "out_name": "policy_rollout_retrieval_residual_k4s040_median_safe_margin0p20.json",
+  "num_completed": 3,
+  "baseline_h4_policy_success": 0.2967,
+  "baseline_h16_rank_checkpoint_success": 0.29739130434782607,
+  "mean_success": 0.3466666666666667,
+  "std_success": 0.01907766106887459,
+  "mean_progress": 0.5649579251622829,
+  "mean_action_mse_to_best": 0.39526754156834837,
+  "gain_vs_h4": 0.04996666666666666,
+  "gain_vs_h16_rank_checkpoint": 0.04927536231884061,
+  "rows": [
+    {
+      "seed": 0,
+      "path": "/scratch/knguy52/dovla/experiments/dovla_h16_policy_ckpt_runs/near_miss_policy_bc5/seed_0/policy_rollout_retrieval_residual_k4s040_median_safe_margin0p20.json",
+      "num_groups": 575,
+      "selection_mode": "retrieval_residual",
+      "num_candidates": 6,
+      "candidate_sigma": 0.0,
+      "selection_margin": 0.2,
+      "prepend_policy_candidate": false,
+      "field_optim_steps": 0,
+      "field_optim_step_size": 0.0,
+      "field_optim_trust_radius": 0.0,
+      "field_optim_l2_penalty": 0.0,
+      "retrieval_neighbors": 4,
+      "retrieval_metric": "raw",
+      "retrieval_type_min_success": 0.0,
+      "retrieval_residual_scale": 0.4,
+      "retrieval_residual_anchor": "expert",
+      "retrieval_residual_reduce": "median_by_type",
+      "policy_rollout_success_rate": 0.3356521739130435,
+      "policy_rollout_progress": 0.5488967806696082,
+      "oracle_success_rate": 0.8573913043478261,
+      "action_mse_to_best": 0.38073945276034266,
+      "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": 0.3026365274921633,
+          "expert_success_rate": 0.9375,
+          "num_groups": 208,
+          "oracle_success_rate": 0.9471153846153846,
+          "policy_expert_regret": 0.9680189866787539,
+          "policy_oracle_regret": 0.9814418541052594,
+          "policy_rollout_progress": 0.6140553494557165,
+          "policy_rollout_success_rate": 0.33653846153846156,
+          "restore_max_error": 4.76837158203125e-07
+        },
+        "PullCube-v1": {
+          "action_mse_to_best": 0.6018173015349871,
+          "expert_success_rate": 0.19480519480519481,
+          "num_groups": 77,
+          "oracle_success_rate": 0.36363636363636365,
+          "policy_expert_regret": 0.22048042928981412,
+          "policy_oracle_regret": 0.4152111829494356,
+          "policy_rollout_progress": 0.32426945280658637,
+          "policy_rollout_success_rate": 0.22077922077922077,
+          "restore_max_error": 4.76837158203125e-07
+        },
+        "PushCube-v1": {
+          "action_mse_to_best": 0.3646613056276073,
+          "expert_success_rate": 0.8617021276595744,
+          "num_groups": 94,
+          "oracle_success_rate": 0.9893617021276596,
+          "policy_expert_regret": 0.5719942746961371,
+          "policy_oracle_regret": 0.6683580540913217,
+          "policy_rollout_progress": 0.6827057756959124,
+          "policy_rollout_success_rate": 0.648936170212766,
+          "restore_max_error": 4.76837158203125e-07
+        },
+        "StackCube-v1": {
+          "action_mse_to_best": 0.44795973062741035,
+          "expert_success_rate": 0.6767676767676768,
+          "num_groups": 99,
+          "oracle_success_rate": 0.8585858585858586,
+          "policy_expert_regret": 0.9758518865012159,
+          "policy_oracle_regret": 1.1501580606205295,
+          "policy_rollout_progress": 0.43842845402582725,
+          "policy_rollout_success_rate": 0.23232323232323232,
+          "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_k4s040_median_safe_margin0p20.json",
+      "num_groups": 575,
+      "selection_mode": "retrieval_residual",
+      "num_candidates": 6,
+      "candidate_sigma": 0.0,
+      "selection_margin": 0.2,
+      "prepend_policy_candidate": false,
+      "field_optim_steps": 0,
+      "field_optim_step_size": 0.0,
+      "field_optim_trust_radius": 0.0,
+      "field_optim_l2_penalty": 0.0,
+      "retrieval_neighbors": 4,
+      "retrieval_metric": "raw",
+      "retrieval_type_min_success": 0.0,
+      "retrieval_residual_scale": 0.4,
+      "retrieval_residual_anchor": "expert",
+      "retrieval_residual_reduce": "median_by_type",
+      "policy_rollout_success_rate": 0.3356521739130435,
+      "policy_rollout_progress": 0.5610073493064746,
+      "oracle_success_rate": 0.8695652173913043,
+      "action_mse_to_best": 0.3881998635018649,
+      "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": 0.32557108186482064,
+          "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": 0.31965635613659804,
+          "expert_success_rate": 0.9402173913043478,
+          "num_groups": 184,
+          "oracle_success_rate": 0.9456521739130435,
+          "policy_expert_regret": 1.0787787894900565,
+          "policy_oracle_regret": 1.0788027939649865,
+          "policy_rollout_progress": 0.571857842343414,
+          "policy_rollout_success_rate": 0.28804347826086957,
+          "restore_max_error": 4.76837158203125e-07
+        },
+        "PullCube-v1": {
+          "action_mse_to_best": 0.5726699156588629,
+          "expert_success_rate": 0.25,
+          "num_groups": 76,
+          "oracle_success_rate": 0.40789473684210525,
+          "policy_expert_regret": 0.3254129353322481,
+          "policy_oracle_regret": 0.5486240120800702,
+          "policy_rollout_progress": 0.2771305680569065,
+          "policy_rollout_success_rate": 0.17105263157894737,
+          "restore_max_error": 4.76837158203125e-07
+        },
+        "PushCube-v1": {
+          "action_mse_to_best": 0.3578395257192152,
+          "expert_success_rate": 0.8198198198198198,
+          "num_groups": 111,
+          "oracle_success_rate": 1.0,
+          "policy_expert_regret": 0.34759203673483013,
+          "policy_oracle_regret": 0.41432771757916287,
+          "policy_rollout_progress": 0.8018884986370534,
+          "policy_rollout_success_rate": 0.7837837837837838,
+          "restore_max_error": 4.76837158203125e-07
+        },
+        "StackCube-v1": {
+          "action_mse_to_best": 0.48753317350337466,
+          "expert_success_rate": 0.6923076923076923,
+          "num_groups": 91,
+          "oracle_success_rate": 0.8571428571428571,
+          "policy_expert_regret": 1.1282761976286606,
+          "policy_oracle_regret": 1.300423951758133,
+          "policy_rollout_progress": 0.39018137540136066,
+          "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_k4s040_median_safe_margin0p20.json",
+      "num_groups": 575,
+      "selection_mode": "retrieval_residual",
+      "num_candidates": 6,
+      "candidate_sigma": 0.0,
+      "selection_margin": 0.2,
+      "prepend_policy_candidate": false,
+      "field_optim_steps": 0,
+      "field_optim_step_size": 0.0,
+      "field_optim_trust_radius": 0.0,
+      "field_optim_l2_penalty": 0.0,
+      "retrieval_neighbors": 4,
+      "retrieval_metric": "raw",
+      "retrieval_type_min_success": 0.0,
+      "retrieval_residual_scale": 0.4,
+      "retrieval_residual_anchor": "expert",
+      "retrieval_residual_reduce": "median_by_type",
+      "policy_rollout_success_rate": 0.36869565217391304,
+      "policy_rollout_progress": 0.5849696455107661,
+      "oracle_success_rate": 0.8765217391304347,
+      "action_mse_to_best": 0.41686330844283753,
+      "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": 0.35425220065614366,
+          "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": 0.33620506880635564,
+          "expert_success_rate": 0.9444444444444444,
+          "num_groups": 198,
+          "oracle_success_rate": 0.9595959595959596,
+          "policy_expert_regret": 0.9879054814163182,
+          "policy_oracle_regret": 0.9949889319450266,
+          "policy_rollout_progress": 0.6336337373676625,
+          "policy_rollout_success_rate": 0.3333333333333333,
+          "restore_max_error": 4.76837158203125e-07
+        },
+        "PullCube-v1": {
+          "action_mse_to_best": 0.6281907881299654,
+          "expert_success_rate": 0.24444444444444444,
+          "num_groups": 90,
+          "oracle_success_rate": 0.4666666666666667,
+          "policy_expert_regret": 0.2925979837637189,
+          "policy_oracle_regret": 0.5006783538463424,
+          "policy_rollout_progress": 0.33872392334182627,
+          "policy_rollout_success_rate": 0.23333333333333334,
+          "restore_max_error": 4.0978193283081055e-07
+        },
+        "PushCube-v1": {
+          "action_mse_to_best": 0.38383023912953856,
+          "expert_success_rate": 0.8514851485148515,
+          "num_groups": 101,
+          "oracle_success_rate": 1.0,
+          "policy_expert_regret": 0.3618683649761842,
+          "policy_oracle_regret": 0.41677160133229624,
+          "policy_rollout_progress": 0.8010501808459216,
+          "policy_rollout_success_rate": 0.7821782178217822,
+          "restore_max_error": 4.76837158203125e-07
+        },
+        "StackCube-v1": {
+          "action_mse_to_best": 0.48683958204670086,
+          "expert_success_rate": 0.7666666666666667,
+          "num_groups": 90,
+          "oracle_success_rate": 0.9111111111111111,
+          "policy_expert_regret": 1.1489313981599278,
+          "policy_oracle_regret": 1.3109186838070552,
+          "policy_rollout_progress": 0.4155406923757659,
+          "policy_rollout_success_rate": 0.15555555555555556,
+          "restore_max_error": 4.76837158203125e-07
+        }
+      }
+    }
+  ]
+}

