Auto-sync: 2026-06-28 02:35:57

results/h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_scale0p35_safe_types_summary.json

@@ -0,0 +1,286 @@
+{
+  "run_root": "/scratch/knguy52/dovla/experiments/dovla_h16_policy_ckpt_runs",
+  "objective": "near_miss_policy_bc5",
+  "out_name": "policy_rollout_retrieval_residual_scale0p35_safe_types.json",
+  "num_completed": 3,
+  "baseline_h4_policy_success": 0.2967,
+  "baseline_h16_rank_checkpoint_success": 0.29739130434782607,
+  "mean_success": 0.3373913043478261,
+  "std_success": 0.004601306627938417,
+  "mean_progress": 0.554790610001031,
+  "mean_action_mse_to_best": 0.41091299527203257,
+  "gain_vs_h4": 0.040691304347826085,
+  "gain_vs_h16_rank_checkpoint": 0.040000000000000036,
+  "rows": [
+    {
+      "seed": 0,
+      "path": "/scratch/knguy52/dovla/experiments/dovla_h16_policy_ckpt_runs/near_miss_policy_bc5/seed_0/policy_rollout_retrieval_residual_scale0p35_safe_types.json",
+      "num_groups": 575,
+      "selection_mode": "retrieval_residual",
+      "num_candidates": 16,
+      "candidate_sigma": 0.0,
+      "field_optim_steps": 0,
+      "field_optim_step_size": 0.0,
+      "field_optim_trust_radius": 0.0,
+      "field_optim_l2_penalty": 0.0,
+      "retrieval_neighbors": 1,
+      "retrieval_metric": "raw",
+      "retrieval_type_min_success": 0.0,
+      "retrieval_residual_scale": 0.35,
+      "policy_rollout_success_rate": 0.33217391304347826,
+      "policy_rollout_progress": 0.5461682539368453,
+      "oracle_success_rate": 0.8573913043478261,
+      "action_mse_to_best": 0.3909499480874966,
+      "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": 0.32116451618962527,
+          "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": 0.3100338529867048,
+          "expert_success_rate": 0.9375,
+          "num_groups": 208,
+          "oracle_success_rate": 0.9471153846153846,
+          "policy_expert_regret": 1.022702943139638,
+          "policy_oracle_regret": 1.0294804712399268,
+          "policy_rollout_progress": 0.5900551938595107,
+          "policy_rollout_success_rate": 0.3125,
+          "restore_max_error": 4.76837158203125e-07
+        },
+        "PullCube-v1": {
+          "action_mse_to_best": 0.6079731457806253,
+          "expert_success_rate": 0.19480519480519481,
+          "num_groups": 77,
+          "oracle_success_rate": 0.36363636363636365,
+          "policy_expert_regret": 0.19322007524405982,
+          "policy_oracle_regret": 0.36429626477712934,
+          "policy_rollout_progress": 0.3495055760994747,
+          "policy_rollout_success_rate": 0.24675324675324675,
+          "restore_max_error": 4.76837158203125e-07
+        },
+        "PushCube-v1": {
+          "action_mse_to_best": 0.3693886375886963,
+          "expert_success_rate": 0.8617021276595744,
+          "num_groups": 94,
+          "oracle_success_rate": 0.9893617021276596,
+          "policy_expert_regret": 0.5726731605035194,
+          "policy_oracle_regret": 0.6695628316795572,
+          "policy_rollout_progress": 0.6815009981076768,
+          "policy_rollout_success_rate": 0.648936170212766,
+          "restore_max_error": 4.76837158203125e-07
+        },
+        "StackCube-v1": {
+          "action_mse_to_best": 0.48100764141653224,
+          "expert_success_rate": 0.6767676767676768,
+          "num_groups": 99,
+          "oracle_success_rate": 0.8585858585858586,
+          "policy_expert_regret": 0.9669641819265153,
+          "policy_oracle_regret": 1.1261755732273815,
+          "policy_rollout_progress": 0.4545213539492,
+          "policy_rollout_success_rate": 0.24242424242424243,
+          "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_scale0p35_safe_types.json",
+      "num_groups": 575,
+      "selection_mode": "retrieval_residual",
+      "num_candidates": 16,
+      "candidate_sigma": 0.0,
+      "field_optim_steps": 0,
+      "field_optim_step_size": 0.0,
+      "field_optim_trust_radius": 0.0,
+      "field_optim_l2_penalty": 0.0,
+      "retrieval_neighbors": 1,
+      "retrieval_metric": "raw",
+      "retrieval_type_min_success": 0.0,
+      "retrieval_residual_scale": 0.35,
+      "policy_rollout_success_rate": 0.3391304347826087,
+      "policy_rollout_progress": 0.5584853339502993,
+      "oracle_success_rate": 0.8695652173913043,
+      "action_mse_to_best": 0.4024488061164384,
+      "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": 0.3343446365005411,
+          "expert_success_rate": 0.8584070796460177,
+          "num_groups": 113,
+          "oracle_success_rate": 0.9380530973451328,
+          "policy_expert_regret": 1.0064765876900834,
+          "policy_oracle_regret": 1.0859059261537232,
+          "policy_rollout_progress": 0.6226141650592332,
+          "policy_rollout_success_rate": 0.21238938053097345,
+          "restore_max_error": 4.76837158203125e-07
+        },
+        "PickCube-v1": {
+          "action_mse_to_best": 0.34227522861455445,
+          "expert_success_rate": 0.9402173913043478,
+          "num_groups": 184,
+          "oracle_success_rate": 0.9456521739130435,
+          "policy_expert_regret": 1.0596443941543365,
+          "policy_oracle_regret": 1.0596683986292665,
+          "policy_rollout_progress": 0.5714469643906973,
+          "policy_rollout_success_rate": 0.31521739130434784,
+          "restore_max_error": 4.76837158203125e-07
+        },
+        "PullCube-v1": {
+          "action_mse_to_best": 0.5759768897765561,
+          "expert_success_rate": 0.25,
+          "num_groups": 76,
+          "oracle_success_rate": 0.40789473684210525,
+          "policy_expert_regret": 0.2709970424736017,
+          "policy_oracle_regret": 0.5241831319994832,
+          "policy_rollout_progress": 0.28006213075040204,
+          "policy_rollout_success_rate": 0.17105263157894737,
+          "restore_max_error": 4.76837158203125e-07
+        },
+        "PushCube-v1": {
+          "action_mse_to_best": 0.3723334042472882,
+          "expert_success_rate": 0.8198198198198198,
+          "num_groups": 111,
+          "oracle_success_rate": 1.0,
+          "policy_expert_regret": 0.364276541782929,
+          "policy_oracle_regret": 0.4308047579215454,
+          "policy_rollout_progress": 0.7944204673036799,
+          "policy_rollout_success_rate": 0.7747747747747747,
+          "restore_max_error": 4.76837158203125e-07
+        },
+        "StackCube-v1": {
+          "action_mse_to_best": 0.5004969893719305,
+          "expert_success_rate": 0.6923076923076923,
+          "num_groups": 91,
+          "oracle_success_rate": 0.8571428571428571,
+          "policy_expert_regret": 1.1210726157322035,
+          "policy_oracle_regret": 1.2932203698616762,
+          "policy_rollout_progress": 0.3973849572978177,
+          "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_scale0p35_safe_types.json",
+      "num_groups": 575,
+      "selection_mode": "retrieval_residual",
+      "num_candidates": 16,
+      "candidate_sigma": 0.0,
+      "field_optim_steps": 0,
+      "field_optim_step_size": 0.0,
+      "field_optim_trust_radius": 0.0,
+      "field_optim_l2_penalty": 0.0,
+      "retrieval_neighbors": 1,
+      "retrieval_metric": "raw",
+      "retrieval_type_min_success": 0.0,
+      "retrieval_residual_scale": 0.35,
+      "policy_rollout_success_rate": 0.3408695652173913,
+      "policy_rollout_progress": 0.5597182421159486,
+      "oracle_success_rate": 0.8765217391304347,
+      "action_mse_to_best": 0.43934023161216273,
+      "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": 0.35727682683985523,
+          "expert_success_rate": 0.8229166666666666,
+          "num_groups": 96,
+          "oracle_success_rate": 0.9270833333333334,
+          "policy_expert_regret": 0.8651364526400963,
+          "policy_oracle_regret": 0.9713857987274727,
+          "policy_rollout_progress": 0.6393060578654209,
+          "policy_rollout_success_rate": 0.3229166666666667,
+          "restore_max_error": 3.5762786865234375e-07
+        },
+        "PickCube-v1": {
+          "action_mse_to_best": 0.3646057000703583,
+          "expert_success_rate": 0.9444444444444444,
+          "num_groups": 198,
+          "oracle_success_rate": 0.9595959595959596,
+          "policy_expert_regret": 1.0403793521526485,
+          "policy_oracle_regret": 1.047462802681357,
+          "policy_rollout_progress": 0.6043060528434286,
+          "policy_rollout_success_rate": 0.30303030303030304,
+          "restore_max_error": 4.76837158203125e-07
+        },
+        "PullCube-v1": {
+          "action_mse_to_best": 0.6706881789697541,
+          "expert_success_rate": 0.24444444444444444,
+          "num_groups": 90,
+          "oracle_success_rate": 0.4666666666666667,
+          "policy_expert_regret": 0.319284730259743,
+          "policy_oracle_regret": 0.5459516277950671,
+          "policy_rollout_progress": 0.3085310854638616,
+          "policy_rollout_success_rate": 0.2,
+          "restore_max_error": 4.0978193283081055e-07
+        },
+        "PushCube-v1": {
+          "action_mse_to_best": 0.3935000709820502,
+          "expert_success_rate": 0.8514851485148515,
+          "num_groups": 101,
+          "oracle_success_rate": 1.0,
+          "policy_expert_regret": 0.41857334041949545,
+          "policy_oracle_regret": 0.49220996445948534,
+          "policy_rollout_progress": 0.765215778114772,
+          "policy_rollout_success_rate": 0.7425742574257426,
+          "restore_max_error": 4.76837158203125e-07
+        },
+        "StackCube-v1": {
+          "action_mse_to_best": 0.5113853989996844,
+          "expert_success_rate": 0.7666666666666667,
+          "num_groups": 90,
+          "oracle_success_rate": 0.9111111111111111,
+          "policy_expert_regret": 1.1934410845239958,
+          "policy_oracle_regret": 1.3555614117119048,
+          "policy_rollout_progress": 0.39730464352501765,
+          "policy_rollout_success_rate": 0.13333333333333333,
+          "restore_max_error": 4.76837158203125e-07
+        }
+      }
+    }
+  ]
+}

