Auto-sync: 2026-06-28 08:27:39 (part 2)

results/h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_scale0p35_safe_types_margin0p18_summary.json

@@ -0,0 +1,292 @@
+{
+  "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_margin0p18.json",
+  "num_completed": 3,
+  "baseline_h4_policy_success": 0.2967,
+  "baseline_h16_rank_checkpoint_success": 0.29739130434782607,
+  "mean_success": 0.3466666666666667,
+  "std_success": 0.01479113138802674,
+  "mean_progress": 0.5636368875724271,
+  "mean_action_mse_to_best": 0.396542789398868,
+  "gain_vs_h4": 0.04996666666666666,
+  "gain_vs_h16_rank_checkpoint": 0.04927536231884061,
+  "rows": [
+    {
+      "seed": 0,
+      "path": "/scratch/knguy52/dovla/experiments/dovla_h16_policy_ckpt_runs/near_miss_policy_bc5/seed_0/policy_rollout_retrieval_residual_scale0p35_safe_types_margin0p18.json",
+      "num_groups": 575,
+      "selection_mode": "retrieval_residual",
+      "num_candidates": 16,
+      "candidate_sigma": 0.0,
+      "selection_margin": 0.18,
+      "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,
+      "retrieval_residual_anchor": "expert",
+      "policy_rollout_success_rate": 0.3408695652173913,
+      "policy_rollout_progress": 0.5523332587479736,
+      "oracle_success_rate": 0.8573913043478261,
+      "action_mse_to_best": 0.38042863945031297,
+      "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": 0.3027281228524561,
+          "expert_success_rate": 0.9375,
+          "num_groups": 208,
+          "oracle_success_rate": 0.9471153846153846,
+          "policy_expert_regret": 0.9681357126324796,
+          "policy_oracle_regret": 0.981558580058985,
+          "policy_rollout_progress": 0.6139386235019908,
+          "policy_rollout_success_rate": 0.33653846153846156,
+          "restore_max_error": 4.76837158203125e-07
+        },
+        "PullCube-v1": {
+          "action_mse_to_best": 0.5987401300823534,
+          "expert_success_rate": 0.19480519480519481,
+          "num_groups": 77,
+          "oracle_success_rate": 0.36363636363636365,
+          "policy_expert_regret": 0.19660904705101115,
+          "policy_oracle_regret": 0.3685447470520853,
+          "policy_rollout_progress": 0.3449618627299107,
+          "policy_rollout_success_rate": 0.24675324675324675,
+          "restore_max_error": 4.76837158203125e-07
+        },
+        "PushCube-v1": {
+          "action_mse_to_best": 0.36554105338105497,
+          "expert_success_rate": 0.8617021276595744,
+          "num_groups": 94,
+          "oracle_success_rate": 0.9893617021276596,
+          "policy_expert_regret": 0.5727000404545601,
+          "policy_oracle_regret": 0.669589711630598,
+          "policy_rollout_progress": 0.681474118156636,
+          "policy_rollout_success_rate": 0.648936170212766,
+          "restore_max_error": 4.76837158203125e-07
+        },
+        "StackCube-v1": {
+          "action_mse_to_best": 0.4475200985548924,
+          "expert_success_rate": 0.6767676767676768,
+          "num_groups": 99,
+          "oracle_success_rate": 0.8585858585858586,
+          "policy_expert_regret": 0.9721011399018644,
+          "policy_oracle_regret": 1.1347771097313275,
+          "policy_rollout_progress": 0.44370839481401925,
+          "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_margin0p18.json",
+      "num_groups": 575,
+      "selection_mode": "retrieval_residual",
+      "num_candidates": 16,
+      "candidate_sigma": 0.0,
+      "selection_margin": 0.18,
+      "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,
+      "retrieval_residual_anchor": "expert",
+      "policy_rollout_success_rate": 0.3356521739130435,
+      "policy_rollout_progress": 0.5607220564134743,
+      "oracle_success_rate": 0.8695652173913043,
+      "action_mse_to_best": 0.388851607432184,
+      "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": 0.32021380088332557,
+          "expert_success_rate": 0.9402173913043478,
+          "num_groups": 184,
+          "oracle_success_rate": 0.9456521739130435,
+          "policy_expert_regret": 1.0790262568296622,
+          "policy_oracle_regret": 1.0790502613045923,
+          "policy_rollout_progress": 0.5716103750038082,
+          "policy_rollout_success_rate": 0.28804347826086957,
+          "restore_max_error": 4.76837158203125e-07
+        },
+        "PullCube-v1": {
+          "action_mse_to_best": 0.5809172638937047,
+          "expert_success_rate": 0.25,
+          "num_groups": 76,
+          "oracle_success_rate": 0.40789473684210525,
+          "policy_expert_regret": 0.277883501037171,
+          "policy_oracle_regret": 0.5505282147542426,
+          "policy_rollout_progress": 0.2752263653827341,
+          "policy_rollout_success_rate": 0.17105263157894737,
+          "restore_max_error": 4.76837158203125e-07
+        },
+        "PushCube-v1": {
+          "action_mse_to_best": 0.35491145425685894,
+          "expert_success_rate": 0.8198198198198198,
+          "num_groups": 111,
+          "oracle_success_rate": 1.0,
+          "policy_expert_regret": 0.3476062170020095,
+          "policy_oracle_regret": 0.4141344331406258,
+          "policy_rollout_progress": 0.8020817830755904,
+          "policy_rollout_success_rate": 0.7837837837837838,
+          "restore_max_error": 4.76837158203125e-07
+        },
+        "StackCube-v1": {
+          "action_mse_to_best": 0.4872079030968822,
+          "expert_success_rate": 0.6923076923076923,
+          "num_groups": 91,
+          "oracle_success_rate": 0.8571428571428571,
+          "policy_expert_regret": 1.1282239404025969,
+          "policy_oracle_regret": 1.3003716945320696,
+          "policy_rollout_progress": 0.3902336326274243,
+          "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_margin0p18.json",
+      "num_groups": 575,
+      "selection_mode": "retrieval_residual",
+      "num_candidates": 16,
+      "candidate_sigma": 0.0,
+      "selection_margin": 0.18,
+      "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,
+      "retrieval_residual_anchor": "expert",
+      "policy_rollout_success_rate": 0.3634782608695652,
+      "policy_rollout_progress": 0.5778553475558337,
+      "oracle_success_rate": 0.8765217391304347,
+      "action_mse_to_best": 0.4203481213141071,
+      "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": 0.35413461838713073,
+          "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": 0.3366299307519438,
+          "expert_success_rate": 0.9444444444444444,
+          "num_groups": 198,
+          "oracle_success_rate": 0.9595959595959596,
+          "policy_expert_regret": 0.9904651472702471,
+          "policy_oracle_regret": 0.9975485977989555,
+          "policy_rollout_progress": 0.6289677324733047,
+          "policy_rollout_success_rate": 0.3282828282828283,
+          "restore_max_error": 4.76837158203125e-07
+        },
+        "PullCube-v1": {
+          "action_mse_to_best": 0.6534511918822924,
+          "expert_success_rate": 0.24444444444444444,
+          "num_groups": 90,
+          "oracle_success_rate": 0.4666666666666667,
+          "policy_expert_regret": 0.33460306424508923,
+          "policy_oracle_regret": 0.541839225928258,
+          "policy_rollout_progress": 0.319785273482133,
+          "policy_rollout_success_rate": 0.2111111111111111,
+          "restore_max_error": 4.0978193283081055e-07
+        },
+        "PushCube-v1": {
+          "action_mse_to_best": 0.3817486162527953,
+          "expert_success_rate": 0.8514851485148515,
+          "num_groups": 101,
+          "oracle_success_rate": 1.0,
+          "policy_expert_regret": 0.3810096250312163,
+          "policy_oracle_regret": 0.4359128613873283,
+          "policy_rollout_progress": 0.7918099108898994,
+          "policy_rollout_success_rate": 0.7722772277227723,
+          "restore_max_error": 4.76837158203125e-07
+        },
+        "StackCube-v1": {
+          "action_mse_to_best": 0.48537002878470553,
+          "expert_success_rate": 0.7666666666666667,
+          "num_groups": 90,
+          "oracle_success_rate": 0.9111111111111111,
+          "policy_expert_regret": 1.1566143698162503,
+          "policy_oracle_regret": 1.3056865349411964,
+          "policy_rollout_progress": 0.4096617301305135,
+          "policy_rollout_success_rate": 0.16666666666666666,
+          "restore_max_error": 4.76837158203125e-07
+        }
+      }
+    }
+  ]
+}