results/h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_k4s040_median_safe_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_k4s040_median_safe_margin0p20.json`
+Completed seeds: 3
+Baseline h=4 policy success: 29.67%
+Baseline h=16 rank-checkpoint success: 29.74%
+
+Mean success: 34.67% +/- 1.91%
+Gain vs h=16 rank checkpoint: +4.93%
+Mean progress: 56.50%
+Mean action MSE to best: 0.395
+
+| seed | mode | k | policy cand | retrieval K | retrieval metric | residual anchor | residual reduce | min type success | residual scale | margin | sigma | opt steps | trust | success | progress | oracle | action MSE |
+|---:|---|---:|---|---:|---|---|---|---:|---:|---:|---:|---:|---:|---:|---:|---:|---:|
+| 0 | retrieval_residual | 6 | no | 4 | raw | expert | median_by_type | 0.00 | 0.40 | 0.200 | 0.00 | 0 | 0.00 | 33.57% | 54.89% | 85.74% | 0.381 |
+| 1 | retrieval_residual | 6 | no | 4 | raw | expert | median_by_type | 0.00 | 0.40 | 0.200 | 0.00 | 0 | 0.00 | 33.57% | 56.10% | 86.96% | 0.388 |
+| 2 | retrieval_residual | 6 | no | 4 | raw | expert | median_by_type | 0.00 | 0.40 | 0.200 | 0.00 | 0 | 0.00 | 36.87% | 58.50% | 87.65% | 0.417 |

results/h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_k4s0425_mean_safe_margin0p20_summary.json