results/h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_scale0p35_safe_types_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_scale0p35_safe_types.json`
+Completed seeds: 3
+Baseline h=4 policy success: 29.67%
+Baseline h=16 rank-checkpoint success: 29.74%
+
+Mean success: 33.74% +/- 0.46%
+Gain vs h=16 rank checkpoint: +4.00%
+Mean progress: 55.48%
+Mean action MSE to best: 0.411
+
+| seed | mode | k | retrieval K | retrieval metric | min type success | residual scale | sigma | opt steps | trust | success | progress | oracle | action MSE |
+|---:|---|---:|---:|---|---:|---:|---:|---:|---:|---:|---:|---:|---:|
+| 0 | retrieval_residual | 16 | 1 | raw | 0.00 | 0.35 | 0.00 | 0 | 0.00 | 33.22% | 54.62% | 85.74% | 0.391 |
+| 1 | retrieval_residual | 16 | 1 | raw | 0.00 | 0.35 | 0.00 | 0 | 0.00 | 33.91% | 55.85% | 86.96% | 0.402 |
+| 2 | retrieval_residual | 16 | 1 | raw | 0.00 | 0.35 | 0.00 | 0 | 0.00 | 34.09% | 55.97% | 87.65% | 0.439 |

results/h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_scale0p45_safe_types_summary.json