results/h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_scale0p35_safe_types_margin0p18_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_margin0p18.json`
+Completed seeds: 3
+Baseline h=4 policy success: 29.67%
+Baseline h=16 rank-checkpoint success: 29.74%
+
+Mean success: 34.67% +/- 1.48%
+Gain vs h=16 rank checkpoint: +4.93%
+Mean progress: 56.36%
+Mean action MSE to best: 0.397
+
+| seed | mode | k | retrieval K | retrieval metric | residual anchor | min type success | residual scale | margin | sigma | opt steps | trust | success | progress | oracle | action MSE |
+|---:|---|---:|---:|---|---|---:|---:|---:|---:|---:|---:|---:|---:|---:|---:|
+| 0 | retrieval_residual | 16 | 1 | raw | expert | 0.00 | 0.35 | 0.180 | 0.00 | 0 | 0.00 | 34.09% | 55.23% | 85.74% | 0.380 |
+| 1 | retrieval_residual | 16 | 1 | raw | expert | 0.00 | 0.35 | 0.180 | 0.00 | 0 | 0.00 | 33.57% | 56.07% | 86.96% | 0.389 |
+| 2 | retrieval_residual | 16 | 1 | raw | expert | 0.00 | 0.35 | 0.180 | 0.00 | 0 | 0.00 | 36.35% | 57.79% | 87.65% | 0.420 |

results/h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_scale0p35_safe_types_margin0p22_summary.json

