Auto-sync: 2026-06-28 12:50:49 (part 2)

results/h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_k4s035_safe_margin0p20_mean_by_type_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_k4s035_safe_margin0p20_mean_by_type.json",
+  "num_completed": 3,
+  "baseline_h4_policy_success": 0.2967,
+  "baseline_h16_rank_checkpoint_success": 0.29739130434782607,
+  "mean_success": 0.3472463768115942,
+  "std_success": 0.018757901086798177,
+  "mean_progress": 0.5647900497676259,
+  "mean_action_mse_to_best": 0.3953576808849323,
+  "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_k4s035_safe_margin0p20_mean_by_type.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": 0.35,
+      "retrieval_residual_anchor": "expert",
+      "retrieval_residual_reduce": "mean_by_type",
+      "policy_rollout_success_rate": 0.3391304347826087,
+      "policy_rollout_progress": 0.5509676206642357,
+      "oracle_success_rate": 0.8573913043478261,
+      "action_mse_to_best": 0.3818602439445322,
+      "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": 0.3025822116133685,
+          "expert_success_rate": 0.9375,
+          "num_groups": 208,
+          "oracle_success_rate": 0.9471153846153846,
+          "policy_expert_regret": 0.9611116516976975,
+          "policy_oracle_regret": 0.974534519124203,
+          "policy_rollout_progress": 0.6161549921290805,
+          "policy_rollout_success_rate": 0.34134615384615385,
+          "restore_max_error": 4.76837158203125e-07
+        },
+        "PullCube-v1": {
+          "action_mse_to_best": 0.6023750427094373,
+          "expert_success_rate": 0.19480519480519481,
+          "num_groups": 77,
+          "oracle_success_rate": 0.36363636363636365,
+          "policy_expert_regret": 0.2205158941018359,
+          "policy_oracle_regret": 0.392295054501745,
+          "policy_rollout_progress": 0.33419856826726396,
+          "policy_rollout_success_rate": 0.23376623376623376,
+          "restore_max_error": 4.76837158203125e-07
+        },
+        "PushCube-v1": {
+          "action_mse_to_best": 0.36585109930564746,
+          "expert_success_rate": 0.8617021276595744,
+          "num_groups": 94,
+          "oracle_success_rate": 0.9893617021276596,
+          "policy_expert_regret": 0.5719942746961371,
+          "policy_oracle_regret": 0.66845020953011,
+          "policy_rollout_progress": 0.682613620257124,
+          "policy_rollout_success_rate": 0.648936170212766,
+          "restore_max_error": 4.76837158203125e-07
+        },
+        "StackCube-v1": {
+          "action_mse_to_best": 0.45301999282468147,
+          "expert_success_rate": 0.6767676767676768,
+          "num_groups": 99,
+          "oracle_success_rate": 0.8585858585858586,
+          "policy_expert_regret": 0.9760173747334817,
+          "policy_oracle_regret": 1.1501769692909838,
+          "policy_rollout_progress": 0.43840954535537296,
+          "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_k4s035_safe_margin0p20_mean_by_type.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": 0.35,
+      "retrieval_residual_anchor": "expert",
+      "retrieval_residual_reduce": "mean_by_type",
+      "policy_rollout_success_rate": 0.3339130434782609,
+      "policy_rollout_progress": 0.5598738340061644,
+      "oracle_success_rate": 0.8695652173913043,
+      "action_mse_to_best": 0.38795768892311533,
+      "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": 0.3197257187743874,
+          "expert_success_rate": 0.9402173913043478,
+          "num_groups": 184,
+          "oracle_success_rate": 0.9456521739130435,
+          "policy_expert_regret": 1.0788406150557264,
+          "policy_oracle_regret": 1.0788646195306564,
+          "policy_rollout_progress": 0.5717960167777442,
+          "policy_rollout_success_rate": 0.28804347826086957,
+          "restore_max_error": 4.76837158203125e-07
+        },
+        "PullCube-v1": {
+          "action_mse_to_best": 0.5752323252197943,