@@ -0,0 +1,286 @@
+{
+  "run_root": "/scratch/knguy52/dovla/experiments/dovla_h16_policy_ckpt_runs",
+  "objective": "near_miss_policy_bc5",
+  "out_name": "policy_rollout_retrieval_residual_scale0p45_safe_types.json",
+  "num_completed": 3,
+  "baseline_h4_policy_success": 0.2967,
+  "baseline_h16_rank_checkpoint_success": 0.29739130434782607,
+  "mean_success": 0.33507246376811595,
+  "std_success": 0.0061076253639726,
+  "mean_progress": 0.5519368337724664,
+  "mean_action_mse_to_best": 0.42269574523080083,
+  "gain_vs_h4": 0.03837246376811593,
+  "gain_vs_h16_rank_checkpoint": 0.03768115942028988,
+  "rows": [
+    {
+      "seed": 0,
+      "path": "/scratch/knguy52/dovla/experiments/dovla_h16_policy_ckpt_runs/near_miss_policy_bc5/seed_0/policy_rollout_retrieval_residual_scale0p45_safe_types.json",
+      "num_groups": 575,
+      "selection_mode": "retrieval_residual",
+      "num_candidates": 16,
+      "candidate_sigma": 0.0,
+      "field_optim_steps": 0,
+      "field_optim_step_size": 0.0,
+      "field_optim_trust_radius": 0.0,
+      "field_optim_l2_penalty": 0.0,
+      "retrieval_neighbors": 1,
+      "retrieval_metric": "raw",
+      "retrieval_type_min_success": 0.0,
+      "retrieval_residual_scale": 0.45,
+      "policy_rollout_success_rate": 0.32869565217391306,
+      "policy_rollout_progress": 0.5425697256052407,
+      "oracle_success_rate": 0.8573913043478261,
+      "action_mse_to_best": 0.3970660170467327,
+      "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": 0.3210236891556877,
+          "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": 0.3156295768701686,
+          "expert_success_rate": 0.9375,
+          "num_groups": 208,
+          "oracle_success_rate": 0.9471153846153846,
+          "policy_expert_regret": 1.0185734751150728,
+          "policy_oracle_regret": 1.0319963425415783,
+          "policy_rollout_progress": 0.5875393225578591,
+          "policy_rollout_success_rate": 0.3125,
+          "restore_max_error": 4.76837158203125e-07
+        },
+        "PullCube-v1": {
+          "action_mse_to_best": 0.6230768923248563,
+          "expert_success_rate": 0.19480519480519481,
+          "num_groups": 77,
+          "oracle_success_rate": 0.36363636363636365,
+          "policy_expert_regret": 0.1950893104721455,
+          "policy_oracle_regret": 0.411565554670957,
+          "policy_rollout_progress": 0.32825910912834744,
+          "policy_rollout_success_rate": 0.22077922077922077,
+          "restore_max_error": 4.76837158203125e-07
+        },
+        "PushCube-v1": {
+          "action_mse_to_best": 0.37244691354955767,
+          "expert_success_rate": 0.8617021276595744,
+          "num_groups": 94,
+          "oracle_success_rate": 0.9893617021276596,
+          "policy_expert_regret": 0.5719866145481455,
+          "policy_oracle_regret": 0.668342428480057,
+          "policy_rollout_progress": 0.6827214013071771,
+          "policy_rollout_success_rate": 0.648936170212766,
+          "restore_max_error": 4.76837158203125e-07
+        },
+        "StackCube-v1": {
+          "action_mse_to_best": 0.49026039779901204,
+          "expert_success_rate": 0.6767676767676768,
+          "num_groups": 99,
+          "oracle_success_rate": 0.8585858585858586,
+          "policy_expert_regret": 0.9613405857423339,
+          "policy_oracle_regret": 1.1242125597265031,
+          "policy_rollout_progress": 0.4542729448188435,
+          "policy_rollout_success_rate": 0.24242424242424243,
+          "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_scale0p45_safe_types.json",
+      "num_groups": 575,
+      "selection_mode": "retrieval_residual",
+      "num_candidates": 16,
+      "candidate_sigma": 0.0,
+      "field_optim_steps": 0,
+      "field_optim_step_size": 0.0,
+      "field_optim_trust_radius": 0.0,
+      "field_optim_l2_penalty": 0.0,
+      "retrieval_neighbors": 1,
+      "retrieval_metric": "raw",
+      "retrieval_type_min_success": 0.0,
+      "retrieval_residual_scale": 0.45,
+      "policy_rollout_success_rate": 0.3356521739130435,
+      "policy_rollout_progress": 0.5554060602512049,
+      "oracle_success_rate": 0.8695652173913043,
+      "action_mse_to_best": 0.416635314171081,
+      "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": 0.33157044279067655,
+          "expert_success_rate": 0.8584070796460177,
+          "num_groups": 113,
+          "oracle_success_rate": 0.9380530973451328,
+          "policy_expert_regret": 0.9967565489026298,
+          "policy_oracle_regret": 1.0770300347193154,
+          "policy_rollout_progress": 0.6226404989715171,
+          "policy_rollout_success_rate": 0.22123893805309736,
+          "restore_max_error": 4.76837158203125e-07
+        },
+        "PickCube-v1": {
+          "action_mse_to_best": 0.3666342041131271,
+          "expert_success_rate": 0.9402173913043478,
+          "num_groups": 184,
+          "oracle_success_rate": 0.9456521739130435,
+          "policy_expert_regret": 1.077671109870805,
+          "policy_oracle_regret": 1.0852080148381547,
+          "policy_rollout_progress": 0.5599415776340048,
+          "policy_rollout_success_rate": 0.29347826086956524,
+          "restore_max_error": 4.76837158203125e-07
+        },
+        "PullCube-v1": {
+          "action_mse_to_best": 0.5990303443058541,
+          "expert_success_rate": 0.25,
+          "num_groups": 76,
+          "oracle_success_rate": 0.40789473684210525,
+          "policy_expert_regret": 0.2711587266163214,
+          "policy_oracle_regret": 0.5430230643649242,
+          "policy_rollout_progress": 0.28427036241383147,
+          "policy_rollout_success_rate": 0.17105263157894737,
+          "restore_max_error": 4.76837158203125e-07
+        },
+        "PushCube-v1": {
+          "action_mse_to_best": 0.38113274311160183,
+          "expert_success_rate": 0.8198198198198198,
+          "num_groups": 111,
+          "oracle_success_rate": 1.0,
+          "policy_expert_regret": 0.3637381719576346,
+          "policy_oracle_regret": 0.41341543143933956,
+          "policy_rollout_progress": 0.8028007847768767,
+          "policy_rollout_success_rate": 0.7837837837837838,
+          "restore_max_error": 4.76837158203125e-07
+        },
+        "StackCube-v1": {
+          "action_mse_to_best": 0.5143418835557916,
+          "expert_success_rate": 0.6923076923076923,
+          "num_groups": 91,
+          "oracle_success_rate": 0.8571428571428571,
+          "policy_expert_regret": 1.1310353387188126,
+          "policy_oracle_regret": 1.3031830928482853,
+          "policy_rollout_progress": 0.38742223431120865,
+          "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_scale0p45_safe_types.json",
+      "num_groups": 575,
+      "selection_mode": "retrieval_residual",
+      "num_candidates": 16,
+      "candidate_sigma": 0.0,
+      "field_optim_steps": 0,
+      "field_optim_step_size": 0.0,
+      "field_optim_trust_radius": 0.0,
+      "field_optim_l2_penalty": 0.0,
+      "retrieval_neighbors": 1,
+      "retrieval_metric": "raw",
+      "retrieval_type_min_success": 0.0,
+      "retrieval_residual_scale": 0.45,
+      "policy_rollout_success_rate": 0.3408695652173913,
+      "policy_rollout_progress": 0.5578347154609535,
+      "oracle_success_rate": 0.8765217391304347,
+      "action_mse_to_best": 0.4543859044745888,
+      "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": 0.35995299029067,
+          "expert_success_rate": 0.8229166666666666,
+          "num_groups": 96,
+          "oracle_success_rate": 0.9270833333333334,
+          "policy_expert_regret": 0.8655313507964214,
+          "policy_oracle_regret": 0.9585438085099062,
+          "policy_rollout_progress": 0.6417313814163208,
+          "policy_rollout_success_rate": 0.3333333333333333,
+          "restore_max_error": 3.5762786865234375e-07
+        },
+        "PickCube-v1": {
+          "action_mse_to_best": 0.3838067414295493,
+          "expert_success_rate": 0.9444444444444444,
+          "num_groups": 198,
+          "oracle_success_rate": 0.9595959595959596,
+          "policy_expert_regret": 1.0759286373439763,
+          "policy_oracle_regret": 1.0830120878726845,
+          "policy_rollout_progress": 0.5869200519201431,
+          "policy_rollout_success_rate": 0.2777777777777778,
+          "restore_max_error": 4.76837158203125e-07
+        },
+        "PullCube-v1": {
+          "action_mse_to_best": 0.685707016951508,
+          "expert_success_rate": 0.24444444444444444,
+          "num_groups": 90,
+          "oracle_success_rate": 0.4666666666666667,
+          "policy_expert_regret": 0.33655670198301474,
+          "policy_oracle_regret": 0.5641340424617132,
+          "policy_rollout_progress": 0.30735235404637123,
+          "policy_rollout_success_rate": 0.2,
+          "restore_max_error": 4.0978193283081055e-07
+        },
+        "PushCube-v1": {
+          "action_mse_to_best": 0.40899270544252775,
+          "expert_success_rate": 0.8514851485148515,
+          "num_groups": 101,
+          "oracle_success_rate": 1.0,
+          "policy_expert_regret": 0.41807408616094305,
+          "policy_oracle_regret": 0.45508477829470495,
+          "policy_rollout_progress": 0.7825389840815327,
+          "policy_rollout_success_rate": 0.7623762376237624,
+          "restore_max_error": 4.76837158203125e-07
+        },
+        "StackCube-v1": {
+          "action_mse_to_best": 0.530008649184472,
+          "expert_success_rate": 0.7666666666666667,
+          "num_groups": 90,
+          "oracle_success_rate": 0.9111111111111111,
+          "policy_expert_regret": 1.1618006545636388,
+          "policy_oracle_regret": 1.3237879402107662,
+          "policy_rollout_progress": 0.4026714359720548,
+          "policy_rollout_success_rate": 0.15555555555555556,
+          "restore_max_error": 4.76837158203125e-07
+        }
+      }
+    }
+  ]
+}

results/h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_scale0p45_safe_types_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_scale0p45_safe_types.json`
+Completed seeds: 3
+Baseline h=4 policy success: 29.67%
+Baseline h=16 rank-checkpoint success: 29.74%
+
+Mean success: 33.51% +/- 0.61%
+Gain vs h=16 rank checkpoint: +3.77%
+Mean progress: 55.19%
+Mean action MSE to best: 0.423
+
+| seed | mode | k | retrieval K | retrieval metric | min type success | residual scale | sigma | opt steps | trust | success | progress | oracle | action MSE |
+|---:|---|---:|---:|---|---:|---:|---:|---:|---:|---:|---:|---:|---:|
+| 0 | retrieval_residual | 16 | 1 | raw | 0.00 | 0.45 | 0.00 | 0 | 0.00 | 32.87% | 54.26% | 85.74% | 0.397 |
+| 1 | retrieval_residual | 16 | 1 | raw | 0.00 | 0.45 | 0.00 | 0 | 0.00 | 33.57% | 55.54% | 86.96% | 0.417 |
+| 2 | retrieval_residual | 16 | 1 | raw | 0.00 | 0.45 | 0.00 | 0 | 0.00 | 34.09% | 55.78% | 87.65% | 0.454 |

results/h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_scale0p60_safe_types_summary.json