@@ -0,0 +1,292 @@
+{
+  "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_margin0p22.json",
+  "num_completed": 3,
+  "baseline_h4_policy_success": 0.2967,
+  "baseline_h16_rank_checkpoint_success": 0.29739130434782607,
+  "mean_success": 0.34840579710144925,
+  "std_success": 0.017593032933905524,
+  "mean_progress": 0.5655779171832683,
+  "mean_action_mse_to_best": 0.39568004657392913,
+  "gain_vs_h4": 0.051705797101449236,
+  "gain_vs_h16_rank_checkpoint": 0.051014492753623186,
+  "rows": [
+    {
+      "seed": 0,
+      "path": "/scratch/knguy52/dovla/experiments/dovla_h16_policy_ckpt_runs/near_miss_policy_bc5/seed_0/policy_rollout_retrieval_residual_scale0p35_safe_types_margin0p22.json",
+      "num_groups": 575,
+      "selection_mode": "retrieval_residual",
+      "num_candidates": 16,
+      "candidate_sigma": 0.0,
+      "selection_margin": 0.22,
+      "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,
+      "retrieval_residual_anchor": "expert",
+      "policy_rollout_success_rate": 0.3391304347826087,
+      "policy_rollout_progress": 0.5522342324463407,
+      "oracle_success_rate": 0.8573913043478261,
+      "action_mse_to_best": 0.380270165618833,
+      "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": 0.30268779866254103,
+          "expert_success_rate": 0.9375,
+          "num_groups": 208,
+          "oracle_success_rate": 0.9471153846153846,
+          "policy_expert_regret": 0.9681423121275237,
+          "policy_oracle_regret": 0.9815651795540291,
+          "policy_rollout_progress": 0.6139320240069467,
+          "policy_rollout_success_rate": 0.33653846153846156,
+          "restore_max_error": 4.76837158203125e-07
+        },
+        "PullCube-v1": {
+          "action_mse_to_best": 0.5987401300823534,
+          "expert_success_rate": 0.19480519480519481,
+          "num_groups": 77,
+          "oracle_success_rate": 0.36363636363636365,
+          "policy_expert_regret": 0.19660904705101115,
+          "policy_oracle_regret": 0.3685447470520853,
+          "policy_rollout_progress": 0.3449618627299107,
+          "policy_rollout_success_rate": 0.24675324675324675,
+          "restore_max_error": 4.76837158203125e-07
+        },
+        "PushCube-v1": {
+          "action_mse_to_best": 0.36658212189820216,
+          "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": 0.4456959019669078,
+          "expert_success_rate": 0.6767676767676768,
+          "num_groups": 99,
+          "oracle_success_rate": 0.8585858585858586,
+          "policy_expert_regret": 0.9723617639204468,
+          "policy_oracle_regret": 1.1465213584779488,
+          "policy_rollout_progress": 0.44206515616840786,
+          "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_scale0p35_safe_types_margin0p22.json",
+      "num_groups": 575,
+      "selection_mode": "retrieval_residual",
+      "num_candidates": 16,
+      "candidate_sigma": 0.0,
+      "selection_margin": 0.22,
+      "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,
+      "retrieval_residual_anchor": "expert",
+      "policy_rollout_success_rate": 0.3373913043478261,
+      "policy_rollout_progress": 0.5622359952136227,
+      "oracle_success_rate": 0.8695652173913043,
+      "action_mse_to_best": 0.38927455800059046,
+      "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": 0.3201620632418148,
+          "expert_success_rate": 0.9402173913043478,
+          "num_groups": 184,
+          "oracle_success_rate": 0.9456521739130435,
+          "policy_expert_regret": 1.0790224476393475,
+          "policy_oracle_regret": 1.0790464521142775,
+          "policy_rollout_progress": 0.571614184194123,
+          "policy_rollout_success_rate": 0.28804347826086957,
+          "restore_max_error": 4.76837158203125e-07
+        },
+        "PullCube-v1": {
+          "action_mse_to_best": 0.5809172638937047,
+          "expert_success_rate": 0.25,
+          "num_groups": 76,
+          "oracle_success_rate": 0.40789473684210525,
+          "policy_expert_regret": 0.277883501037171,
+          "policy_oracle_regret": 0.5505282147542426,
+          "policy_rollout_progress": 0.2752263653827341,
+          "policy_rollout_success_rate": 0.17105263157894737,
+          "restore_max_error": 4.76837158203125e-07
+        },
+        "PushCube-v1": {
+          "action_mse_to_best": 0.3571881777061535,
+          "expert_success_rate": 0.8198198198198198,
+          "num_groups": 111,
+          "oracle_success_rate": 1.0,
+          "policy_expert_regret": 0.3476113024595621,
+          "policy_oracle_regret": 0.39728926269857734,
+          "policy_rollout_progress": 0.8099179445086299,
+          "policy_rollout_success_rate": 0.7927927927927928,
+          "restore_max_error": 4.76837158203125e-07
+        },
+        "StackCube-v1": {
+          "action_mse_to_best": 0.4872079030968822,
+          "expert_success_rate": 0.6923076923076923,
+          "num_groups": 91,
+          "oracle_success_rate": 0.8571428571428571,
+          "policy_expert_regret": 1.1282239404025969,
+          "policy_oracle_regret": 1.3003716945320696,
+          "policy_rollout_progress": 0.3902336326274243,
+          "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_margin0p22.json",
+      "num_groups": 575,
+      "selection_mode": "retrieval_residual",
+      "num_candidates": 16,
+      "candidate_sigma": 0.0,
+      "selection_margin": 0.22,
+      "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,
+      "retrieval_residual_anchor": "expert",
+      "policy_rollout_success_rate": 0.36869565217391304,
+      "policy_rollout_progress": 0.5822635238898416,
+      "oracle_success_rate": 0.8765217391304347,
+      "action_mse_to_best": 0.417495416102364,
+      "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": 0.35361564260771655,
+          "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": 0.33572704106042484,
+          "expert_success_rate": 0.9444444444444444,
+          "num_groups": 198,
+          "oracle_success_rate": 0.9595959595959596,
+          "policy_expert_regret": 0.9903708135878498,
+          "policy_oracle_regret": 0.9974542641165581,
+          "policy_rollout_progress": 0.6290620661557021,
+          "policy_rollout_success_rate": 0.3282828282828283,
+          "restore_max_error": 4.76837158203125e-07
+        },
+        "PullCube-v1": {
+          "action_mse_to_best": 0.6387406456801626,
+          "expert_success_rate": 0.24444444444444444,
+          "num_groups": 90,
+          "oracle_success_rate": 0.4666666666666667,
+          "policy_expert_regret": 0.29494567263732785,
+          "policy_oracle_regret": 0.5021084638174418,
+          "policy_rollout_progress": 0.33729381337072684,
+          "policy_rollout_success_rate": 0.23333333333333334,
+          "restore_max_error": 4.0978193283081055e-07
+        },
+        "PushCube-v1": {
+          "action_mse_to_best": 0.38087968056154725,
+          "expert_success_rate": 0.8514851485148515,
+          "num_groups": 101,
+          "oracle_success_rate": 1.0,
+          "policy_expert_regret": 0.36179918213055867,
+          "policy_oracle_regret": 0.41670241848667067,
+          "policy_rollout_progress": 0.8011193636915471,
+          "policy_rollout_success_rate": 0.7821782178217822,
+          "restore_max_error": 4.76837158203125e-07
+        },
+        "StackCube-v1": {
+          "action_mse_to_best": 0.48537002878470553,
+          "expert_success_rate": 0.7666666666666667,
+          "num_groups": 90,
+          "oracle_success_rate": 0.9111111111111111,
+          "policy_expert_regret": 1.1566143698162503,
+          "policy_oracle_regret": 1.3056865349411964,
+          "policy_rollout_progress": 0.4096617301305135,
+          "policy_rollout_success_rate": 0.16666666666666666,
+          "restore_max_error": 4.76837158203125e-07
+        }
+      }
+    }
+  ]
+}

results/h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_scale0p35_safe_types_margin0p22_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_margin0p22.json`
+Completed seeds: 3
+Baseline h=4 policy success: 29.67%
+Baseline h=16 rank-checkpoint success: 29.74%
+
+Mean success: 34.84% +/- 1.76%
+Gain vs h=16 rank checkpoint: +5.10%
+Mean progress: 56.56%
+Mean action MSE to best: 0.396
+
+| seed | mode | k | retrieval K | retrieval metric | residual anchor | min type success | residual scale | margin | sigma | opt steps | trust | success | progress | oracle | action MSE |
+|---:|---|---:|---:|---|---|---:|---:|---:|---:|---:|---:|---:|---:|---:|---:|
+| 0 | retrieval_residual | 16 | 1 | raw | expert | 0.00 | 0.35 | 0.220 | 0.00 | 0 | 0.00 | 33.91% | 55.22% | 85.74% | 0.380 |
+| 1 | retrieval_residual | 16 | 1 | raw | expert | 0.00 | 0.35 | 0.220 | 0.00 | 0 | 0.00 | 33.74% | 56.22% | 86.96% | 0.389 |
+| 2 | retrieval_residual | 16 | 1 | raw | expert | 0.00 | 0.35 | 0.220 | 0.00 | 0 | 0.00 | 36.87% | 58.23% | 87.65% | 0.417 |

results/h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_scale0p35_safe_types_margin0p25_summary.json