+          "expert_success_rate": 0.25,
+          "num_groups": 76,
+          "oracle_success_rate": 0.40789473684210525,
+          "policy_expert_regret": 0.3254193017798427,
+          "policy_oracle_regret": 0.570340347647863,
+          "policy_rollout_progress": 0.26857212722595586,
+          "policy_rollout_success_rate": 0.15789473684210525,
+          "restore_max_error": 4.76837158203125e-07
+        },
+        "PushCube-v1": {
+          "action_mse_to_best": 0.35583734327742644,
+          "expert_success_rate": 0.8198198198198198,
+          "num_groups": 111,
+          "oracle_success_rate": 1.0,
+          "policy_expert_regret": 0.3476062170020095,
+          "policy_oracle_regret": 0.41423631895769825,
+          "policy_rollout_progress": 0.8019798972585179,
+          "policy_rollout_success_rate": 0.7837837837837838,
+          "restore_max_error": 4.76837158203125e-07
+        },
+        "StackCube-v1": {
+          "action_mse_to_best": 0.4861648879679186,
+          "expert_success_rate": 0.6923076923076923,
+          "num_groups": 91,
+          "oracle_success_rate": 0.8571428571428571,
+          "policy_expert_regret": 1.128277287050918,
+          "policy_oracle_regret": 1.3004250411803906,
+          "policy_rollout_progress": 0.39018028597910326,
+          "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_k4s035_safe_margin0p20_mean_by_type.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": 0.35,
+      "retrieval_residual_anchor": "expert",
+      "retrieval_residual_reduce": "mean_by_type",
+      "policy_rollout_success_rate": 0.36869565217391304,
+      "policy_rollout_progress": 0.5835286946324777,
+      "oracle_success_rate": 0.8765217391304347,
+      "action_mse_to_best": 0.41625510978714925,
+      "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": 0.35410993275096797,
+          "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": 0.3353001233190298,
+          "expert_success_rate": 0.9444444444444444,
+          "num_groups": 198,
+          "oracle_success_rate": 0.9595959595959596,
+          "policy_expert_regret": 0.995225173825718,
+          "policy_oracle_regret": 1.0023086243544264,
+          "policy_rollout_progress": 0.6313645500087678,
+          "policy_rollout_success_rate": 0.3282828282828283,
+          "restore_max_error": 4.76837158203125e-07
+        },
+        "PullCube-v1": {
+          "action_mse_to_best": 0.633242882291476,
+          "expert_success_rate": 0.24444444444444444,
+          "num_groups": 90,
+          "oracle_success_rate": 0.4666666666666667,
+          "policy_expert_regret": 0.2953753255169633,
+          "policy_oracle_regret": 0.5035596452722327,
+          "policy_rollout_progress": 0.33584247959295477,
+          "policy_rollout_success_rate": 0.23333333333333334,
+          "restore_max_error": 4.0978193283081055e-07
+        },
+        "PushCube-v1": {
+          "action_mse_to_best": 0.38076651996315114,
+          "expert_success_rate": 0.8514851485148515,
+          "num_groups": 101,
+          "oracle_success_rate": 1.0,
+          "policy_expert_regret": 0.3617186400264797,
+          "policy_oracle_regret": 0.4166218763825917,
+          "policy_rollout_progress": 0.8011999057956262,
+          "policy_rollout_success_rate": 0.7821782178217822,
+          "restore_max_error": 4.76837158203125e-07
+        },
+        "StackCube-v1": {
+          "action_mse_to_best": 0.48348258048709897,
+          "expert_success_rate": 0.7666666666666667,
+          "num_groups": 90,
+          "oracle_success_rate": 0.9111111111111111,
+          "policy_expert_regret": 1.1522358513540691,
+          "policy_oracle_regret": 1.3013080164790154,
+          "policy_rollout_progress": 0.4140402485926946,
+          "policy_rollout_success_rate": 0.16666666666666666,
+          "restore_max_error": 4.76837158203125e-07
+        }
+      }
+    }
+  ]
+}