@@ -0,0 +1,286 @@
+{
+  "run_root": "/scratch/knguy52/dovla/experiments/dovla_h16_policy_ckpt_runs",
+  "objective": "near_miss_policy_bc5",
+  "out_name": "policy_rollout_retrieval_residual_scale0p60_safe_types.json",
+  "num_completed": 3,
+  "baseline_h4_policy_success": 0.2967,
+  "baseline_h16_rank_checkpoint_success": 0.29739130434782607,
+  "mean_success": 0.3339130434782609,
+  "std_success": 0.007969696860792757,
+  "mean_progress": 0.5559090871016994,
+  "mean_action_mse_to_best": 0.4327756830790768,
+  "gain_vs_h4": 0.03721304347826088,
+  "gain_vs_h16_rank_checkpoint": 0.03652173913043483,
+  "rows": [
+    {
+      "seed": 0,
+      "path": "/scratch/knguy52/dovla/experiments/dovla_h16_policy_ckpt_runs/near_miss_policy_bc5/seed_0/policy_rollout_retrieval_residual_scale0p60_safe_types.json",
+      "num_groups": 575,
+      "selection_mode": "retrieval_residual",
+      "num_candidates": 16,
+      "candidate_sigma": 0.0,
+      "field_optim_steps": 0,
+      "field_optim_step_size": 0.0,
+      "field_optim_trust_radius": 0.0,
+      "field_optim_l2_penalty": 0.0,
+      "retrieval_neighbors": 1,
+      "retrieval_metric": "raw",
+      "retrieval_type_min_success": 0.0,
+      "retrieval_residual_scale": 0.6,
+      "policy_rollout_success_rate": 0.3356521739130435,
+      "policy_rollout_progress": 0.5462673410558668,
+      "oracle_success_rate": 0.8573913043478261,
+      "action_mse_to_best": 0.41363262383267285,
+      "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": 0.33272739927543654,
+          "expert_success_rate": 0.8865979381443299,
+          "num_groups": 97,
+          "oracle_success_rate": 0.9278350515463918,
+          "policy_expert_regret": 1.0792435525926118,
+          "policy_oracle_regret": 1.1056590303010547,
+          "policy_rollout_progress": 0.5705533410163269,
+          "policy_rollout_success_rate": 0.2268041237113402,
+          "restore_max_error": 4.76837158203125e-07
+        },
+        "PickCube-v1": {
+          "action_mse_to_best": 0.319189935313681,
+          "expert_success_rate": 0.9375,
+          "num_groups": 208,
+          "oracle_success_rate": 0.9471153846153846,
+          "policy_expert_regret": 0.9982658499457802,
+          "policy_oracle_regret": 1.0116887173722857,
+          "policy_rollout_progress": 0.5934238708040749,
+          "policy_rollout_success_rate": 0.3269230769230769,
+          "restore_max_error": 4.76837158203125e-07
+        },
+        "PullCube-v1": {
+          "action_mse_to_best": 0.6416945716777405,
+          "expert_success_rate": 0.19480519480519481,
+          "num_groups": 77,
+          "oracle_success_rate": 0.36363636363636365,
+          "policy_expert_regret": 0.21982169559213344,
+          "policy_oracle_regret": 0.4149192056056209,
+          "policy_rollout_progress": 0.32457477085038344,
+          "policy_rollout_success_rate": 0.22077922077922077,
+          "restore_max_error": 4.76837158203125e-07
+        },
+        "PushCube-v1": {
+          "action_mse_to_best": 0.3743356650576313,
+          "expert_success_rate": 0.8617021276595744,
+          "num_groups": 94,
+          "oracle_success_rate": 0.9893617021276596,
+          "policy_expert_regret": 0.5719938382823416,
+          "policy_oracle_regret": 0.6495445995254719,
+          "policy_rollout_progress": 0.6908809323894217,
+          "policy_rollout_success_rate": 0.6595744680851063,
+          "restore_max_error": 4.76837158203125e-07
+        },
+        "StackCube-v1": {
+          "action_mse_to_best": 0.5512591908507124,
+          "expert_success_rate": 0.6767676767676768,
+          "num_groups": 99,
+          "oracle_success_rate": 0.8585858585858586,
+          "policy_expert_regret": 0.9571640711240094,
+          "policy_oracle_regret": 1.119972202512953,
+          "policy_rollout_progress": 0.45851330203239365,
+          "policy_rollout_success_rate": 0.24242424242424243,
+          "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_scale0p60_safe_types.json",
+      "num_groups": 575,
+      "selection_mode": "retrieval_residual",
+      "num_candidates": 16,
+      "candidate_sigma": 0.0,
+      "field_optim_steps": 0,
+      "field_optim_step_size": 0.0,
+      "field_optim_trust_radius": 0.0,
+      "field_optim_l2_penalty": 0.0,
+      "retrieval_neighbors": 1,
+      "retrieval_metric": "raw",
+      "retrieval_type_min_success": 0.0,
+      "retrieval_residual_scale": 0.6,
+      "policy_rollout_success_rate": 0.32521739130434785,
+      "policy_rollout_progress": 0.5533821567909225,
+      "oracle_success_rate": 0.8695652173913043,
+      "action_mse_to_best": 0.4267575810363759,
+      "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": 0.3373497151481235,
+          "expert_success_rate": 0.8584070796460177,
+          "num_groups": 113,
+          "oracle_success_rate": 0.9380530973451328,
+          "policy_expert_regret": 1.0041353826501729,
+          "policy_oracle_regret": 1.0833511376275426,
+          "policy_rollout_progress": 0.6276101456806723,
+          "policy_rollout_success_rate": 0.22123893805309736,
+          "restore_max_error": 4.76837158203125e-07
+        },
+        "PickCube-v1": {
+          "action_mse_to_best": 0.3536196441146667,
+          "expert_success_rate": 0.9402173913043478,
+          "num_groups": 184,
+          "oracle_success_rate": 0.9456521739130435,
+          "policy_expert_regret": 1.1217444096489444,
+          "policy_oracle_regret": 1.1217684141238746,
+          "policy_rollout_progress": 0.550555091391763,
+          "policy_rollout_success_rate": 0.266304347826087,
+          "restore_max_error": 4.76837158203125e-07
+        },
+        "PullCube-v1": {
+          "action_mse_to_best": 0.6196842101451597,
+          "expert_success_rate": 0.25,
+          "num_groups": 76,
+          "oracle_success_rate": 0.40789473684210525,
+          "policy_expert_regret": 0.2989049281704387,
+          "policy_oracle_regret": 0.5711722960082912,
+          "policy_rollout_progress": 0.2677701984311601,
+          "policy_rollout_success_rate": 0.15789473684210525,
+          "restore_max_error": 4.76837158203125e-07
+        },
+        "PushCube-v1": {
+          "action_mse_to_best": 0.40791176601841644,
+          "expert_success_rate": 0.8198198198198198,
+          "num_groups": 111,
+          "oracle_success_rate": 1.0,
+          "policy_expert_regret": 0.34764513883504783,
+          "policy_oracle_regret": 0.39757217507104614,
+          "policy_rollout_progress": 0.8096350321361611,
+          "policy_rollout_success_rate": 0.7927927927927928,
+          "restore_max_error": 4.76837158203125e-07
+        },
+        "StackCube-v1": {
+          "action_mse_to_best": 0.5475260523956392,
+          "expert_success_rate": 0.6923076923076923,
+          "num_groups": 91,
+          "oracle_success_rate": 0.8571428571428571,
+          "policy_expert_regret": 1.136560504580592,
+          "policy_oracle_regret": 1.3087082587100647,
+          "policy_rollout_progress": 0.3928860794384401,
+          "policy_rollout_success_rate": 0.14285714285714285,
+          "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_scale0p60_safe_types.json",
+      "num_groups": 575,
+      "selection_mode": "retrieval_residual",
+      "num_candidates": 16,
+      "candidate_sigma": 0.0,
+      "field_optim_steps": 0,
+      "field_optim_step_size": 0.0,
+      "field_optim_trust_radius": 0.0,
+      "field_optim_l2_penalty": 0.0,
+      "retrieval_neighbors": 1,
+      "retrieval_metric": "raw",
+      "retrieval_type_min_success": 0.0,
+      "retrieval_residual_scale": 0.6,
+      "policy_rollout_success_rate": 0.3408695652173913,
+      "policy_rollout_progress": 0.568077763458309,
+      "oracle_success_rate": 0.8765217391304347,
+      "action_mse_to_best": 0.4579368443681818,
+      "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": 0.3749906163769386,
+          "expert_success_rate": 0.8229166666666666,
+          "num_groups": 96,
+          "oracle_success_rate": 0.9270833333333334,
+          "policy_expert_regret": 0.910795712377876,
+          "policy_oracle_regret": 1.0306057695609827,
+          "policy_rollout_progress": 0.6217527536985775,
+          "policy_rollout_success_rate": 0.28125,
+          "restore_max_error": 3.5762786865234375e-07
+        },
+        "PickCube-v1": {
+          "action_mse_to_best": 0.3769689722908567,
+          "expert_success_rate": 0.9444444444444444,
+          "num_groups": 198,
+          "oracle_success_rate": 0.9595959595959596,
+          "policy_expert_regret": 1.0650019904485706,
+          "policy_oracle_regret": 1.072085440977279,
+          "policy_rollout_progress": 0.5978466988155486,
+          "policy_rollout_success_rate": 0.2777777777777778,
+          "restore_max_error": 4.76837158203125e-07
+        },
+        "PullCube-v1": {
+          "action_mse_to_best": 0.6606155963407623,
+          "expert_success_rate": 0.24444444444444444,
+          "num_groups": 90,
+          "oracle_success_rate": 0.4666666666666667,
+          "policy_expert_regret": 0.2944552243169811,
+          "policy_oracle_regret": 0.5017143970148431,
+          "policy_rollout_progress": 0.3372603339039617,
+          "policy_rollout_success_rate": 0.23333333333333334,
+          "restore_max_error": 4.0978193283081055e-07
+        },
+        "PushCube-v1": {
+          "action_mse_to_best": 0.4256735229403666,
+          "expert_success_rate": 0.8514851485148515,
+          "num_groups": 101,
+          "oracle_success_rate": 1.0,
+          "policy_expert_regret": 0.3820718020790874,
+          "policy_oracle_regret": 0.4366904573865456,
+          "policy_rollout_progress": 0.7910323148906822,
+          "policy_rollout_success_rate": 0.7722772277227723,
+          "restore_max_error": 4.76837158203125e-07
+        },
+        "StackCube-v1": {
+          "action_mse_to_best": 0.558070003758702,
+          "expert_success_rate": 0.7666666666666667,
+          "num_groups": 90,
+          "oracle_success_rate": 0.9111111111111111,
+          "policy_expert_regret": 1.1442130752735669,
+          "policy_oracle_regret": 1.293285240398513,
+          "policy_rollout_progress": 0.4259456601407793,
+          "policy_rollout_success_rate": 0.16666666666666666,
+          "restore_max_error": 4.76837158203125e-07
+        }
+      }
+    }
+  ]
+}