@@ -0,0 +1,292 @@
+{
+  "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_margin0p25.json",
+  "num_completed": 3,
+  "baseline_h4_policy_success": 0.2967,
+  "baseline_h16_rank_checkpoint_success": 0.29739130434782607,
+  "mean_success": 0.3455072463768116,
+  "std_success": 0.017157853434143458,
+  "mean_progress": 0.5634927251384764,
+  "mean_action_mse_to_best": 0.39447076573751977,
+  "gain_vs_h4": 0.04880724637681161,
+  "gain_vs_h16_rank_checkpoint": 0.04811594202898556,
+  "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_margin0p25.json",
+      "num_groups": 575,
+      "selection_mode": "retrieval_residual",
+      "num_candidates": 16,
+      "candidate_sigma": 0.0,
+      "selection_margin": 0.25,
+      "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,
+      "retrieval_residual_anchor": "expert",
+      "policy_rollout_success_rate": 0.3373913043478261,
+      "policy_rollout_progress": 0.5501761496791859,
+      "oracle_success_rate": 0.8573913043478261,
+      "action_mse_to_best": 0.3798463192703607,
+      "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": 0.30271882717855847,
+          "expert_success_rate": 0.9375,
+          "num_groups": 208,
+          "oracle_success_rate": 0.9471153846153846,
+          "policy_expert_regret": 0.9681180751022811,
+          "policy_oracle_regret": 0.9815409425287865,
+          "policy_rollout_progress": 0.6139562610321894,
+          "policy_rollout_success_rate": 0.33653846153846156,
+          "restore_max_error": 4.76837158203125e-07
+        },
+        "PullCube-v1": {
+          "action_mse_to_best": 0.600519543924889,
+          "expert_success_rate": 0.19480519480519481,
+          "num_groups": 77,
+          "oracle_success_rate": 0.36363636363636365,
+          "policy_expert_regret": 0.19700447897209072,
+          "policy_oracle_regret": 0.39191550494461963,
+          "policy_rollout_progress": 0.33457811782438934,
+          "policy_rollout_success_rate": 0.23376623376623376,
+          "restore_max_error": 4.76837158203125e-07
+        },
+        "PushCube-v1": {
+          "action_mse_to_best": 0.3656082003912393,
+          "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": 0.4427097219070702,
+          "expert_success_rate": 0.6767676767676768,
+          "num_groups": 99,
+          "oracle_success_rate": 0.8585858585858586,
+          "policy_expert_regret": 0.9762899511390262,
+          "policy_oracle_regret": 1.1504495456965282,
+          "policy_rollout_progress": 0.4381369689498285,
+          "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_scale0p35_safe_types_margin0p25.json",
+      "num_groups": 575,
+      "selection_mode": "retrieval_residual",
+      "num_candidates": 16,
+      "candidate_sigma": 0.0,
+      "selection_margin": 0.25,
+      "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,
+      "retrieval_residual_anchor": "expert",
+      "policy_rollout_success_rate": 0.3339130434782609,
+      "policy_rollout_progress": 0.559359166495826,
+      "oracle_success_rate": 0.8695652173913043,
+      "action_mse_to_best": 0.38780134577142156,
+      "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": 0.3198142160418565,
+          "expert_success_rate": 0.9402173913043478,
+          "num_groups": 184,
+          "oracle_success_rate": 0.9456521739130435,
+          "policy_expert_regret": 1.0790035518028005,
+          "policy_oracle_regret": 1.0790275562777305,
+          "policy_rollout_progress": 0.57163308003067,
+          "policy_rollout_success_rate": 0.28804347826086957,
+          "restore_max_error": 4.76837158203125e-07
+        },
+        "PullCube-v1": {
+          "action_mse_to_best": 0.57169899932648,
+          "expert_success_rate": 0.25,
+          "num_groups": 76,
+          "oracle_success_rate": 0.40789473684210525,
+          "policy_expert_regret": 0.3012117788587746,
+          "policy_oracle_regret": 0.5738564925758463,
+          "policy_rollout_progress": 0.2650559822979726,
+          "policy_rollout_success_rate": 0.15789473684210525,
+          "restore_max_error": 4.76837158203125e-07
+        },
+        "PushCube-v1": {
+          "action_mse_to_best": 0.3564448890586694,
+          "expert_success_rate": 0.8198198198198198,
+          "num_groups": 111,
+          "oracle_success_rate": 1.0,
+          "policy_expert_regret": 0.3483100165654947,
+          "policy_oracle_regret": 0.4142685801059276,
+          "policy_rollout_progress": 0.8019476361102886,
+          "policy_rollout_success_rate": 0.7837837837837838,
+          "restore_max_error": 4.76837158203125e-07
+        },
+        "StackCube-v1": {
+          "action_mse_to_best": 0.4872079030968822,
+          "expert_success_rate": 0.6923076923076923,
+          "num_groups": 91,
+          "oracle_success_rate": 0.8571428571428571,
+          "policy_expert_regret": 1.1282239404025969,
+          "policy_oracle_regret": 1.3003716945320696,
+          "policy_rollout_progress": 0.3902336326274243,
+          "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_margin0p25.json",
+      "num_groups": 575,
+      "selection_mode": "retrieval_residual",
+      "num_candidates": 16,
+      "candidate_sigma": 0.0,
+      "selection_margin": 0.25,
+      "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,
+      "retrieval_residual_anchor": "expert",
+      "policy_rollout_success_rate": 0.3652173913043478,
+      "policy_rollout_progress": 0.5809428592404172,
+      "oracle_success_rate": 0.8765217391304347,
+      "action_mse_to_best": 0.41576463217077697,
+      "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": 0.35361564260771655,
+          "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": 0.3354226778440102,
+          "expert_success_rate": 0.9444444444444444,
+          "num_groups": 198,
+          "oracle_success_rate": 0.9595959595959596,
+          "policy_expert_regret": 0.9955595935121028,
+          "policy_oracle_regret": 1.0026430440408112,
+          "policy_rollout_progress": 0.6289237912819542,
+          "policy_rollout_success_rate": 0.32323232323232326,
+          "restore_max_error": 4.76837158203125e-07
+        },
+        "PullCube-v1": {
+          "action_mse_to_best": 0.6295425814059046,
+          "expert_success_rate": 0.24444444444444444,
+          "num_groups": 90,
+          "oracle_success_rate": 0.4666666666666667,
+          "policy_expert_regret": 0.2935961757958972,
+          "policy_oracle_regret": 0.5007589669760111,
+          "policy_rollout_progress": 0.3386433102121575,
+          "policy_rollout_success_rate": 0.23333333333333334,
+          "restore_max_error": 4.0978193283081055e-07
+        },
+        "PushCube-v1": {
+          "action_mse_to_best": 0.3798191750256142,
+          "expert_success_rate": 0.8514851485148515,
+          "num_groups": 101,
+          "oracle_success_rate": 1.0,
+          "policy_expert_regret": 0.36240387021905124,
+          "policy_oracle_regret": 0.43505349238910296,
+          "policy_rollout_progress": 0.7926692798881247,
+          "policy_rollout_success_rate": 0.7722772277227723,
+          "restore_max_error": 4.76837158203125e-07
+        },
+        "StackCube-v1": {
+          "action_mse_to_best": 0.48537002878470553,
+          "expert_success_rate": 0.7666666666666667,
+          "num_groups": 90,
+          "oracle_success_rate": 0.9111111111111111,
+          "policy_expert_regret": 1.1566143698162503,
+          "policy_oracle_regret": 1.3056865349411964,
+          "policy_rollout_progress": 0.4096617301305135,
+          "policy_rollout_success_rate": 0.16666666666666666,
+          "restore_max_error": 4.76837158203125e-07
+        }
+      }
+    }
+  ]
+}