@@ -0,0 +1,298 @@
+{
+  "run_root": "/scratch/knguy52/dovla/experiments/dovla_h16_policy_ckpt_runs",
+  "objective": "near_miss_policy_bc5",
+  "out_name": "policy_rollout_retrieval_residual_k4s0425_mean_safe_margin0p20.json",
+  "num_completed": 3,
+  "baseline_h4_policy_success": 0.2967,
+  "baseline_h16_rank_checkpoint_success": 0.29739130434782607,
+  "mean_success": 0.3472463768115942,
+  "std_success": 0.017069486219519373,
+  "mean_progress": 0.5651525005180357,
+  "mean_action_mse_to_best": 0.3958947908273642,
+  "gain_vs_h4": 0.050546376811594185,
+  "gain_vs_h16_rank_checkpoint": 0.049855072463768135,
+  "rows": [
+    {
+      "seed": 0,
+      "path": "/scratch/knguy52/dovla/experiments/dovla_h16_policy_ckpt_runs/near_miss_policy_bc5/seed_0/policy_rollout_retrieval_residual_k4s0425_mean_safe_margin0p20.json",
+      "num_groups": 575,
+      "selection_mode": "retrieval_residual",
+      "num_candidates": 6,
+      "candidate_sigma": 0.0,
+      "selection_margin": 0.2,
+      "prepend_policy_candidate": false,
+      "field_optim_steps": 0,
+      "field_optim_step_size": 0.0,
+      "field_optim_trust_radius": 0.0,
+      "field_optim_l2_penalty": 0.0,
+      "retrieval_neighbors": 4,
+      "retrieval_metric": "raw",
+      "retrieval_type_min_success": 0.0,
+      "retrieval_residual_scale": 0.425,
+      "retrieval_residual_anchor": "expert",
+      "retrieval_residual_reduce": "mean_by_type",
+      "policy_rollout_success_rate": 0.3373913043478261,
+      "policy_rollout_progress": 0.5503261472296942,
+      "oracle_success_rate": 0.8573913043478261,
+      "action_mse_to_best": 0.38102783091609244,
+      "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": 0.30265720665025025,
+          "expert_success_rate": 0.9375,
+          "num_groups": 208,
+          "oracle_success_rate": 0.9471153846153846,
+          "policy_expert_regret": 0.9611825004148369,
+          "policy_oracle_regret": 0.9746053678413423,
+          "policy_rollout_progress": 0.6160841434119413,
+          "policy_rollout_success_rate": 0.34134615384615385,
+          "restore_max_error": 4.76837158203125e-07
+        },
+        "PullCube-v1": {
+          "action_mse_to_best": 0.6002108526307267,
+          "expert_success_rate": 0.19480519480519481,
+          "num_groups": 77,
+          "oracle_success_rate": 0.36363636363636365,
+          "policy_expert_regret": 0.22002532334647493,
+          "policy_oracle_regret": 0.4150555553186878,
+          "policy_rollout_progress": 0.32442508043733415,
+          "policy_rollout_success_rate": 0.22077922077922077,
+          "restore_max_error": 4.76837158203125e-07
+        },
+        "PushCube-v1": {
+          "action_mse_to_best": 0.36597024978316844,
+          "expert_success_rate": 0.8617021276595744,
+          "num_groups": 94,
+          "oracle_success_rate": 0.9893617021276596,
+          "policy_expert_regret": 0.5719942746961371,
+          "policy_oracle_regret": 0.668446716317471,
+          "policy_rollout_progress": 0.6826171134697631,
+          "policy_rollout_success_rate": 0.648936170212766,
+          "restore_max_error": 4.76837158203125e-07
+        },
+        "StackCube-v1": {
+          "action_mse_to_best": 0.4495978315332622,
+          "expert_success_rate": 0.6767676767676768,
+          "num_groups": 99,
+          "oracle_success_rate": 0.8585858585858586,
+          "policy_expert_regret": 0.9719959655494401,
+          "policy_oracle_regret": 1.1461555601069422,
+          "policy_rollout_progress": 0.4424309545394146,
+          "policy_rollout_success_rate": 0.23232323232323232,
+          "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_k4s0425_mean_safe_margin0p20.json",
+      "num_groups": 575,
+      "selection_mode": "retrieval_residual",
+      "num_candidates": 6,
+      "candidate_sigma": 0.0,
+      "selection_margin": 0.2,
+      "prepend_policy_candidate": false,
+      "field_optim_steps": 0,
+      "field_optim_step_size": 0.0,
+      "field_optim_trust_radius": 0.0,
+      "field_optim_l2_penalty": 0.0,
+      "retrieval_neighbors": 4,
+      "retrieval_metric": "raw",
+      "retrieval_type_min_success": 0.0,
+      "retrieval_residual_scale": 0.425,
+      "retrieval_residual_anchor": "expert",
+      "retrieval_residual_reduce": "mean_by_type",
+      "policy_rollout_success_rate": 0.3373913043478261,
+      "policy_rollout_progress": 0.5627563353595526,
+      "oracle_success_rate": 0.8695652173913043,
+      "action_mse_to_best": 0.3898331182320481,
+      "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": 0.32557108186482064,
+          "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": 0.3196794526125102,
+          "expert_success_rate": 0.9402173913043478,
+          "num_groups": 184,
+          "oracle_success_rate": 0.9456521739130435,
+          "policy_expert_regret": 1.0788223734429188,
+          "policy_oracle_regret": 1.0788463779178488,
+          "policy_rollout_progress": 0.5718142583905517,
+          "policy_rollout_success_rate": 0.28804347826086957,
+          "restore_max_error": 4.76837158203125e-07
+        },
+        "PullCube-v1": {
+          "action_mse_to_best": 0.5720450864418557,
+          "expert_success_rate": 0.25,
+          "num_groups": 76,
+          "oracle_success_rate": 0.40789473684210525,
+          "policy_expert_regret": 0.32437335417949054,
+          "policy_oracle_regret": 0.5472229318448195,
+          "policy_rollout_progress": 0.2785316482921572,
+          "policy_rollout_success_rate": 0.17105263157894737,
+          "restore_max_error": 4.76837158203125e-07
+        },
+        "PushCube-v1": {
+          "action_mse_to_best": 0.3580907076127357,
+          "expert_success_rate": 0.8198198198198198,
+          "num_groups": 111,
+          "oracle_success_rate": 1.0,
+          "policy_expert_regret": 0.34753666885264284,
+          "policy_oracle_regret": 0.39731630092268594,
+          "policy_rollout_progress": 0.8098909062845213,
+          "policy_rollout_success_rate": 0.7927927927927928,
+          "restore_max_error": 4.76837158203125e-07
+        },
+        "StackCube-v1": {
+          "action_mse_to_best": 0.498021937773696,
+          "expert_success_rate": 0.6923076923076923,
+          "num_groups": 91,
+          "oracle_success_rate": 0.8571428571428571,
+          "policy_expert_regret": 1.1280680978363686,
+          "policy_oracle_regret": 1.3002158519658413,
+          "policy_rollout_progress": 0.39038947519365247,
+          "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_k4s0425_mean_safe_margin0p20.json",
+      "num_groups": 575,
+      "selection_mode": "retrieval_residual",
+      "num_candidates": 6,
+      "candidate_sigma": 0.0,
+      "selection_margin": 0.2,
+      "prepend_policy_candidate": false,
+      "field_optim_steps": 0,
+      "field_optim_step_size": 0.0,
+      "field_optim_trust_radius": 0.0,
+      "field_optim_l2_penalty": 0.0,
+      "retrieval_neighbors": 4,
+      "retrieval_metric": "raw",
+      "retrieval_type_min_success": 0.0,
+      "retrieval_residual_scale": 0.425,
+      "retrieval_residual_anchor": "expert",
+      "retrieval_residual_reduce": "mean_by_type",
+      "policy_rollout_success_rate": 0.36695652173913046,
+      "policy_rollout_progress": 0.5823750189648601,
+      "oracle_success_rate": 0.8765217391304347,
+      "action_mse_to_best": 0.41682342333395195,
+      "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": 0.3543375191172042,
+          "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": 0.33537142891248667,
+          "expert_success_rate": 0.9444444444444444,
+          "num_groups": 198,
+          "oracle_success_rate": 0.9595959595959596,
+          "policy_expert_regret": 0.9878710091603223,
+          "policy_oracle_regret": 0.9949544596890307,
+          "policy_rollout_progress": 0.6336682096236583,
+          "policy_rollout_success_rate": 0.3333333333333333,
+          "restore_max_error": 4.76837158203125e-07
+        },
+        "PullCube-v1": {
+          "action_mse_to_best": 0.6285684901807044,
+          "expert_success_rate": 0.24444444444444444,
+          "num_groups": 90,
+          "oracle_success_rate": 0.4666666666666667,
+          "policy_expert_regret": 0.2926993225873553,
+          "policy_oracle_regret": 0.5009603909389221,
+          "policy_rollout_progress": 0.33844188624924654,
+          "policy_rollout_success_rate": 0.23333333333333334,
+          "restore_max_error": 4.0978193283081055e-07
+        },
+        "PushCube-v1": {
+          "action_mse_to_best": 0.38641081691378415,
+          "expert_success_rate": 0.8514851485148515,
+          "num_groups": 101,
+          "oracle_success_rate": 1.0,
+          "policy_expert_regret": 0.38059869836462606,
+          "policy_oracle_regret": 0.4532478929746269,
+          "policy_rollout_progress": 0.7843758694016107,
+          "policy_rollout_success_rate": 0.7623762376237624,
+          "restore_max_error": 4.76837158203125e-07
+        },
+        "StackCube-v1": {
+          "action_mse_to_best": 0.48505407813936474,
+          "expert_success_rate": 0.7666666666666667,
+          "num_groups": 90,
+          "oracle_success_rate": 0.9111111111111111,
+          "policy_expert_regret": 1.1483937073085042,
+          "policy_oracle_regret": 1.2974658724334505,
+          "policy_rollout_progress": 0.41788239263825944,
+          "policy_rollout_success_rate": 0.16666666666666666,
+          "restore_max_error": 4.76837158203125e-07
+        }
+      }
+    }
+  ]
+}