results/h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_scale0p60_safe_types_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_scale0p60_safe_types.json`
+Completed seeds: 3
+Baseline h=4 policy success: 29.67%
+Baseline h=16 rank-checkpoint success: 29.74%
+
+Mean success: 33.39% +/- 0.80%
+Gain vs h=16 rank checkpoint: +3.65%
+Mean progress: 55.59%
+Mean action MSE to best: 0.433
+
+| seed | mode | k | retrieval K | retrieval metric | min type success | residual scale | sigma | opt steps | trust | success | progress | oracle | action MSE |
+|---:|---|---:|---:|---|---:|---:|---:|---:|---:|---:|---:|---:|---:|
+| 0 | retrieval_residual | 16 | 1 | raw | 0.00 | 0.60 | 0.00 | 0 | 0.00 | 33.57% | 54.63% | 85.74% | 0.414 |
+| 1 | retrieval_residual | 16 | 1 | raw | 0.00 | 0.60 | 0.00 | 0 | 0.00 | 32.52% | 55.34% | 86.96% | 0.427 |
+| 2 | retrieval_residual | 16 | 1 | raw | 0.00 | 0.60 | 0.00 | 0 | 0.00 | 34.09% | 56.81% | 87.65% | 0.458 |

results/h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_scale0p70_safe_types_summary.json