results/h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_scale0p35_safe_types_margin0p25_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_margin0p25.json`
+Completed seeds: 3
+Baseline h=4 policy success: 29.67%
+Baseline h=16 rank-checkpoint success: 29.74%
+
+Mean success: 34.55% +/- 1.72%
+Gain vs h=16 rank checkpoint: +4.81%
+Mean progress: 56.35%
+Mean action MSE to best: 0.394
+
+| seed | mode | k | retrieval K | retrieval metric | residual anchor | min type success | residual scale | margin | sigma | opt steps | trust | success | progress | oracle | action MSE |
+|---:|---|---:|---:|---|---|---:|---:|---:|---:|---:|---:|---:|---:|---:|---:|
+| 0 | retrieval_residual | 16 | 1 | raw | expert | 0.00 | 0.35 | 0.250 | 0.00 | 0 | 0.00 | 33.74% | 55.02% | 85.74% | 0.380 |
+| 1 | retrieval_residual | 16 | 1 | raw | expert | 0.00 | 0.35 | 0.250 | 0.00 | 0 | 0.00 | 33.39% | 55.94% | 86.96% | 0.388 |
+| 2 | retrieval_residual | 16 | 1 | raw | expert | 0.00 | 0.35 | 0.250 | 0.00 | 0 | 0.00 | 36.52% | 58.09% | 87.65% | 0.416 |

results/paper_core_results.md

@@ -22,16 +22,19 @@ baseline is the h=16 rank-checkpoint online rollout (`29.74%`).
 | Field-selected no-expert policy + field, seed-0 train map | No | No | 27.65% | -2.09 pp | Field scoring around that student remains below baseline |
 | Field-selected no-expert policy, aligned allmap | No | No | 28.00% | -1.74 pp | Full train/val target coverage does not fix the field-teacher student |
 | Field-selected no-expert policy + field, aligned allmap | No | No | 26.49% | -3.25 pp | Field scoring around the aligned student remains below baseline |
+| Residual-tangent distillation policy, aligned allmap | No | No | 28.87% | -0.87 pp | Low pseudo-target BC loss does not translate into rollout success |
 | Train-state residual retrieval | No | No | 32.12% | +2.38 pp | Transferred counterfactual residuals are a positive clean bridge |
 | Train-state residual retrieval, scale 0.25 | No | No | 32.93% | +3.19 pp | Smaller tangent step ties the previous clean best |
 | Train-state residual retrieval, scale 0.50 | No | No | 33.33% | +3.59 pp | Calibrated local tangent transport |
 | Train-state residual retrieval, no random residuals | No | No | 33.45% | +3.71 pp | Removing anti-goal random residuals helps slightly |
 | Train-state residual retrieval, no random/wrong-direction residuals | No | No | 33.57% | +3.83 pp | Anti-goal family masking improves the clean bridge |
 | Train-state residual retrieval, policy/no-op/wrong-gripper residuals | No | No | 33.68% | +3.94 pp | Typed family mask improves clean bridge |
-| Train-state residual retrieval, policy/no-op/wrong-gripper, scale 0.35 | No | No | 33.74% | +4.00 pp | Current best deployment-clean diagnostic |
+| 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 | Current best deployment-clean diagnostic; abstains unless field advantage beats policy |
+| 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 |
-| Train-state residual retrieval, train family reliability prior | No | No | 33.33% | +3.59 pp | Train terminal-success thresholds through 0.75 do not filter enough |
+| Train-state residual retrieval, repaired train family reliability prior | No | No | 33.28-33.33% | +3.54-3.59 pp | Train terminal-success thresholds do not recover the typed safe mask |
 | Train-state residual retrieval, scale 0.75 | No | No | 32.70% | +2.96 pp | Larger tangent steps begin to lose success |
 | Train-state residual retrieval, scale 1.25 | No | No | 32.52% | +2.78 pp | Further scale increase does not help |
 | Residual+Gaussian hybrid, K32 sigma0.35 | No | No | 31.30% | +1.57 pp | Adding policy-centered Gaussian proposals dilutes residual transport |
@@ -56,18 +59,21 @@ Suggested main-table rows:
 8. Field-selected no-expert policy + field, aligned allmap
 9. Train-state residual retrieval, scale 0.50
 10. Train-state residual retrieval, typed safe families at scale 0.35
-11. Residual+Gaussian hybrid, K32 sigma0.35
-12. Lattice, near-miss only
-13. Lattice, no expert
-14. Lattice, full
-15. Oracle ceiling
+11. Train-state residual retrieval, typed safe families + advantage margin 0.20
+12. Residual-tangent distillation policy
+13. Residual+Gaussian hybrid, K32 sigma0.35
+14. Lattice, near-miss only
+15. Lattice, no expert
+16. Lattice, full
+17. Oracle ceiling
 