results/h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_k4s035_safe_margin0p20_mean_by_type_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_k4s035_safe_margin0p20_mean_by_type.json`
+Completed seeds: 3
+Baseline h=4 policy success: 29.67%
+Baseline h=16 rank-checkpoint success: 29.74%
+
+Mean success: 34.72% +/- 1.88%
+Gain vs h=16 rank checkpoint: +4.99%
+Mean progress: 56.48%
+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 | mean_by_type | 0.00 | 0.35 | 0.200 | 0.00 | 0 | 0.00 | 33.91% | 55.10% | 85.74% | 0.382 |
+| 1 | retrieval_residual | 6 | no | 4 | raw | expert | mean_by_type | 0.00 | 0.35 | 0.200 | 0.00 | 0 | 0.00 | 33.39% | 55.99% | 86.96% | 0.388 |
+| 2 | retrieval_residual | 6 | no | 4 | raw | expert | mean_by_type | 0.00 | 0.35 | 0.200 | 0.00 | 0 | 0.00 | 36.87% | 58.35% | 87.65% | 0.416 |

results/h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_k4s040_safe_margin0p20_mean_by_type_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_safe_margin0p20_mean_by_type.json",
+  "num_completed": 3,
+  "baseline_h4_policy_success": 0.2967,
+  "baseline_h16_rank_checkpoint_success": 0.29739130434782607,
+  "mean_success": 0.34956521739130436,
+  "std_success": 0.018073573644196997,
+  "mean_progress": 0.5664916701836668,
+  "mean_action_mse_to_best": 0.3951832763526751,
+  "gain_vs_h4": 0.05286521739130434,
+  "gain_vs_h16_rank_checkpoint": 0.05217391304347829,
+  "rows": [
+    {
+      "seed": 0,
+      "path": "/scratch/knguy52/dovla/experiments/dovla_h16_policy_ckpt_runs/near_miss_policy_bc5/seed_0/policy_rollout_retrieval_residual_k4s040_safe_margin0p20_mean_by_type.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": "mean_by_type",
+      "policy_rollout_success_rate": 0.3391304347826087,
+      "policy_rollout_progress": 0.550848600073596,
+      "oracle_success_rate": 0.8573913043478261,
+      "action_mse_to_best": 0.3816084327896976,
+      "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": 0.30260570952668786,
+          "expert_success_rate": 0.9375,
+          "num_groups": 208,
+          "oracle_success_rate": 0.9471153846153846,
+          "policy_expert_regret": 0.9611076272379321,
+          "policy_oracle_regret": 0.9745304946644375,
+          "policy_rollout_progress": 0.6161590165888461,
+          "policy_rollout_success_rate": 0.34134615384615385,
+          "restore_max_error": 4.76837158203125e-07
+        },
+        "PullCube-v1": {
+          "action_mse_to_best": 0.6002476983643198,
+          "expert_success_rate": 0.19480519480519481,
+          "num_groups": 77,
+          "oracle_success_rate": 0.36363636363636365,
+          "policy_expert_regret": 0.22172368054471717,
+          "policy_oracle_regret": 0.3935028409446263,
+          "policy_rollout_progress": 0.3329907818243827,
+          "policy_rollout_success_rate": 0.23376623376623376,
+          "restore_max_error": 4.76837158203125e-07
+        },
+        "PushCube-v1": {
+          "action_mse_to_best": 0.36465501638644554,
+          "expert_success_rate": 0.8617021276595744,
+          "num_groups": 94,
+          "oracle_success_rate": 0.9893617021276596,
+          "policy_expert_regret": 0.5719942746961371,
+          "policy_oracle_regret": 0.6683293720509144,
+          "policy_rollout_progress": 0.6827344577363197,
+          "policy_rollout_success_rate": 0.648936170212766,
+          "restore_max_error": 4.76837158203125e-07
+        },
+        "StackCube-v1": {
+          "action_mse_to_best": 0.45429835978406247,
+          "expert_success_rate": 0.6767676767676768,
+          "num_groups": 99,
+          "oracle_success_rate": 0.8585858585858586,
+          "policy_expert_regret": 0.9758924564929924,
+          "policy_oracle_regret": 1.1500520510504944,
+          "policy_rollout_progress": 0.43853446359586234,
+          "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_safe_margin0p20_mean_by_type.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": "mean_by_type",
+      "policy_rollout_success_rate": 0.3391304347826087,
+      "policy_rollout_progress": 0.5633281786707433,
+      "oracle_success_rate": 0.8695652173913043,
+      "action_mse_to_best": 0.38771247025741185,
+      "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": 0.31969433684792853,
+          "expert_success_rate": 0.9402173913043478,