@@ -0,0 +1,286 @@
+{
+  "run_root": "/scratch/knguy52/dovla/experiments/dovla_h16_policy_ckpt_runs",
+  "objective": "near_miss_policy_bc5",
+  "out_name": "policy_rollout_retrieval_residual_scale0p70_safe_types.json",
+  "num_completed": 3,
+  "baseline_h4_policy_success": 0.2967,
+  "baseline_h16_rank_checkpoint_success": 0.29739130434782607,
+  "mean_success": 0.33159420289855074,
+  "std_success": 0.00957838356049755,
+  "mean_progress": 0.5543574167716492,
+  "mean_action_mse_to_best": 0.4590698274988951,
+  "gain_vs_h4": 0.03489420289855072,
+  "gain_vs_h16_rank_checkpoint": 0.03420289855072467,
+  "rows": [
+    {
+      "seed": 0,
+      "path": "/scratch/knguy52/dovla/experiments/dovla_h16_policy_ckpt_runs/near_miss_policy_bc5/seed_0/policy_rollout_retrieval_residual_scale0p70_safe_types.json",
+      "num_groups": 575,
+      "selection_mode": "retrieval_residual",
+      "num_candidates": 16,
+      "candidate_sigma": 0.0,
+      "field_optim_steps": 0,
+      "field_optim_step_size": 0.0,
+      "field_optim_trust_radius": 0.0,
+      "field_optim_l2_penalty": 0.0,
+      "retrieval_neighbors": 1,
+      "retrieval_metric": "raw",
+      "retrieval_type_min_success": 0.0,
+      "retrieval_residual_scale": 0.7,
+      "policy_rollout_success_rate": 0.32521739130434785,
+      "policy_rollout_progress": 0.5439443577093113,
+      "oracle_success_rate": 0.8573913043478261,
+      "action_mse_to_best": 0.45933493631041566,
+      "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": 0.4931159399008167,
+          "expert_success_rate": 0.8865979381443299,
+          "num_groups": 97,
+          "oracle_success_rate": 0.9278350515463918,
+          "policy_expert_regret": 1.1324601213342136,
+          "policy_oracle_regret": 1.1720696789087708,
+          "policy_rollout_progress": 0.5453798058106727,
+          "policy_rollout_success_rate": 0.18556701030927836,
+          "restore_max_error": 4.76837158203125e-07
+        },
+        "PickCube-v1": {
+          "action_mse_to_best": 0.3283846854733733,
+          "expert_success_rate": 0.9375,
+          "num_groups": 208,
+          "oracle_success_rate": 0.9471153846153846,
+          "policy_expert_regret": 1.0071088800815722,
+          "policy_oracle_regret": 1.0205317475080777,
+          "policy_rollout_progress": 0.5941962252836674,
+          "policy_rollout_success_rate": 0.3173076923076923,
+          "restore_max_error": 4.76837158203125e-07
+        },
+        "PullCube-v1": {
+          "action_mse_to_best": 0.6572364510653855,
+          "expert_success_rate": 0.19480519480519481,
+          "num_groups": 77,
+          "oracle_success_rate": 0.36363636363636365,
+          "policy_expert_regret": 0.21923505429287332,
+          "policy_oracle_regret": 0.43562990036647914,
+          "policy_rollout_progress": 0.3168113688702465,
+          "policy_rollout_success_rate": 0.2077922077922078,
+          "restore_max_error": 4.76837158203125e-07
+        },
+        "PushCube-v1": {
+          "action_mse_to_best": 0.3978812553067791,
+          "expert_success_rate": 0.8617021276595744,
+          "num_groups": 94,
+          "oracle_success_rate": 0.9893617021276596,
+          "policy_expert_regret": 0.5339092030487163,
+          "policy_oracle_regret": 0.6114599642918465,
+          "policy_rollout_progress": 0.7076889718783662,
+          "policy_rollout_success_rate": 0.6808510638297872,
+          "restore_max_error": 4.76837158203125e-07
+        },
+        "StackCube-v1": {
+          "action_mse_to_best": 0.6057907363512751,
+          "expert_success_rate": 0.6767676767676768,
+          "num_groups": 99,
+          "oracle_success_rate": 0.8585858585858586,
+          "policy_expert_regret": 0.9716372298772888,
+          "policy_oracle_regret": 1.1306861774487929,
+          "policy_rollout_progress": 0.45814253434990393,
+          "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_scale0p70_safe_types.json",
+      "num_groups": 575,
+      "selection_mode": "retrieval_residual",
+      "num_candidates": 16,
+      "candidate_sigma": 0.0,
+      "field_optim_steps": 0,
+      "field_optim_step_size": 0.0,
+      "field_optim_trust_radius": 0.0,
+      "field_optim_l2_penalty": 0.0,
+      "retrieval_neighbors": 1,
+      "retrieval_metric": "raw",
+      "retrieval_type_min_success": 0.0,
+      "retrieval_residual_scale": 0.7,
+      "policy_rollout_success_rate": 0.3269565217391304,
+      "policy_rollout_progress": 0.5545947884606278,
+      "oracle_success_rate": 0.8695652173913043,
+      "action_mse_to_best": 0.4435377200887255,
+      "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": 0.33948258955888777,
+          "expert_success_rate": 0.8584070796460177,
+          "num_groups": 113,
+          "oracle_success_rate": 0.9380530973451328,
+          "policy_expert_regret": 1.0030630766290478,
+          "policy_oracle_regret": 1.0841963428311643,
+          "policy_rollout_progress": 0.6267649404770506,
+          "policy_rollout_success_rate": 0.22123893805309736,
+          "restore_max_error": 4.76837158203125e-07
+        },
+        "PickCube-v1": {
+          "action_mse_to_best": 0.374743456689074,
+          "expert_success_rate": 0.9402173913043478,
+          "num_groups": 184,
+          "oracle_success_rate": 0.9456521739130435,
+          "policy_expert_regret": 1.1397522360430625,
+          "policy_oracle_regret": 1.1397762405179928,
+          "policy_rollout_progress": 0.5434168302150362,
+          "policy_rollout_success_rate": 0.2554347826086957,
+          "restore_max_error": 4.76837158203125e-07
+        },
+        "PullCube-v1": {
+          "action_mse_to_best": 0.6159595158931456,
+          "expert_success_rate": 0.25,
+          "num_groups": 76,
+          "oracle_success_rate": 0.40789473684210525,
+          "policy_expert_regret": 0.2513730524745035,
+          "policy_oracle_regret": 0.5007567832883644,
+          "policy_rollout_progress": 0.2984259322922873,
+          "policy_rollout_success_rate": 0.19736842105263158,
+          "restore_max_error": 4.76837158203125e-07
+        },
+        "PushCube-v1": {
+          "action_mse_to_best": 0.4280071382721265,
+          "expert_success_rate": 0.8198198198198198,
+          "num_groups": 111,
+          "oracle_success_rate": 1.0,
+          "policy_expert_regret": 0.3479747251347379,
+          "policy_oracle_regret": 0.4141823646721539,
+          "policy_rollout_progress": 0.8020338515440623,
+          "policy_rollout_success_rate": 0.7837837837837838,
+          "restore_max_error": 4.76837158203125e-07
+        },
+        "StackCube-v1": {
+          "action_mse_to_best": 0.5867928004833859,
+          "expert_success_rate": 0.6923076923076923,
+          "num_groups": 91,
+          "oracle_success_rate": 0.8571428571428571,
+          "policy_expert_regret": 1.1187571339554838,
+          "policy_oracle_regret": 1.2909048880849565,
+          "policy_rollout_progress": 0.3997004390745373,
+          "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_scale0p70_safe_types.json",
+      "num_groups": 575,
+      "selection_mode": "retrieval_residual",
+      "num_candidates": 16,
+      "candidate_sigma": 0.0,
+      "field_optim_steps": 0,
+      "field_optim_step_size": 0.0,
+      "field_optim_trust_radius": 0.0,
+      "field_optim_l2_penalty": 0.0,
+      "retrieval_neighbors": 1,
+      "retrieval_metric": "raw",
+      "retrieval_type_min_success": 0.0,
+      "retrieval_residual_scale": 0.7,
+      "policy_rollout_success_rate": 0.3426086956521739,
+      "policy_rollout_progress": 0.5645331041450086,
+      "oracle_success_rate": 0.8765217391304347,
+      "action_mse_to_best": 0.47433682609754413,
+      "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": 0.39612667114609695,
+          "expert_success_rate": 0.8229166666666666,
+          "num_groups": 96,
+          "oracle_success_rate": 0.9270833333333334,
+          "policy_expert_regret": 0.8934455127455294,
+          "policy_oracle_regret": 1.0133062531240284,
+          "policy_rollout_progress": 0.6182189368021985,
+          "policy_rollout_success_rate": 0.3020833333333333,
+          "restore_max_error": 3.5762786865234375e-07
+        },
+        "PickCube-v1": {
+          "action_mse_to_best": 0.39531708213340755,
+          "expert_success_rate": 0.9444444444444444,
+          "num_groups": 198,
+          "oracle_success_rate": 0.9595959595959596,
+          "policy_expert_regret": 1.0540599665519865,
+          "policy_oracle_regret": 1.061143417080695,
+          "policy_rollout_progress": 0.6037382176616276,
+          "policy_rollout_success_rate": 0.2828282828282828,
+          "restore_max_error": 4.76837158203125e-07
+        },
+        "PullCube-v1": {
+          "action_mse_to_best": 0.6620482171575228,
+          "expert_success_rate": 0.24444444444444444,
+          "num_groups": 90,
+          "oracle_success_rate": 0.4666666666666667,
+          "policy_expert_regret": 0.29207462275193796,
+          "policy_oracle_regret": 0.49979850695365,
+          "policy_rollout_progress": 0.3396940696156687,
+          "policy_rollout_success_rate": 0.23333333333333334,
+          "restore_max_error": 4.0978193283081055e-07
+        },
+        "PushCube-v1": {
+          "action_mse_to_best": 0.4231516254125255,
+          "expert_success_rate": 0.8514851485148515,
+          "num_groups": 101,
+          "oracle_success_rate": 1.0,
+          "policy_expert_regret": 0.40259190110286863,
+          "policy_oracle_regret": 0.47524152327292035,
+          "policy_rollout_progress": 0.7722832292023272,
+          "policy_rollout_success_rate": 0.7524752475247525,
+          "restore_max_error": 4.76837158203125e-07
+        },
+        "StackCube-v1": {
+          "action_mse_to_best": 0.6013342066978414,
+          "expert_success_rate": 0.7666666666666667,
+          "num_groups": 90,
+          "oracle_success_rate": 0.9111111111111111,
+          "policy_expert_regret": 1.153561906185415,
+          "policy_oracle_regret": 1.3026340713103612,
+          "policy_rollout_progress": 0.4127141937613487,
+          "policy_rollout_success_rate": 0.16666666666666666,
+          "restore_max_error": 4.76837158203125e-07
+        }
+      }
+    }
+  ]
+}

results/h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_scale0p70_safe_types_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_scale0p70_safe_types.json`
+Completed seeds: 3
+Baseline h=4 policy success: 29.67%
+Baseline h=16 rank-checkpoint success: 29.74%
+
+Mean success: 33.16% +/- 0.96%
+Gain vs h=16 rank checkpoint: +3.42%
+Mean progress: 55.44%
+Mean action MSE to best: 0.459
+
+| seed | mode | k | retrieval K | retrieval metric | min type success | residual scale | sigma | opt steps | trust | success | progress | oracle | action MSE |
+|---:|---|---:|---:|---|---:|---:|---:|---:|---:|---:|---:|---:|---:|
+| 0 | retrieval_residual | 16 | 1 | raw | 0.00 | 0.70 | 0.00 | 0 | 0.00 | 32.52% | 54.39% | 85.74% | 0.459 |
+| 1 | retrieval_residual | 16 | 1 | raw | 0.00 | 0.70 | 0.00 | 0 | 0.00 | 32.70% | 55.46% | 86.96% | 0.444 |
+| 2 | retrieval_residual | 16 | 1 | raw | 0.00 | 0.70 | 0.00 | 0 | 0.00 | 34.26% | 56.45% | 87.65% | 0.474 |