results/h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_k4s0425_mean_safe_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_k4s0425_mean_safe_margin0p20.json`
+Completed seeds: 3
+Baseline h=4 policy success: 29.67%
+Baseline h=16 rank-checkpoint success: 29.74%
+
+Mean success: 34.72% +/- 1.71%
+Gain vs h=16 rank checkpoint: +4.99%
+Mean progress: 56.52%
+Mean action MSE to best: 0.396
+
+| seed | mode | k | policy cand | retrieval K | retrieval metric | residual anchor | residual reduce | min type success | residual scale | margin | sigma | opt steps | trust | success | progress | oracle | action MSE |
+|---:|---|---:|---|---:|---|---|---|---:|---:|---:|---:|---:|---:|---:|---:|---:|---:|
+| 0 | retrieval_residual | 6 | no | 4 | raw | expert | mean_by_type | 0.00 | 0.42 | 0.200 | 0.00 | 0 | 0.00 | 33.74% | 55.03% | 85.74% | 0.381 |
+| 1 | retrieval_residual | 6 | no | 4 | raw | expert | mean_by_type | 0.00 | 0.42 | 0.200 | 0.00 | 0 | 0.00 | 33.74% | 56.28% | 86.96% | 0.390 |
+| 2 | retrieval_residual | 6 | no | 4 | raw | expert | mean_by_type | 0.00 | 0.42 | 0.200 | 0.00 | 0 | 0.00 | 36.70% | 58.24% | 87.65% | 0.417 |

results/h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_k4s045_mean_safe_margin0p20_summary.json

@@ -0,0 +1,298 @@
+{
+  "run_root": "/scratch/knguy52/dovla/experiments/dovla_h16_policy_ckpt_runs",
+  "objective": "near_miss_policy_bc5",
+  "out_name": "policy_rollout_retrieval_residual_k4s045_mean_safe_margin0p20.json",
+  "num_completed": 3,
+  "baseline_h4_policy_success": 0.2967,
+  "baseline_h16_rank_checkpoint_success": 0.29739130434782607,
+  "mean_success": 0.34840579710144925,
+  "std_success": 0.017593032933905524,
+  "mean_progress": 0.5664884964474323,
+  "mean_action_mse_to_best": 0.39641575963168907,
+  "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_k4s045_mean_safe_margin0p20.json",
+      "num_groups": 575,
+      "selection_mode": "retrieval_residual",
+      "num_candidates": 6,
+      "candidate_sigma": 0.0,
+      "selection_margin": 0.2,
+      "prepend_policy_candidate": false,
+      "field_optim_steps": 0,
+      "field_optim_step_size": 0.0,
+      "field_optim_trust_radius": 0.0,
+      "field_optim_l2_penalty": 0.0,
+      "retrieval_neighbors": 4,
+      "retrieval_metric": "raw",
+      "retrieval_type_min_success": 0.0,
+      "retrieval_residual_scale": 0.45,
+      "retrieval_residual_anchor": "expert",
+      "retrieval_residual_reduce": "mean_by_type",
+      "policy_rollout_success_rate": 0.3373913043478261,
+      "policy_rollout_progress": 0.5511278174764922,
+      "oracle_success_rate": 0.8573913043478261,
+      "action_mse_to_best": 0.3815302481591378,
+      "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": 0.3026830694781473,
+          "expert_success_rate": 0.9375,
+          "num_groups": 208,
+          "oracle_success_rate": 0.9471153846153846,
+          "policy_expert_regret": 0.9611414949982785,
+          "policy_oracle_regret": 0.974564362424784,
+          "policy_rollout_progress": 0.6161251488284996,
+          "policy_rollout_success_rate": 0.34134615384615385,
+          "restore_max_error": 4.76837158203125e-07
+        },
+        "PullCube-v1": {
+          "action_mse_to_best": 0.6002162871035662,
+          "expert_success_rate": 0.19480519480519481,
+          "num_groups": 77,
+          "oracle_success_rate": 0.36363636363636365,
+          "policy_expert_regret": 0.2200784192317734,
+          "policy_oracle_regret": 0.41524309277014393,
+          "policy_rollout_progress": 0.32423754298587804,
+          "policy_rollout_success_rate": 0.22077922077922077,
+          "restore_max_error": 4.76837158203125e-07
+        },
+        "PushCube-v1": {
+          "action_mse_to_best": 0.3662071306892532,
+          "expert_success_rate": 0.8617021276595744,
+          "num_groups": 94,
+          "oracle_success_rate": 0.9893617021276596,
+          "policy_expert_regret": 0.5719942746961371,
+          "policy_oracle_regret": 0.6683686906670002,
+          "policy_rollout_progress": 0.6826951391202338,
+          "policy_rollout_success_rate": 0.648936170212766,
+          "restore_max_error": 4.76837158203125e-07
+        },
+        "StackCube-v1": {
+          "action_mse_to_best": 0.45223242938142233,
+          "expert_success_rate": 0.6767676767676768,
+          "num_groups": 99,
+          "oracle_success_rate": 0.8585858585858586,
+          "policy_expert_regret": 0.9673541752677975,
+          "policy_oracle_regret": 1.1415137698252995,
+          "policy_rollout_progress": 0.4470727448210572,
+          "policy_rollout_success_rate": 0.23232323232323232,
+          "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_k4s045_mean_safe_margin0p20.json",
+      "num_groups": 575,
+      "selection_mode": "retrieval_residual",
+      "num_candidates": 6,
+      "candidate_sigma": 0.0,
+      "selection_margin": 0.2,
+      "prepend_policy_candidate": false,
+      "field_optim_steps": 0,
+      "field_optim_step_size": 0.0,
+      "field_optim_trust_radius": 0.0,
+      "field_optim_l2_penalty": 0.0,
+      "retrieval_neighbors": 4,
+      "retrieval_metric": "raw",
+      "retrieval_type_min_success": 0.0,
+      "retrieval_residual_scale": 0.45,
+      "retrieval_residual_anchor": "expert",
+      "retrieval_residual_reduce": "mean_by_type",
+      "policy_rollout_success_rate": 0.3391304347826087,
+      "policy_rollout_progress": 0.564176576020925,
+      "oracle_success_rate": 0.8695652173913043,
+      "action_mse_to_best": 0.3899652783669855,
+      "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": 0.32557108186482064,
+          "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": 0.31963159356509213,
+          "expert_success_rate": 0.9402173913043478,
+          "num_groups": 184,
+          "oracle_success_rate": 0.9456521739130435,
+          "policy_expert_regret": 1.078835941274124,
+          "policy_oracle_regret": 1.0788599457490542,
+          "policy_rollout_progress": 0.5718006905593463,
+          "policy_rollout_success_rate": 0.28804347826086957,
+          "restore_max_error": 4.76837158203125e-07
+        },
+        "PullCube-v1": {
+          "action_mse_to_best": 0.5725811961057939,
+          "expert_success_rate": 0.25,
+          "num_groups": 76,
+          "oracle_success_rate": 0.40789473684210525,
+          "policy_expert_regret": 0.32446301054503573,
+          "policy_oracle_regret": 0.5471641206868777,
+          "policy_rollout_progress": 0.278590459450099,
+          "policy_rollout_success_rate": 0.17105263157894737,
+          "restore_max_error": 4.76837158203125e-07
+        },
+        "PushCube-v1": {
+          "action_mse_to_best": 0.35770909411010443,
+          "expert_success_rate": 0.8198198198198198,
+          "num_groups": 111,
+          "oracle_success_rate": 1.0,
+          "policy_expert_regret": 0.3474828630417317,
+          "policy_oracle_regret": 0.3809633700697272,
+          "policy_rollout_progress": 0.817234828128471,
+          "policy_rollout_success_rate": 0.8018018018018018,
+          "restore_max_error": 4.76837158203125e-07
+        },
+        "StackCube-v1": {
+          "action_mse_to_best": 0.4989715301544293,
+          "expert_success_rate": 0.6923076923076923,
+          "num_groups": 91,
+          "oracle_success_rate": 0.8571428571428571,
+          "policy_expert_regret": 1.128073703456711,
+          "policy_oracle_regret": 1.3002214575861837,
+          "policy_rollout_progress": 0.39038386957331017,
+          "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_k4s045_mean_safe_margin0p20.json",
+      "num_groups": 575,
+      "selection_mode": "retrieval_residual",
+      "num_candidates": 6,
+      "candidate_sigma": 0.0,
+      "selection_margin": 0.2,
+      "prepend_policy_candidate": false,
+      "field_optim_steps": 0,
+      "field_optim_step_size": 0.0,
+      "field_optim_trust_radius": 0.0,
+      "field_optim_l2_penalty": 0.0,
+      "retrieval_neighbors": 4,
+      "retrieval_metric": "raw",
+      "retrieval_type_min_success": 0.0,
+      "retrieval_residual_scale": 0.45,
+      "retrieval_residual_anchor": "expert",
+      "retrieval_residual_reduce": "mean_by_type",
+      "policy_rollout_success_rate": 0.36869565217391304,
+      "policy_rollout_progress": 0.5841610958448797,
+      "oracle_success_rate": 0.8765217391304347,
+      "action_mse_to_best": 0.41775175236894385,
+      "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": 0.354422265435763,
+          "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": 0.3356626883638327,
+          "expert_success_rate": 0.9444444444444444,
+          "num_groups": 198,
+          "oracle_success_rate": 0.9595959595959596,
+          "policy_expert_regret": 0.9807057304100858,
+          "policy_oracle_regret": 0.9877891809387942,
+          "policy_rollout_progress": 0.6357829833233898,
+          "policy_rollout_success_rate": 0.3383838383838384,
+          "restore_max_error": 4.76837158203125e-07
+        },
+        "PullCube-v1": {
+          "action_mse_to_best": 0.6288446952899297,
+          "expert_success_rate": 0.24444444444444444,
+          "num_groups": 90,
+          "oracle_success_rate": 0.4666666666666667,
+          "policy_expert_regret": 0.2923288299667547,
+          "policy_oracle_regret": 0.5005992201536275,
+          "policy_rollout_progress": 0.33880305703454117,
+          "policy_rollout_success_rate": 0.23333333333333334,
+          "restore_max_error": 4.0978193283081055e-07
+        },
+        "PushCube-v1": {
+          "action_mse_to_best": 0.38862679190564864,
+          "expert_success_rate": 0.8514851485148515,
+          "num_groups": 101,
+          "oracle_success_rate": 1.0,
+          "policy_expert_regret": 0.38059502485955116,
+          "policy_oracle_regret": 0.4354982612156632,
+          "policy_rollout_progress": 0.7922245110615646,
+          "policy_rollout_success_rate": 0.7722772277227723,
+          "restore_max_error": 4.76837158203125e-07
+        },
+        "StackCube-v1": {
+          "action_mse_to_best": 0.48749088084118236,
+          "expert_success_rate": 0.7666666666666667,
+          "num_groups": 90,
+          "oracle_success_rate": 0.9111111111111111,
+          "policy_expert_regret": 1.1490002441737386,
+          "policy_oracle_regret": 1.310987529820866,
+          "policy_rollout_progress": 0.415471846361955,
+          "policy_rollout_success_rate": 0.15555555555555556,
+          "restore_max_error": 4.76837158203125e-07
+        }
+      }
+    }
+  ]
+}