+          "num_groups": 184,
+          "oracle_success_rate": 0.9456521739130435,
+          "policy_expert_regret": 1.0709968252087254,
+          "policy_oracle_regret": 1.0710208296836556,
+          "policy_rollout_progress": 0.5742050240160492,
+          "policy_rollout_success_rate": 0.29347826086956524,
+          "restore_max_error": 4.76837158203125e-07
+        },
+        "PullCube-v1": {
+          "action_mse_to_best": 0.5715771013576734,
+          "expert_success_rate": 0.25,
+          "num_groups": 76,
+          "oracle_success_rate": 0.40789473684210525,
+          "policy_expert_regret": 0.3008137478640205,
+          "policy_oracle_regret": 0.5236700972248065,
+          "policy_rollout_progress": 0.2889265881753282,
+          "policy_rollout_success_rate": 0.18421052631578946,
+          "restore_max_error": 4.76837158203125e-07
+        },
+        "PushCube-v1": {
+          "action_mse_to_best": 0.35610941270584456,
+          "expert_success_rate": 0.8198198198198198,
+          "num_groups": 111,
+          "oracle_success_rate": 1.0,
+          "policy_expert_regret": 0.3476062170020095,
+          "policy_oracle_regret": 0.4142807445547602,
+          "policy_rollout_progress": 0.801935471661456,
+          "policy_rollout_success_rate": 0.7837837837837838,
+          "restore_max_error": 4.76837158203125e-07
+        },
+        "StackCube-v1": {
+          "action_mse_to_best": 0.4873997324586406,
+          "expert_success_rate": 0.6923076923076923,
+          "num_groups": 91,
+          "oracle_success_rate": 0.8571428571428571,
+          "policy_expert_regret": 1.1282664842002994,
+          "policy_oracle_regret": 1.300414238329772,
+          "policy_rollout_progress": 0.3901910888297217,
+          "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_safe_margin0p20_mean_by_type.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": "mean_by_type",
+      "policy_rollout_success_rate": 0.37043478260869567,
+      "policy_rollout_progress": 0.5852982318066611,
+      "oracle_success_rate": 0.8765217391304347,
+      "action_mse_to_best": 0.41622892601091577,
+      "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": 0.3542572152073262,
+          "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": 0.33534634657053636,
+          "expert_success_rate": 0.9444444444444444,
+          "num_groups": 198,
+          "oracle_success_rate": 0.9595959595959596,
+          "policy_expert_regret": 0.9878635107153895,
+          "policy_oracle_regret": 0.9949469612440979,
+          "policy_rollout_progress": 0.6336757080685912,
+          "policy_rollout_success_rate": 0.3333333333333333,
+          "restore_max_error": 4.76837158203125e-07
+        },
+        "PullCube-v1": {
+          "action_mse_to_best": 0.6283310342166158,
+          "expert_success_rate": 0.24444444444444444,
+          "num_groups": 90,
+          "oracle_success_rate": 0.4666666666666667,
+          "policy_expert_regret": 0.2927713212729537,
+          "policy_oracle_regret": 0.5009592348571661,
+          "policy_rollout_progress": 0.3384430423310025,
+          "policy_rollout_success_rate": 0.23333333333333334,
+          "restore_max_error": 4.0978193283081055e-07
+        },
+        "PushCube-v1": {
+          "action_mse_to_best": 0.38386758350499783,
+          "expert_success_rate": 0.8514851485148515,
+          "num_groups": 101,
+          "oracle_success_rate": 1.0,
+          "policy_expert_regret": 0.3619124246115732,
+          "policy_oracle_regret": 0.4168156609676852,
+          "policy_rollout_progress": 0.8010061212105326,
+          "policy_rollout_success_rate": 0.7821782178217822,
+          "restore_max_error": 4.76837158203125e-07
+        },
+        "StackCube-v1": {
+          "action_mse_to_best": 0.48448826846563153,
+          "expert_success_rate": 0.7666666666666667,
+          "num_groups": 90,
+          "oracle_success_rate": 0.9111111111111111,
+          "policy_expert_regret": 1.1483981160653962,
+          "policy_oracle_regret": 1.2974702811903425,
+          "policy_rollout_progress": 0.4178779838813676,
+          "policy_rollout_success_rate": 0.16666666666666666,
+          "restore_max_error": 4.76837158203125e-07
+        }
+      }
+    }
+  ]
+}