results/paper_table_status.json

@@ -432,6 +432,25 @@
       "story_role": "typed tangent scale fine sweep",
       "fallback_success": null,
       "pending_job": "14859503/14859504",
+      "path_exists": true,
+      "status": "complete",
+      "success": 0.3373913043478261,
+      "std_success": 0.004601306627938417,
+      "completed_seeds": null,
+      "num_completed": 3,
+      "best_config": null,
+      "gain_vs_h16_policy": 0.040000000000000036
+    },
+    {
+      "key": "retrieval_residual_scale030_safe_types",
+      "label": "Train-state residual retrieval, scale 0.30, policy/no-op/wrong-gripper residuals",
+      "path": "h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_scale0p30_safe_types_summary.json",
+      "clean_deployment": "yes",
+      "same_state_proposals": "no",
+      "expert_proposal": "no",
+      "story_role": "typed tangent scale zoom sweep",
+      "fallback_success": null,
+      "pending_job": "14859589/14859590",
       "path_exists": false,
       "status": "pending",
       "success": null,
@@ -442,15 +461,15 @@
       "gain_vs_h16_policy": null
     },
     {
-      "key": "retrieval_residual_scale045_safe_types",
-      "label": "Train-state residual retrieval, scale 0.45, policy/no-op/wrong-gripper residuals",
-      "path": "h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_scale0p45_safe_types_summary.json",
+      "key": "retrieval_residual_scale0325_safe_types",
+      "label": "Train-state residual retrieval, scale 0.325, policy/no-op/wrong-gripper residuals",
+      "path": "h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_scale0p325_safe_types_summary.json",
       "clean_deployment": "yes",
       "same_state_proposals": "no",
       "expert_proposal": "no",
-      "story_role": "typed tangent scale fine sweep",
+      "story_role": "typed tangent scale zoom sweep",
       "fallback_success": null,
-      "pending_job": "14859505/14859506",
+      "pending_job": "14859591/14859592",
       "path_exists": false,
       "status": "pending",
       "success": null,
@@ -461,15 +480,15 @@
       "gain_vs_h16_policy": null
     },
     {
-      "key": "retrieval_residual_scale060_safe_types",
-      "label": "Train-state residual retrieval, scale 0.60, policy/no-op/wrong-gripper residuals",
-      "path": "h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_scale0p60_safe_types_summary.json",
+      "key": "retrieval_residual_scale0375_safe_types",
+      "label": "Train-state residual retrieval, scale 0.375, policy/no-op/wrong-gripper residuals",
+      "path": "h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_scale0p375_safe_types_summary.json",
       "clean_deployment": "yes",
       "same_state_proposals": "no",
       "expert_proposal": "no",
-      "story_role": "typed tangent scale fine sweep",
+      "story_role": "typed tangent scale zoom sweep",
       "fallback_success": null,
-      "pending_job": "14859507/14859508",
+      "pending_job": "14859593/14859594",
       "path_exists": false,
       "status": "pending",
       "success": null,
@@ -480,15 +499,15 @@
       "gain_vs_h16_policy": null
     },
     {
-      "key": "retrieval_residual_scale070_safe_types",
-      "label": "Train-state residual retrieval, scale 0.70, policy/no-op/wrong-gripper residuals",
-      "path": "h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_scale0p70_safe_types_summary.json",
+      "key": "retrieval_residual_scale040_safe_types",
+      "label": "Train-state residual retrieval, scale 0.40, policy/no-op/wrong-gripper residuals",
+      "path": "h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_scale0p40_safe_types_summary.json",
       "clean_deployment": "yes",
       "same_state_proposals": "no",
       "expert_proposal": "no",
-      "story_role": "typed tangent scale fine sweep",
+      "story_role": "typed tangent scale zoom sweep",
       "fallback_success": null,
-      "pending_job": "14859509/14859510",
+      "pending_job": "14859595/14859596",
       "path_exists": false,
       "status": "pending",
       "success": null,
@@ -498,6 +517,63 @@
       "best_config": null,
       "gain_vs_h16_policy": null
     },
+    {
+      "key": "retrieval_residual_scale045_safe_types",
+      "label": "Train-state residual retrieval, scale 0.45, policy/no-op/wrong-gripper residuals",
+      "path": "h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_scale0p45_safe_types_summary.json",
+      "clean_deployment": "yes",
+      "same_state_proposals": "no",
+      "expert_proposal": "no",
+      "story_role": "typed tangent scale fine sweep",
+      "fallback_success": null,
+      "pending_job": "14859505/14859506",
+      "path_exists": true,
+      "status": "complete",
+      "success": 0.33507246376811595,
+      "std_success": 0.0061076253639726,
+      "completed_seeds": null,
+      "num_completed": 3,
+      "best_config": null,
+      "gain_vs_h16_policy": 0.03768115942028988
+    },
+    {
+      "key": "retrieval_residual_scale060_safe_types",
+      "label": "Train-state residual retrieval, scale 0.60, policy/no-op/wrong-gripper residuals",
+      "path": "h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_scale0p60_safe_types_summary.json",
+      "clean_deployment": "yes",
+      "same_state_proposals": "no",
+      "expert_proposal": "no",
+      "story_role": "typed tangent scale fine sweep",
+      "fallback_success": null,
+      "pending_job": "14859507/14859508",
+      "path_exists": true,
+      "status": "complete",
+      "success": 0.3339130434782609,
+      "std_success": 0.007969696860792757,
+      "completed_seeds": null,
+      "num_completed": 3,
+      "best_config": null,
+      "gain_vs_h16_policy": 0.03652173913043483
+    },
+    {
+      "key": "retrieval_residual_scale070_safe_types",
+      "label": "Train-state residual retrieval, scale 0.70, policy/no-op/wrong-gripper residuals",
+      "path": "h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_scale0p70_safe_types_summary.json",
+      "clean_deployment": "yes",
+      "same_state_proposals": "no",
+      "expert_proposal": "no",
+      "story_role": "typed tangent scale fine sweep",
+      "fallback_success": null,
+      "pending_job": "14859509/14859510",
+      "path_exists": true,
+      "status": "complete",
+      "success": 0.33159420289855074,
+      "std_success": 0.00957838356049755,
+      "completed_seeds": null,
+      "num_completed": 3,
+      "best_config": null,
+      "gain_vs_h16_policy": 0.03420289855072467
+    },
     {
       "key": "retrieval_residual_scale050_type_success010",
       "label": "Train-state residual retrieval, scale 0.50, train family success >= 0.10",
@@ -766,23 +842,23 @@
     }
   ],
   "best_clean": {
-    "key": "retrieval_residual_scale050_safe_types",
-    "label": "Train-state residual retrieval, scale 0.50, policy/no-op/wrong-gripper residuals",
-    "path": "h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_scale0p50_safe_types_summary.json",
+    "key": "retrieval_residual_scale035_safe_types",
+    "label": "Train-state residual retrieval, scale 0.35, policy/no-op/wrong-gripper residuals",
+    "path": "h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_scale0p35_safe_types_summary.json",
     "clean_deployment": "yes",
     "same_state_proposals": "no",
     "expert_proposal": "no",
-    "story_role": "typed tangent-family mask ablation",
+    "story_role": "typed tangent scale fine sweep",
     "fallback_success": null,
-    "pending_job": "14859195/14859196",
+    "pending_job": "14859503/14859504",
     "path_exists": true,
     "status": "complete",
-    "success": 0.3368115942028986,
-    "std_success": 0.006584241560348136,
+    "success": 0.3373913043478261,
+    "std_success": 0.004601306627938417,
     "completed_seeds": null,
     "num_completed": 3,
     "best_config": null,
-    "gain_vs_h16_policy": 0.03942028985507251
+    "gain_vs_h16_policy": 0.040000000000000036
   },
   "best_mechanism_no_expert": {
     "key": "no_expert_lattice",
@@ -807,7 +883,7 @@
     "Use no-expert same-state lattice as the conservative mechanism result, not as deployment-clean inference.",
     "Use full lattice only as an upper result because it includes expert proposals.",
     "Do not claim external SOTA from this table alone; add current external baselines separately.",
-    "Current best clean deployment row is Train-state residual retrieval, scale 0.50, policy/no-op/wrong-gripper residuals at 33.68%.",
+    "Current best clean deployment row is Train-state residual retrieval, scale 0.35, policy/no-op/wrong-gripper residuals at 33.74%.",
     "Trust-region field optimization should be framed as a negative/diagnostic ablation.",
     "Train-state counterfactual residual retrieval is a positive clean bridge but remains below the current clean best.",
     "KNN counterfactual residual retrieval is a positive clean bridge but remains below the current clean best."