results/h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_k4s045_mean_safe_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_k4s045_mean_safe_margin0p20.json`
+Completed seeds: 3
+Baseline h=4 policy success: 29.67%
+Baseline h=16 rank-checkpoint success: 29.74%
+
+Mean success: 34.84% +/- 1.76%
+Gain vs h=16 rank checkpoint: +5.10%
+Mean progress: 56.65%
+Mean action MSE to best: 0.396
+
+| seed | mode | k | policy cand | retrieval K | retrieval metric | residual anchor | residual reduce | min type success | residual scale | margin | sigma | opt steps | trust | success | progress | oracle | action MSE |
+|---:|---|---:|---|---:|---|---|---|---:|---:|---:|---:|---:|---:|---:|---:|---:|---:|
+| 0 | retrieval_residual | 6 | no | 4 | raw | expert | mean_by_type | 0.00 | 0.45 | 0.200 | 0.00 | 0 | 0.00 | 33.74% | 55.11% | 85.74% | 0.382 |
+| 1 | retrieval_residual | 6 | no | 4 | raw | expert | mean_by_type | 0.00 | 0.45 | 0.200 | 0.00 | 0 | 0.00 | 33.91% | 56.42% | 86.96% | 0.390 |
+| 2 | retrieval_residual | 6 | no | 4 | raw | expert | mean_by_type | 0.00 | 0.45 | 0.200 | 0.00 | 0 | 0.00 | 36.87% | 58.42% | 87.65% | 0.418 |

results/h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_k8s040_mean_safe_margin0p20_summary.json