results/h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_k4s040_safe_margin0p20_mean_by_type_summary.md

@@ -0,0 +1,19 @@
+# h=16 Best-Policy Checkpoint Rollout
+
+Run root: `/scratch/knguy52/dovla/experiments/dovla_h16_policy_ckpt_runs`
+Objective: `near_miss_policy_bc5`
+Result file: `policy_rollout_retrieval_residual_k4s040_safe_margin0p20_mean_by_type.json`
+Completed seeds: 3
+Baseline h=4 policy success: 29.67%
+Baseline h=16 rank-checkpoint success: 29.74%
+
+Mean success: 34.96% +/- 1.81%
+Gain vs h=16 rank checkpoint: +5.22%
+Mean progress: 56.65%
+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 | mean_by_type | 0.00 | 0.40 | 0.200 | 0.00 | 0 | 0.00 | 33.91% | 55.08% | 85.74% | 0.382 |
+| 1 | retrieval_residual | 6 | no | 4 | raw | expert | mean_by_type | 0.00 | 0.40 | 0.200 | 0.00 | 0 | 0.00 | 33.91% | 56.33% | 86.96% | 0.388 |
+| 2 | retrieval_residual | 6 | no | 4 | raw | expert | mean_by_type | 0.00 | 0.40 | 0.200 | 0.00 | 0 | 0.00 | 37.04% | 58.53% | 87.65% | 0.416 |