results/paper_table_status.md

@@ -25,10 +25,14 @@ Baseline h=16 policy: 29.74%
 | retrieval_residual_scale050_no_random_wrongdir | Train-state residual retrieval, scale 0.50, no random/wrong-direction residuals | complete | 33.57% | +3.83 pp | yes | no | no | anti-goal residual family mask ablation |
 | retrieval_residual_scale025_no_random_wrongdir | Train-state residual retrieval, scale 0.25, no random/wrong-direction residuals | complete | 33.45% | +3.71 pp | yes | no | no | anti-goal residual family mask ablation |
 | retrieval_residual_scale050_safe_types | Train-state residual retrieval, scale 0.50, policy/no-op/wrong-gripper residuals | complete | 33.68% | +3.94 pp | yes | no | no | typed tangent-family mask ablation |
-| retrieval_residual_scale035_safe_types | Train-state residual retrieval, scale 0.35, policy/no-op/wrong-gripper residuals | pending 14859503/14859504 | pending | pending | yes | no | no | typed tangent scale fine sweep |
-| retrieval_residual_scale045_safe_types | Train-state residual retrieval, scale 0.45, policy/no-op/wrong-gripper residuals | pending 14859505/14859506 | pending | pending | yes | no | no | typed tangent scale fine sweep |
-| retrieval_residual_scale060_safe_types | Train-state residual retrieval, scale 0.60, policy/no-op/wrong-gripper residuals | pending 14859507/14859508 | pending | pending | yes | no | no | typed tangent scale fine sweep |
-| retrieval_residual_scale070_safe_types | Train-state residual retrieval, scale 0.70, policy/no-op/wrong-gripper residuals | pending 14859509/14859510 | pending | pending | yes | no | no | typed tangent scale fine sweep |
+| retrieval_residual_scale035_safe_types | Train-state residual retrieval, scale 0.35, policy/no-op/wrong-gripper residuals | complete | 33.74% | +4.00 pp | yes | no | no | typed tangent scale fine sweep |
+| retrieval_residual_scale030_safe_types | Train-state residual retrieval, scale 0.30, policy/no-op/wrong-gripper residuals | pending 14859589/14859590 | pending | pending | yes | no | no | typed tangent scale zoom sweep |
+| retrieval_residual_scale0325_safe_types | Train-state residual retrieval, scale 0.325, policy/no-op/wrong-gripper residuals | pending 14859591/14859592 | pending | pending | yes | no | no | typed tangent scale zoom sweep |
+| retrieval_residual_scale0375_safe_types | Train-state residual retrieval, scale 0.375, policy/no-op/wrong-gripper residuals | pending 14859593/14859594 | pending | pending | yes | no | no | typed tangent scale zoom sweep |
+| retrieval_residual_scale040_safe_types | Train-state residual retrieval, scale 0.40, policy/no-op/wrong-gripper residuals | pending 14859595/14859596 | pending | pending | yes | no | no | typed tangent scale zoom sweep |
+| retrieval_residual_scale045_safe_types | Train-state residual retrieval, scale 0.45, policy/no-op/wrong-gripper residuals | complete | 33.51% | +3.77 pp | yes | no | no | typed tangent scale fine sweep |
+| retrieval_residual_scale060_safe_types | Train-state residual retrieval, scale 0.60, policy/no-op/wrong-gripper residuals | complete | 33.39% | +3.65 pp | yes | no | no | typed tangent scale fine sweep |
+| retrieval_residual_scale070_safe_types | Train-state residual retrieval, scale 0.70, policy/no-op/wrong-gripper residuals | complete | 33.16% | +3.42 pp | yes | no | no | typed tangent scale fine sweep |
 | retrieval_residual_scale050_type_success010 | Train-state residual retrieval, scale 0.50, train family success >= 0.10 | complete | 33.33% | +3.59 pp | yes | no | no | train-split residual family reliability prior |
 | retrieval_residual_scale050_type_success025 | Train-state residual retrieval, scale 0.50, train family success >= 0.25 | complete | 33.33% | +3.59 pp | yes | no | no | train-split residual family reliability prior |
 | retrieval_residual_scale050_type_success050 | Train-state residual retrieval, scale 0.50, train family success >= 0.50 | complete | 33.33% | +3.59 pp | yes | no | no | train-split residual family reliability prior |
@@ -49,7 +53,7 @@ Baseline h=16 policy: 29.74%
 - Use no-expert same-state lattice as the conservative mechanism result, not as deployment-clean inference.
 - Use full lattice only as an upper result because it includes expert proposals.
 - Do not claim external SOTA from this table alone; add current external baselines separately.
-- Current best clean deployment row is Train-state residual retrieval, scale 0.50, policy/no-op/wrong-gripper residuals at 33.68%.
+- Current best clean deployment row is Train-state residual retrieval, scale 0.35, policy/no-op/wrong-gripper residuals at 33.74%.
 - Trust-region field optimization should be framed as a negative/diagnostic ablation.
 - Train-state counterfactual residual retrieval is a positive clean bridge but remains below the current clean best.
 - KNN counterfactual residual retrieval is a positive clean bridge but remains below the current clean best.

scripts/build_paper_table_status.py

@@ -245,6 +245,46 @@ SPECS = [
         story_role="typed tangent scale fine sweep",
         pending_job="14859503/14859504",
     ),
+    ResultSpec(
+        key="retrieval_residual_scale030_safe_types",
+        label="Train-state residual retrieval, scale 0.30, policy/no-op/wrong-gripper residuals",
+        path="h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_scale0p30_safe_types_summary.json",
+        clean_deployment="yes",
+        same_state_proposals="no",
+        expert_proposal="no",
+        story_role="typed tangent scale zoom sweep",
+        pending_job="14859589/14859590",
+    ),
+    ResultSpec(
+        key="retrieval_residual_scale0325_safe_types",
+        label="Train-state residual retrieval, scale 0.325, policy/no-op/wrong-gripper residuals",
+        path="h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_scale0p325_safe_types_summary.json",
+        clean_deployment="yes",
+        same_state_proposals="no",
+        expert_proposal="no",
+        story_role="typed tangent scale zoom sweep",
+        pending_job="14859591/14859592",
+    ),
+    ResultSpec(
+        key="retrieval_residual_scale0375_safe_types",
+        label="Train-state residual retrieval, scale 0.375, policy/no-op/wrong-gripper residuals",
+        path="h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_scale0p375_safe_types_summary.json",
+        clean_deployment="yes",
+        same_state_proposals="no",
+        expert_proposal="no",
+        story_role="typed tangent scale zoom sweep",
+        pending_job="14859593/14859594",
+    ),
+    ResultSpec(
+        key="retrieval_residual_scale040_safe_types",
+        label="Train-state residual retrieval, scale 0.40, policy/no-op/wrong-gripper residuals",
+        path="h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_scale0p40_safe_types_summary.json",
+        clean_deployment="yes",
+        same_state_proposals="no",
+        expert_proposal="no",
+        story_role="typed tangent scale zoom sweep",
+        pending_job="14859595/14859596",
+    ),
     ResultSpec(
         key="retrieval_residual_scale045_safe_types",
         label="Train-state residual retrieval, scale 0.45, policy/no-op/wrong-gripper residuals",