@@ -0,0 +1,298 @@
+{
+  "run_root": "/scratch/knguy52/dovla/experiments/dovla_h16_policy_ckpt_runs",
+  "objective": "near_miss_policy_bc5",
+  "out_name": "policy_rollout_retrieval_residual_k8s040_mean_safe_margin0p20.json",
+  "num_completed": 3,
+  "baseline_h4_policy_success": 0.2967,
+  "baseline_h16_rank_checkpoint_success": 0.29739130434782607,
+  "mean_success": 0.3466666666666667,
+  "std_success": 0.01759303293390554,
+  "mean_progress": 0.5649502666741792,
+  "mean_action_mse_to_best": 0.39602510291867066,
+  "gain_vs_h4": 0.04996666666666666,
+  "gain_vs_h16_rank_checkpoint": 0.04927536231884061,
+  "rows": [
+    {
+      "seed": 0,
+      "path": "/scratch/knguy52/dovla/experiments/dovla_h16_policy_ckpt_runs/near_miss_policy_bc5/seed_0/policy_rollout_retrieval_residual_k8s040_mean_safe_margin0p20.json",
+      "num_groups": 575,
+      "selection_mode": "retrieval_residual",
+      "num_candidates": 6,
+      "candidate_sigma": 0.0,
+      "selection_margin": 0.2,
+      "prepend_policy_candidate": false,
+      "field_optim_steps": 0,
+      "field_optim_step_size": 0.0,
+      "field_optim_trust_radius": 0.0,
+      "field_optim_l2_penalty": 0.0,
+      "retrieval_neighbors": 8,
+      "retrieval_metric": "raw",
+      "retrieval_type_min_success": 0.0,
+      "retrieval_residual_scale": 0.4,
+      "retrieval_residual_anchor": "expert",
+      "retrieval_residual_reduce": "mean_by_type",
+      "policy_rollout_success_rate": 0.3373913043478261,
+      "policy_rollout_progress": 0.5501821484829744,
+      "oracle_success_rate": 0.8573913043478261,
+      "action_mse_to_best": 0.3825344390527386,
+      "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": 0.3026177089016598,
+          "expert_success_rate": 0.9375,
+          "num_groups": 208,
+          "oracle_success_rate": 0.9471153846153846,
+          "policy_expert_regret": 0.9752939623207427,
+          "policy_oracle_regret": 0.9887168297472482,
+          "policy_rollout_progress": 0.61158806612142,
+          "policy_rollout_success_rate": 0.3317307692307692,
+          "restore_max_error": 4.76837158203125e-07
+        },
+        "PullCube-v1": {
+          "action_mse_to_best": 0.5987851029479658,
+          "expert_success_rate": 0.19480519480519481,
+          "num_groups": 77,
+          "oracle_success_rate": 0.36363636363636365,
+          "policy_expert_regret": 0.22290432959335385,
+          "policy_oracle_regret": 0.3960622181910296,
+          "policy_rollout_progress": 0.3304314045779794,
+          "policy_rollout_success_rate": 0.23376623376623376,
+          "restore_max_error": 4.76837158203125e-07
+        },
+        "PushCube-v1": {
+          "action_mse_to_best": 0.36385857944000277,
+          "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": 0.46154525525152984,
+          "expert_success_rate": 0.6767676767676768,
+          "num_groups": 99,
+          "oracle_success_rate": 0.8585858585858586,
+          "policy_expert_regret": 0.9759036435021294,
+          "policy_oracle_regret": 1.1500632380596314,
+          "policy_rollout_progress": 0.43852327658672524,
+          "policy_rollout_success_rate": 0.23232323232323232,
+          "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_k8s040_mean_safe_margin0p20.json",
+      "num_groups": 575,
+      "selection_mode": "retrieval_residual",
+      "num_candidates": 6,
+      "candidate_sigma": 0.0,
+      "selection_margin": 0.2,
+      "prepend_policy_candidate": false,
+      "field_optim_steps": 0,
+      "field_optim_step_size": 0.0,
+      "field_optim_trust_radius": 0.0,
+      "field_optim_l2_penalty": 0.0,
+      "retrieval_neighbors": 8,
+      "retrieval_metric": "raw",
+      "retrieval_type_min_success": 0.0,
+      "retrieval_residual_scale": 0.4,
+      "retrieval_residual_anchor": "expert",
+      "retrieval_residual_reduce": "mean_by_type",
+      "policy_rollout_success_rate": 0.3356521739130435,
+      "policy_rollout_progress": 0.5618221815643103,
+      "oracle_success_rate": 0.8695652173913043,
+      "action_mse_to_best": 0.38842434496011424,
+      "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": 0.32557108186482064,
+          "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": 0.3196088104994725,
+          "expert_success_rate": 0.9402173913043478,
+          "num_groups": 184,
+          "oracle_success_rate": 0.9456521739130435,
+          "policy_expert_regret": 1.0786293483359497,
+          "policy_oracle_regret": 1.0786533528108797,
+          "policy_rollout_progress": 0.5720072834975207,
+          "policy_rollout_success_rate": 0.28804347826086957,
+          "restore_max_error": 4.76837158203125e-07
+        },
+        "PullCube-v1": {
+          "action_mse_to_best": 0.5658362184308077,
+          "expert_success_rate": 0.25,
+          "num_groups": 76,
+          "oracle_success_rate": 0.40789473684210525,
+          "policy_expert_regret": 0.32608815941861585,
+          "policy_oracle_regret": 0.5482076280260164,
+          "policy_rollout_progress": 0.2775469521109603,
+          "policy_rollout_success_rate": 0.17105263157894737,
+          "restore_max_error": 4.76837158203125e-07
+        },
+        "PushCube-v1": {
+          "action_mse_to_best": 0.35549953717503463,
+          "expert_success_rate": 0.8198198198198198,
+          "num_groups": 111,
+          "oracle_success_rate": 1.0,
+          "policy_expert_regret": 0.3476062170020095,
+          "policy_oracle_regret": 0.41422087330001967,
+          "policy_rollout_progress": 0.8019953429161966,
+          "policy_rollout_success_rate": 0.7837837837837838,
+          "restore_max_error": 4.76837158203125e-07
+        },
+        "StackCube-v1": {
+          "action_mse_to_best": 0.4976092718930526,
+          "expert_success_rate": 0.6923076923076923,
+          "num_groups": 91,
+          "oracle_success_rate": 0.8571428571428571,
+          "policy_expert_regret": 1.1239077748505624,
+          "policy_oracle_regret": 1.296055528980035,
+          "policy_rollout_progress": 0.3945497981794588,
+          "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_k8s040_mean_safe_margin0p20.json",
+      "num_groups": 575,
+      "selection_mode": "retrieval_residual",
+      "num_candidates": 6,
+      "candidate_sigma": 0.0,
+      "selection_margin": 0.2,
+      "prepend_policy_candidate": false,
+      "field_optim_steps": 0,
+      "field_optim_step_size": 0.0,
+      "field_optim_trust_radius": 0.0,
+      "field_optim_l2_penalty": 0.0,
+      "retrieval_neighbors": 8,
+      "retrieval_metric": "raw",
+      "retrieval_type_min_success": 0.0,
+      "retrieval_residual_scale": 0.4,
+      "retrieval_residual_anchor": "expert",
+      "retrieval_residual_reduce": "mean_by_type",
+      "policy_rollout_success_rate": 0.36695652173913046,
+      "policy_rollout_progress": 0.5828464699752531,
+      "oracle_success_rate": 0.8765217391304347,
+      "action_mse_to_best": 0.4171165247431592,
+      "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": 0.3542259271149912,
+          "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": 0.335321919202353,
+          "expert_success_rate": 0.9444444444444444,
+          "num_groups": 198,
+          "oracle_success_rate": 0.9595959595959596,
+          "policy_expert_regret": 0.995215258382336,
+          "policy_oracle_regret": 1.0022987089110444,
+          "policy_rollout_progress": 0.6313744654521497,
+          "policy_rollout_success_rate": 0.3282828282828283,
+          "restore_max_error": 4.76837158203125e-07
+        },
+        "PullCube-v1": {
+          "action_mse_to_best": 0.6327366744478543,
+          "expert_success_rate": 0.24444444444444444,
+          "num_groups": 90,
+          "oracle_success_rate": 0.4666666666666667,
+          "policy_expert_regret": 0.29303795304838587,
+          "policy_oracle_regret": 0.5010531740952906,
+          "policy_rollout_progress": 0.3383491030928781,
+          "policy_rollout_success_rate": 0.23333333333333334,
+          "restore_max_error": 4.0978193283081055e-07
+        },
+        "PushCube-v1": {
+          "action_mse_to_best": 0.3853067491609271,
+          "expert_success_rate": 0.8514851485148515,
+          "num_groups": 101,
+          "oracle_success_rate": 1.0,
+          "policy_expert_regret": 0.3810940955830093,
+          "policy_oracle_regret": 0.43599733193912127,
+          "policy_rollout_progress": 0.7917254403381064,
+          "policy_rollout_success_rate": 0.7722772277227723,
+          "restore_max_error": 4.76837158203125e-07
+        },
+        "StackCube-v1": {
+          "action_mse_to_best": 0.48422544840723275,
+          "expert_success_rate": 0.7666666666666667,
+          "num_groups": 90,
+          "oracle_success_rate": 0.9111111111111111,
+          "policy_expert_regret": 1.1484904906815954,
+          "policy_oracle_regret": 1.2975626558065414,
+          "policy_rollout_progress": 0.4177856092651685,
+          "policy_rollout_success_rate": 0.16666666666666666,
+          "restore_max_error": 4.76837158203125e-07
+        }
+      }
+    }
+  ]
+}