results/paper_core_results.md

@@ -32,6 +32,7 @@ baseline is the h=16 rank-checkpoint online rollout (`29.74%`).
 | Train-state residual retrieval, policy/no-op/wrong-gripper, scale 0.35 | No | No | 33.74% | +4.00 pp | Typed tangent transport before abstention |
 | Train-state residual retrieval, safe residuals + advantage margin 0.20 | No | No | 34.84% | +5.10 pp | Abstains unless field advantage beats policy |
 | K2 train-state residual retrieval, safe residuals + advantage margin 0.20 | No | No | 35.01% | +5.28 pp | Current best deployment-clean diagnostic; abstention makes a small train-neighborhood useful |
+| K4 train-state residual retrieval, safe residuals + mean-by-type tangent consensus | No | No | 34.96% | +5.22 pp | Near-tie clean diagnostic; consensus denoising does not beat raw K2 residuals |
 | Policy-relative residual anchor, safe residuals | No | No | 33.74% | +4.00 pp | Policy-relative anchoring ties but does not improve expert-relative residuals |
 | Train-state residual retrieval, z-score metric | No | No | 32.23% | +2.49 pp | State normalization hurts nearest tangent retrieval here |
 | Train-state residual retrieval, z-score metric + anti-goal mask | No | No | 32.75% | +3.01 pp | Masking helps z-score but remains below raw |
@@ -45,6 +46,7 @@ baseline is the h=16 rank-checkpoint online rollout (`29.74%`).
 | Lattice, no expert/no near-miss | Yes | No | 25.57% | -4.17 pp | Non-local negatives do not help |
 | Lattice, near-miss only | Yes | No | 55.94% | +26.20 pp | Local counterfactual proposals carry the gain |
 | Lattice, no expert | Yes | No | 56.99% | +27.25 pp | Reviewer-safe main result |
+| Lattice, no expert + policy baseline candidate | Yes | No | 40.70% | +10.96 pp | Policy fallback collapses same-state selection; proposal geometry is the mechanism |
 | Lattice, full | Yes | Yes | 69.33% | +39.59 pp | Upper deployment result with expert proposal |
 | Oracle ceiling | Yes | Yes | 86.78% | +57.04 pp | Remaining headroom |
 
@@ -62,12 +64,14 @@ Suggested main-table rows:
 10. Train-state residual retrieval, typed safe families at scale 0.35
 11. Train-state residual retrieval, typed safe families + advantage margin 0.20
 12. K2 train-state residual retrieval, typed safe families + advantage margin 0.20
-13. Residual-tangent distillation policy
-14. Residual+Gaussian hybrid, K32 sigma0.35
-15. Lattice, near-miss only
-16. Lattice, no expert
-17. Lattice, full
-18. Oracle ceiling
+13. K4 train-state residual retrieval, mean-by-type tangent consensus
+14. Residual-tangent distillation policy
+15. Residual+Gaussian hybrid, K32 sigma0.35
+16. Lattice, near-miss only
+17. Lattice, no expert
+18. Lattice, no expert + policy baseline candidate
+19. Lattice, full
+20. Oracle ceiling
 
 Suggested claim:
 
@@ -75,7 +79,8 @@ Suggested claim:
 > selection rule. Deployment-clean K2 typed counterfactual residual transport with advantage
 > abstention gives the strongest clean gain so far, while ungated KNN residual
 > retrieval, field-gradient ascent, broader non-expert BC targets, field-teacher/tangent distillation, z-score retrieval,
-> train-family reliability priors, policy-relative anchoring, and residual+Gaussian hybrids fail.
+> train-family reliability priors, policy-relative anchoring, residual+Gaussian hybrids,
+> tangent consensus, and same-state policy-baseline fallback fail to improve the main rows.
 > The large effect appears only when the field is queried on
 > same-state intervention proposals, and the mechanism is isolated to local near-miss
 > counterfactual geometry.