 Suggested claim:
 
 > DoVLA-CIL is not a better behavior-cloning policy; it is a local counterfactual action
-> selection rule. Deployment-clean typed counterfactual residual transport gives the strongest
-> clean gain so far, while field-gradient ascent, KNN residual retrieval, broader non-expert BC
-> targets, field-teacher distillation, z-score retrieval, train-family reliability priors, and
-> residual+Gaussian hybrids fail. The large effect appears only when the field is queried on
+> selection rule. Deployment-clean typed counterfactual residual transport with advantage
+> abstention gives the strongest clean gain so far, while field-gradient ascent, KNN residual
+> retrieval, broader non-expert BC targets, field-teacher/tangent distillation, z-score retrieval,
+> train-family reliability priors, policy-relative anchoring, and residual+Gaussian hybrids fail.
+> 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

@@ -16,7 +16,7 @@ when queried on proposal geometry that matches those local counterfactuals.
 | Same-state local counterfactual proposals are the mechanism | near-miss-only lattice is 55.94%; removing expert+near_miss drops to 25.57% | Strongly supported |
 | Conservative same-state result is large | no-expert lattice is 56.99% vs 29.74% policy | Main result |
 | Full lattice gives upper result | full lattice is 69.33%, oracle is 86.78% | Strong but label expert proposal clearly |
-| Deployment-clean proposal is currently a bottleneck | best clean residual transport is 33.74%, far below 56.99% | Supported |
+| Deployment-clean proposal is currently a bottleneck | best clean residual transport with counterfactual-advantage abstention is 34.84%, far below 56.99% | Supported |
 | Gradient-based field optimization does not solve the clean proposal gap | `field_optim` best observed result is 25.39% | Negative diagnostic |
 | A broader non-expert proposal target does not reduce the proposal gap | direct broad non-expert policy is 27.88%; with field scoring it is 26.49% | Negative diagnostic |
 | Counterfactual residuals transfer better than absolute retrieved actions | nearest residual retrieval is 32.12% vs absolute retrieval 28.93%; KNN4 residual drops to 29.91% | Supported as a clean bridge |
@@ -25,8 +25,11 @@ when queried on proposal geometry that matches those local counterfactuals.
 | Seed-0 train-split field-teacher distillation does not solve the proposal gap | direct student is 26.84%; with field scoring it is 27.65% | Negative diagnostic |
 | 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% | Current best clean result |
 | 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 | thresholds through 0.75 do not filter the bad families; rows stay at 33.33% | 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 |
+| Policy-relative residual anchoring does not improve the bridge | policy-anchor safe residual transport ties 33.74% rather than improving expert-anchor residuals | Negative diagnostic |
 
 ## Main Table Candidate
 
@@ -45,13 +48,15 @@ clean proposal result, the intended main rows are:
 9. Train-state residual retrieval: 32.12%
 10. Train-state residual retrieval, scale 0.50: 33.33%
 11. Train-state residual retrieval, typed safe families: 33.74%
-12. Z-score residual retrieval: 32.23-32.81%
-13. Train-family reliability prior: 32.93-33.33%
-14. Residual+Gaussian hybrid K32/K64: 31.30% / 30.90%
-15. Lattice, near-miss only: 55.94%
-16. Lattice, no expert: 56.99%
-17. Lattice, full: 69.33%
-18. Oracle ceiling: 86.78%
+12. Train-state residual retrieval, typed safe families + advantage margin: 34.84%
+13. Residual-tangent distillation policy: 28.87%
+14. Z-score residual retrieval: 32.23-32.81%
+15. Train-family reliability prior: 33.28-33.33%
+16. Residual+Gaussian hybrid K32/K64: 31.30% / 30.90%
+17. Lattice, near-miss only: 55.94%
+18. Lattice, no expert: 56.99%
+19. Lattice, full: 69.33%
+20. Oracle ceiling: 86.78%
 
 ## Novelty Framing
 
@@ -79,7 +84,7 @@ test-time search. The cleaner novelty is:
 
 ## Job Status
 
-Last checked: `2026-06-28 06:24 UTC`. No DoVLA jobs are currently queued.
+Last checked: `2026-06-28 12:25 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%.
@@ -106,13 +111,23 @@ Last checked: `2026-06-28 06:24 UTC`. No DoVLA jobs are currently queued.
 - `14859503`-`14859597`: completed typed-safe residual scale fine/zoom sweep.
   Scales `0.325`, `0.35`, and `0.40` tie as the best clean rows at 33.74%;
   scales above `0.50` fall back.
+- `14862455`-`14862460`: completed residual-tangent target export, 3-seed
+  distillation, and direct/best-rank rollouts. The aligned tangent student is
+  negative: best-policy rollout reaches 28.87%, and best-rank reaches 27.48%.
+- `14862605`-`14862612`: completed policy-relative residual-anchor and repaired
+  train-family reliability diagnostics. Policy anchoring ties the old 33.74%
+  best, while repaired reliability thresholds at scale `0.35` reach only
+  33.33%/33.28%.
+- `14862635`-`14862828`: completed counterfactual-advantage margin sweeps.
+  The current best clean row is typed-safe residual transport at scale `0.35`
+  with margin `0.20` or `0.22`: 34.84% mean success (+5.10 pp vs h=16).
 