results/h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_k8s040_mean_safe_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_k8s040_mean_safe_margin0p20.json`
+Completed seeds: 3
+Baseline h=4 policy success: 29.67%
+Baseline h=16 rank-checkpoint success: 29.74%
+
+Mean success: 34.67% +/- 1.76%
+Gain vs h=16 rank checkpoint: +4.93%
+Mean progress: 56.50%
+Mean action MSE to best: 0.396
+
+| seed | mode | k | policy cand | retrieval K | retrieval metric | residual anchor | residual reduce | min type success | residual scale | margin | sigma | opt steps | trust | success | progress | oracle | action MSE |
+|---:|---|---:|---|---:|---|---|---|---:|---:|---:|---:|---:|---:|---:|---:|---:|---:|
+| 0 | retrieval_residual | 6 | no | 8 | raw | expert | mean_by_type | 0.00 | 0.40 | 0.200 | 0.00 | 0 | 0.00 | 33.74% | 55.02% | 85.74% | 0.383 |
+| 1 | retrieval_residual | 6 | no | 8 | raw | expert | mean_by_type | 0.00 | 0.40 | 0.200 | 0.00 | 0 | 0.00 | 33.57% | 56.18% | 86.96% | 0.388 |
+| 2 | retrieval_residual | 6 | no | 8 | raw | expert | mean_by_type | 0.00 | 0.40 | 0.200 | 0.00 | 0 | 0.00 | 36.70% | 58.28% | 87.65% | 0.417 |

results/paper_story_memo.md

@@ -89,7 +89,8 @@ test-time search. The cleaner novelty is:
 
 ## Job Status
 
-Last checked: `2026-06-28 16:46 UTC`. No DoVLA jobs are currently queued.
+Last checked: `2026-06-28 17:14 UTC`. Ray-search jobs are queued, pending on
+GPU priority.
 
 - `14858328`-`14858333`: completed train-split `field_selected_noexpert_bc5`;
   direct rollout is 26.84%, field-guided best is 27.65%.
@@ -136,6 +137,18 @@ Last checked: `2026-06-28 16:46 UTC`. No DoVLA jobs are currently queued.
 - `14868693`-`14868700`: completed clean KNN residual mean-by-type consensus
   sweep. K4, scale `0.40`, margin `0.20` reaches 34.96%, a near tie but still
   below the 35.01% K2 raw residual best.
+- `14868798`-`14868805`: completed consensus follow-up. K4 mean scales `0.425`
+  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`/`14868995`/`14868997`/`14868999`: pending counterfactual tangent
+  ray-search rollouts for K1/K2/K4 safe residual retrieval with scale grids.
+  Summary jobs `14868994`/`14868996`/`14868998`/`14869000` and table rebuild
+  `14869605` are dependency-gated on those rollouts.
+- `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
+  index/type accounting path before rollout.
 
 ## Decision Notes
 

results/paper_table_status.json

@@ -517,6 +517,82 @@
       "best_config": null,
       "gain_vs_h16_policy": 0.05275362318840582
     },
+    {
+      "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",
+      "path": "h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_k1grid_tight_safe_ray_margin0p20_summary.json",
+      "clean_deployment": "yes",
+      "same_state_proposals": "no",
+      "expert_proposal": "no",
+      "story_role": "counterfactual tangent ray-search diagnostic",
+      "fallback_success": null,
+      "pending_job": "14868993/14868994",
+      "path_exists": false,
+      "status": "pending",
+      "success": null,
+      "std_success": null,
+      "completed_seeds": null,
+      "num_completed": null,
+      "best_config": null,
+      "gain_vs_h16_policy": null
+    },
+    {
+      "key": "retrieval_residual_k2grid_tight_safe_ray_margin020",
+      "label": "K2 train-state residual ray search, safe residuals, scales 0.30/0.40/0.50, advantage margin 0.20",
+      "path": "h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_k2grid_tight_safe_ray_margin0p20_summary.json",
+      "clean_deployment": "yes",
+      "same_state_proposals": "no",
+      "expert_proposal": "no",
+      "story_role": "counterfactual tangent ray-search diagnostic",
+      "fallback_success": null,
+      "pending_job": "14868995/14868996",
+      "path_exists": false,
+      "status": "pending",
+      "success": null,
+      "std_success": null,
+      "completed_seeds": null,
+      "num_completed": null,
+      "best_config": null,
+      "gain_vs_h16_policy": null
+    },
+    {
+      "key": "retrieval_residual_k2grid_broad_safe_ray_margin020",
+      "label": "K2 train-state residual ray search, safe residuals, scales 0.20/0.35/0.50/0.65, advantage margin 0.20",
+      "path": "h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_k2grid_broad_safe_ray_margin0p20_summary.json",
+      "clean_deployment": "yes",
+      "same_state_proposals": "no",
+      "expert_proposal": "no",
+      "story_role": "counterfactual tangent ray-search diagnostic",
+      "fallback_success": null,
+      "pending_job": "14868997/14868998",
+      "path_exists": false,
+      "status": "pending",
+      "success": null,
+      "std_success": null,
+      "completed_seeds": null,
+      "num_completed": null,
+      "best_config": null,
+      "gain_vs_h16_policy": null
+    },
+    {
+      "key": "retrieval_residual_k4grid_tight_safe_ray_margin020",
+      "label": "K4 train-state residual ray search, safe residuals, scales 0.30/0.40/0.50, advantage margin 0.20",
+      "path": "h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_k4grid_tight_safe_ray_margin0p20_summary.json",
+      "clean_deployment": "yes",
+      "same_state_proposals": "no",
+      "expert_proposal": "no",
+      "story_role": "counterfactual tangent ray-search diagnostic",
+      "fallback_success": null,
+      "pending_job": "14868999/14869000",
+      "path_exists": false,
+      "status": "pending",
+      "success": null,
+      "std_success": null,
+      "completed_seeds": null,
+      "num_completed": null,
+      "best_config": null,
+      "gain_vs_h16_policy": null
+    },
     {
       "key": "retrieval_residual_k4_scale040_safe_margin020_mean_by_type",
       "label": "K4 train-state residual retrieval, scale 0.40, safe residuals, mean-by-type tangent consensus",