results/paper_story_memo.md

@@ -26,6 +26,8 @@ when queried on proposal geometry that matches those local counterfactuals.
 | All-split field-teacher distillation does not fix checkpointing/coverage | allmap direct is 28.00%; field-guided best is 26.49% despite 100% target coverage | Negative diagnostic |
 | Residual family consistency improves clean transport | policy/no-op/wrong-gripper typed residuals reach 33.74%, above raw 33.33% | Supported as diagnostic |
 | Counterfactual advantage abstention improves clean transport | requiring field advantage over the zero-residual policy raises typed residual transport to 34.84%, and K2 retrieval reaches 35.01% | Current best clean result |
+| Tangent consensus is close but does not beat raw K2 residuals | K4 mean-by-type residual consensus reaches 34.96%, just below the 35.01% K2 raw residual row | Near-tie diagnostic |
+| Policy fallback is not the same-state mechanism | adding a policy baseline candidate to the no-expert same-state lattice drops 56.99% to 40.70% even with margin 0.00 | Negative diagnostic |
 | Z-score retrieval metric does not help | z-score rows reach 32.23-32.81%, below raw retrieval | Negative diagnostic |
 | Train-split residual family reliability does not recover the typed mask | after fixing threshold pass-through, scale-0.35 thresholds 0.10/0.25 reach 33.33%/33.28%, below typed safe residuals | Negative diagnostic |
 | Residual-tangent distillation does not solve clean proposal generation | aligned allmap tangent student reaches 28.87% despite low pseudo-target BC loss | Negative diagnostic |
@@ -50,14 +52,16 @@ clean proposal result, the intended main rows are:
 11. Train-state residual retrieval, typed safe families: 33.74%
 12. Train-state residual retrieval, typed safe families + advantage margin: 34.84%
 13. K2 train-state residual retrieval, typed safe families + advantage margin: 35.01%
-14. Residual-tangent distillation policy: 28.87%
-15. Z-score residual retrieval: 32.23-32.81%
-16. Train-family reliability prior: 33.28-33.33%
-17. Residual+Gaussian hybrid K32/K64: 31.30% / 30.90%
-18. Lattice, near-miss only: 55.94%
-19. Lattice, no expert: 56.99%
-20. Lattice, full: 69.33%
-21. Oracle ceiling: 86.78%
+14. K4 mean-by-type tangent consensus: 34.96%
+15. Residual-tangent distillation policy: 28.87%
+16. Z-score residual retrieval: 32.23-32.81%
+17. Train-family reliability prior: 33.28-33.33%
+18. Residual+Gaussian hybrid K32/K64: 31.30% / 30.90%
+19. Lattice, near-miss only: 55.94%
+20. Lattice, no expert: 56.99%
+21. Lattice, no expert + policy baseline candidate: 40.70%
+22. Lattice, full: 69.33%
+23. Oracle ceiling: 86.78%
 
 ## Novelty Framing
 
@@ -85,7 +89,7 @@ test-time search. The cleaner novelty is:
 
 ## Job Status
 
-Last checked: `2026-06-28 12:25 UTC`. No DoVLA jobs are currently queued.
+Last checked: `2026-06-28 16:46 UTC`. No DoVLA jobs are currently queued.
 
 - `14858328`-`14858333`: completed train-split `field_selected_noexpert_bc5`;
   direct rollout is 26.84%, field-guided best is 27.65%.
@@ -125,6 +129,13 @@ Last checked: `2026-06-28 12:25 UTC`. No DoVLA jobs are currently queued.
 - `14862857`-`14862939`: completed KNN-with-abstention sweeps. K2 residual
   retrieval at scale `0.40`, margin `0.20` is the current best clean row:
   35.01% mean success (+5.28 pp vs h=16).
+- `14868661`-`14868668`: completed same-state no-expert lattice with a prepended
+  policy baseline candidate. The best setting, margin `0.00`, reaches only
+  40.70%, far below the no-expert lattice's 56.99%; policy fallback should be
+  framed as a negative diagnostic.
+- `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.
 
 ## Decision Notes
 
@@ -133,5 +144,6 @@ Last checked: `2026-06-28 12:25 UTC`. No DoVLA jobs are currently queued.
 - Use K2 typed safe residual transport with advantage abstention (35.01%) only as the current best clean
   deployment diagnostic, not as a SOTA claim.
 - Treat z-score retrieval, repaired train-family reliability priors, Gaussian hybrids,
-  field optimization, field-teacher/tangent distillation, and policy-relative anchoring as negative diagnostics
+  field optimization, field-teacher/tangent distillation, policy-relative anchoring, tangent consensus,
+  and same-state policy-baseline fallback as negative or near-tie diagnostics
   that sharpen the story around local counterfactual proposal geometry.