 ## Decision Notes
 
 - Promote same-state no-expert lattice (56.99%) as the conservative mechanism
   result.
-- Use typed safe residual transport (33.74%) only as the current best clean
+- Use typed safe residual transport with advantage abstention (34.84%) only as the current best clean
   deployment diagnostic, not as a SOTA claim.
-- Treat z-score retrieval, train-family reliability priors, Gaussian hybrids,
-  field optimization, and field-teacher distillation as negative diagnostics
+- Treat z-score retrieval, repaired train-family reliability priors, Gaussian hybrids,
+  field optimization, field-teacher/tangent distillation, and policy-relative anchoring as negative diagnostics
   that sharpen the story around local counterfactual proposal geometry.

results/paper_table_status.json

@@ -232,6 +232,25 @@
       "best_config": "k16_sigma0.20",
       "gain_vs_h16_policy": -0.03246376811594198
     },
+    {
+      "key": "retrieval_residual_tangent_distill_allmap",
+      "label": "Residual-tangent distillation policy, aligned validation",
+      "path": "h16_policy_ckpt_residual_tangent_bc5_allmap_v2_best_policy_summary.json",
+      "clean_deployment": "yes",
+      "same_state_proposals": "no",
+      "expert_proposal": "no",
+      "story_role": "negative student of transported tangent teacher",
+      "fallback_success": null,
+      "pending_job": "14862455/14862456/14862457/14862458",
+      "path_exists": true,
+      "status": "complete",
+      "success": 0.288695652173913,
+      "std_success": 0.023909090582378275,
+      "completed_seeds": null,
+      "num_completed": 3,
+      "best_config": null,
+      "gain_vs_h16_policy": -0.008695652173913049
+    },
     {
       "key": "retrieval_residual",
       "label": "Train-state counterfactual residual retrieval",
@@ -441,6 +460,63 @@
       "best_config": null,
       "gain_vs_h16_policy": 0.040000000000000036
     },
+    {
+      "key": "retrieval_residual_scale035_safe_margin020",
+      "label": "Train-state residual retrieval, scale 0.35, safe residuals, advantage margin 0.20",
+      "path": "h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_scale0p35_safe_types_margin0p20_summary.json",
+      "clean_deployment": "yes",
+      "same_state_proposals": "no",
+      "expert_proposal": "no",
+      "story_role": "counterfactual advantage abstention",
+      "fallback_success": null,
+      "pending_job": "14862714/14862715",
+      "path_exists": true,
+      "status": "complete",
+      "success": 0.34840579710144925,
+      "std_success": 0.017593032933905524,
+      "completed_seeds": null,
+      "num_completed": 3,
+      "best_config": null,
+      "gain_vs_h16_policy": 0.051014492753623186
+    },
+    {
+      "key": "retrieval_residual_scale050_safe_margin020",
+      "label": "Train-state residual retrieval, scale 0.50, safe residuals, advantage margin 0.20",
+      "path": "h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_scale0p50_safe_types_margin0p20_summary.json",
+      "clean_deployment": "yes",
+      "same_state_proposals": "no",
+      "expert_proposal": "no",
+      "story_role": "counterfactual advantage abstention scale tie",
+      "fallback_success": null,
+      "pending_job": "14862802/14862803",
+      "path_exists": true,
+      "status": "complete",
+      "success": 0.34840579710144925,
+      "std_success": 0.017593032933905524,
+      "completed_seeds": null,
+      "num_completed": 3,
+      "best_config": null,
+      "gain_vs_h16_policy": 0.051014492753623186
+    },
+    {
+      "key": "retrieval_residual_policy_anchor_scale035_safe",
+      "label": "Policy-relative train-state residual retrieval, scale 0.35, safe non-expert residuals",
+      "path": "h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_policy_anchor_scale0p35_safe_noexpert_summary.json",
+      "clean_deployment": "yes",
+      "same_state_proposals": "no",
+      "expert_proposal": "no",
+      "story_role": "policy-relative tangent anchor diagnostic",
+      "fallback_success": null,
+      "pending_job": "14862605/14862606",
+      "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",
@@ -669,6 +745,44 @@
       "best_config": null,
       "gain_vs_h16_policy": 0.03188405797101451
     },
+    {
+      "key": "retrieval_residual_scale035_type_success010",
+      "label": "Train-state residual retrieval, scale 0.35, train family success >= 0.10",
+      "path": "h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_scale0p35_type_success010_summary.json",
+      "clean_deployment": "yes",
+      "same_state_proposals": "no",
+      "expert_proposal": "no",
+      "story_role": "repaired train-split reliability-prior diagnostic",
+      "fallback_success": null,
+      "pending_job": "14862609/14862610",
+      "path_exists": true,
+      "status": "complete",
+      "success": 0.3333333333333333,
+      "std_success": 0.012819330079707808,
+      "completed_seeds": null,
+      "num_completed": 3,
+      "best_config": null,
+      "gain_vs_h16_policy": 0.035942028985507246
+    },
+    {
+      "key": "retrieval_residual_scale035_type_success025",
+      "label": "Train-state residual retrieval, scale 0.35, train family success >= 0.25",
+      "path": "h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_scale0p35_type_success025_summary.json",
+      "clean_deployment": "yes",
+      "same_state_proposals": "no",
+      "expert_proposal": "no",
+      "story_role": "repaired train-split reliability-prior diagnostic",
+      "fallback_success": null,
+      "pending_job": "14862611/14862612",
+      "path_exists": true,
+      "status": "complete",
+      "success": 0.3327536231884058,
+      "std_success": 0.01458521231931493,
+      "completed_seeds": null,
+      "num_completed": 3,
+      "best_config": null,
+      "gain_vs_h16_policy": 0.03536231884057972
+    },
     {
       "key": "retrieval_residual_scale075",
       "label": "Train-state residual retrieval, scale 0.75",
@@ -842,23 +956,23 @@
     }
   ],
   "best_clean": {
-    "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",
+    "key": "retrieval_residual_scale035_safe_margin020",
+    "label": "Train-state residual retrieval, scale 0.35, safe residuals, advantage margin 0.20",
+    "path": "h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_scale0p35_safe_types_margin0p20_summary.json",
     "clean_deployment": "yes",
     "same_state_proposals": "no",
     "expert_proposal": "no",
-    "story_role": "typed tangent scale fine sweep",
+    "story_role": "counterfactual advantage abstention",
     "fallback_success": null,
-    "pending_job": "14859503/14859504",
+    "pending_job": "14862714/14862715",
     "path_exists": true,
     "status": "complete",
-    "success": 0.3373913043478261,
-    "std_success": 0.004601306627938417,
+    "success": 0.34840579710144925,
+    "std_success": 0.017593032933905524,
     "completed_seeds": null,
     "num_completed": 3,
     "best_config": null,
-    "gain_vs_h16_policy": 0.040000000000000036
+    "gain_vs_h16_policy": 0.051014492753623186
   },
   "best_mechanism_no_expert": {
     "key": "no_expert_lattice",
@@ -883,7 +997,7 @@
     "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 0.35, policy/no-op/wrong-gripper residuals at 33.74%.",
+    "Current best clean deployment row is Train-state residual retrieval, scale 0.35, safe residuals, advantage margin 0.20 at 34.84%.",
     "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