results/paper_table_status.md

@@ -30,6 +30,10 @@ Baseline h=16 policy: 29.74%
 | retrieval_residual_scale035_safe_margin020 | Train-state residual retrieval, scale 0.35, safe residuals, advantage margin 0.20 | complete | 34.84% | +5.10 pp | yes | no | no | counterfactual advantage abstention |
 | retrieval_residual_scale050_safe_margin020 | Train-state residual retrieval, scale 0.50, safe residuals, advantage margin 0.20 | complete | 34.84% | +5.10 pp | yes | no | no | counterfactual advantage abstention scale tie |
 | retrieval_residual_knn2_scale040_safe_margin020 | K2 train-state residual retrieval, scale 0.40, safe residuals, advantage margin 0.20 | complete | 35.01% | +5.28 pp | yes | no | no | best clean counterfactual advantage abstention |
+| 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 | pending 14868993/14868994 | pending | pending | 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 | pending 14868995/14868996 | pending | pending | 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 | pending 14868997/14868998 | 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 | pending 14868999/14869000 | pending | pending | 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 |

scripts/build_paper_table_status.py

@@ -285,6 +285,46 @@ SPECS = [
         story_role="best clean counterfactual advantage abstention",
         pending_job="14862936/14862937",
     ),
+    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",
+        path="h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_k1grid_tight_safe_ray_margin0p20_summary.json",
+        clean_deployment="yes",
+        same_state_proposals="no",
+        expert_proposal="no",
+        story_role="counterfactual tangent ray-search diagnostic",
+        pending_job="14868993/14868994",
+    ),
+    ResultSpec(
+        key="retrieval_residual_k2grid_tight_safe_ray_margin020",
+        label="K2 train-state residual ray search, safe residuals, scales 0.30/0.40/0.50, advantage margin 0.20",
+        path="h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_k2grid_tight_safe_ray_margin0p20_summary.json",
+        clean_deployment="yes",
+        same_state_proposals="no",
+        expert_proposal="no",
+        story_role="counterfactual tangent ray-search diagnostic",
+        pending_job="14868995/14868996",
+    ),
+    ResultSpec(
+        key="retrieval_residual_k2grid_broad_safe_ray_margin020",
+        label="K2 train-state residual ray search, safe residuals, scales 0.20/0.35/0.50/0.65, advantage margin 0.20",
+        path="h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_k2grid_broad_safe_ray_margin0p20_summary.json",
+        clean_deployment="yes",
+        same_state_proposals="no",
+        expert_proposal="no",
+        story_role="counterfactual tangent ray-search diagnostic",
+        pending_job="14868997/14868998",
+    ),
+    ResultSpec(
+        key="retrieval_residual_k4grid_tight_safe_ray_margin020",
+        label="K4 train-state residual ray search, safe residuals, scales 0.30/0.40/0.50, advantage margin 0.20",
+        path="h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_k4grid_tight_safe_ray_margin0p20_summary.json",
+        clean_deployment="yes",
+        same_state_proposals="no",
+        expert_proposal="no",
+        story_role="counterfactual tangent ray-search diagnostic",
+        pending_job="14868999/14869000",
+    ),
     ResultSpec(
         key="retrieval_residual_k4_scale040_safe_margin020_mean_by_type",
         label="K4 train-state residual retrieval, scale 0.40, safe residuals, mean-by-type tangent consensus",

scripts/eval_maniskill_policy_rollout.py

@@ -134,6 +134,12 @@ def main(argv: list[str] | None = None) -> int:
         default=1.0,
         help="Scale applied to train-state residuals before adding them to the policy mean.",
     )
+    parser.add_argument(
+        "--retrieval-residual-scales",
+        default="",
+        help="Optional comma-separated scale grid for counterfactual residual ray search. "
+        "When set, this overrides --retrieval-residual-scale during candidate generation.",
+    )
     parser.add_argument(
         "--retrieval-residual-anchor",
         choices=("expert", "policy"),
@@ -156,6 +162,14 @@ def main(argv: list[str] | None = None) -> int:
     lattice_exclude_types = tuple(
         item.strip() for item in args.lattice_exclude_types.split(",") if item.strip()
     )
+    try:
+        retrieval_residual_scales = tuple(
+            float(item.strip())
+            for item in args.retrieval_residual_scales.split(",")
+            if item.strip()
+        )
+    except ValueError as exc:
+        raise SystemExit("--retrieval-residual-scales must contain comma-separated floats") from exc
     result = evaluate_maniskill_policy_rollout(
         args.checkpoint,
         args.dataset,
@@ -180,6 +194,7 @@ def main(argv: list[str] | None = None) -> int:
         retrieval_metric=args.retrieval_metric,
         retrieval_type_min_success=args.retrieval_type_min_success,
         retrieval_residual_scale=args.retrieval_residual_scale,
+        retrieval_residual_scales=retrieval_residual_scales,
         retrieval_residual_anchor=args.retrieval_residual_anchor,
         retrieval_residual_reduce=args.retrieval_residual_reduce,
         lattice_exclude_types=lattice_exclude_types,

scripts/slurm/eval_maniskill_policy_rollout.sbatch

@@ -54,6 +54,10 @@ RETRIEVAL_NEIGHBORS="${RETRIEVAL_NEIGHBORS:-1}"
 RETRIEVAL_METRIC="${RETRIEVAL_METRIC:-raw}"
 RETRIEVAL_TYPE_MIN_SUCCESS="${RETRIEVAL_TYPE_MIN_SUCCESS:-0.0}"
 RETRIEVAL_RESIDUAL_SCALE="${RETRIEVAL_RESIDUAL_SCALE:-1.0}"
+RETRIEVAL_RESIDUAL_SCALES="${RETRIEVAL_RESIDUAL_SCALES:-}"
+if [[ -n "${RETRIEVAL_RESIDUAL_SCALES_COLON:-}" ]]; then
+  RETRIEVAL_RESIDUAL_SCALES="${RETRIEVAL_RESIDUAL_SCALES_COLON//:/,}"
+fi
 RETRIEVAL_RESIDUAL_ANCHOR="${RETRIEVAL_RESIDUAL_ANCHOR:-expert}"
 RETRIEVAL_RESIDUAL_REDUCE="${RETRIEVAL_RESIDUAL_REDUCE:-none}"
 LATTICE_EXCLUDE_TYPES="${LATTICE_EXCLUDE_TYPES:-}"
@@ -111,6 +115,7 @@ apptainer exec --nv \
     --retrieval-metric "$RETRIEVAL_METRIC" \
     --retrieval-type-min-success "$RETRIEVAL_TYPE_MIN_SUCCESS" \
     --retrieval-residual-scale "$RETRIEVAL_RESIDUAL_SCALE" \
+    --retrieval-residual-scales "$RETRIEVAL_RESIDUAL_SCALES" \
     --retrieval-residual-anchor "$RETRIEVAL_RESIDUAL_ANCHOR" \
     --retrieval-residual-reduce "$RETRIEVAL_RESIDUAL_REDUCE" \
     --lattice-exclude-types "$LATTICE_EXCLUDE_TYPES" \