@@ -15,6 +15,7 @@ Baseline h=16 policy: 29.74%
 | field_selected_noexpert_policy_field | Field-selected no-expert distillation + field | complete k8_sigma0.10 | 27.65% | -2.09 pp | yes | no | no | student proposal with field scoring |
 | field_selected_noexpert_policy_allmap | Field-selected no-expert distillation policy, aligned validation | complete | 28.00% | -1.74 pp | yes | no | no | field-teacher student with aligned checkpoint selection |
 | field_selected_noexpert_policy_allmap_field | Field-selected no-expert distillation + field, aligned validation | complete k16_sigma0.20 | 26.49% | -3.25 pp | yes | no | no | aligned field-teacher student with field scoring |
+| retrieval_residual_tangent_distill_allmap | Residual-tangent distillation policy, aligned validation | complete | 28.87% | -0.87 pp | yes | no | no | negative student of transported tangent teacher |
 | retrieval_residual | Train-state counterfactual residual retrieval | complete | 32.12% | +2.38 pp | yes | no | no | transferable local tangent proposal |
 | retrieval_residual_scale025 | Train-state residual retrieval, scale 0.25 | complete | 32.93% | +3.19 pp | yes | no | no | tangent transport scale ablation |
 | retrieval_residual_scale050 | Train-state residual retrieval, scale 0.50 | complete | 33.33% | +3.59 pp | yes | no | no | tangent transport scale ablation |
@@ -26,6 +27,9 @@ Baseline h=16 policy: 29.74%
 | 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 | complete | 33.74% | +4.00 pp | yes | no | no | typed tangent scale fine sweep |
+| retrieval_residual_scale035_safe_margin020 | Train-state residual retrieval, scale 0.35, safe residuals, advantage margin 0.20 | complete | 34.84% | +5.10 pp | yes | no | no | counterfactual advantage abstention |
+| retrieval_residual_scale050_safe_margin020 | Train-state residual retrieval, scale 0.50, safe residuals, advantage margin 0.20 | complete | 34.84% | +5.10 pp | yes | no | no | counterfactual advantage abstention scale tie |
+| retrieval_residual_policy_anchor_scale035_safe | Policy-relative train-state residual retrieval, scale 0.35, safe non-expert residuals | complete | 33.74% | +4.00 pp | yes | no | no | policy-relative tangent anchor diagnostic |
 | retrieval_residual_scale030_safe_types | Train-state residual retrieval, scale 0.30, policy/no-op/wrong-gripper residuals | complete | 33.51% | +3.77 pp | yes | no | no | typed tangent scale zoom sweep |
 | retrieval_residual_scale0325_safe_types | Train-state residual retrieval, scale 0.325, policy/no-op/wrong-gripper residuals | complete | 33.74% | +4.00 pp | 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 | complete | 33.51% | +3.77 pp | yes | no | no | typed tangent scale zoom sweep |
@@ -38,6 +42,8 @@ Baseline h=16 policy: 29.74%
 | 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 |
 | retrieval_residual_scale050_type_success075 | Train-state residual retrieval, scale 0.50, train family success >= 0.75 | complete | 33.33% | +3.59 pp | yes | no | no | train-split residual family reliability prior |
 | retrieval_residual_scale025_type_success025 | Train-state residual retrieval, scale 0.25, train family success >= 0.25 | complete | 32.93% | +3.19 pp | yes | no | no | train-split residual family reliability prior |
+| retrieval_residual_scale035_type_success010 | Train-state residual retrieval, scale 0.35, train family success >= 0.10 | complete | 33.33% | +3.59 pp | yes | no | no | repaired train-split reliability-prior diagnostic |
+| retrieval_residual_scale035_type_success025 | Train-state residual retrieval, scale 0.35, train family success >= 0.25 | complete | 33.28% | +3.54 pp | yes | no | no | repaired train-split reliability-prior diagnostic |
 | retrieval_residual_scale075 | Train-state residual retrieval, scale 0.75 | complete | 32.70% | +2.96 pp | yes | no | no | tangent transport scale ablation |
 | retrieval_residual_scale125 | Train-state residual retrieval, scale 1.25 | complete | 32.52% | +2.78 pp | yes | no | no | tangent transport scale ablation |
 | retrieval_residual_hybrid_k32 | Train-state residual + Gaussian proposals, K32 sigma0.35 | complete | 31.30% | +1.57 pp | yes | no | no | hybrid tangent/local proposal bridge |
@@ -53,7 +59,7 @@ 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 0.35, policy/no-op/wrong-gripper residuals at 33.74%.
+- Current best clean deployment row is Train-state residual retrieval, scale 0.35, safe residuals, advantage margin 0.20 at 34.84%.
 - 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

@@ -135,6 +135,16 @@ SPECS = [
         story_role="aligned field-teacher student with field scoring",
         pending_job="14858449/14858450/14858453/14858454",
     ),
+    ResultSpec(
+        key="retrieval_residual_tangent_distill_allmap",
+        label="Residual-tangent distillation policy, aligned validation",
+        path="h16_policy_ckpt_residual_tangent_bc5_allmap_v2_best_policy_summary.json",
+        clean_deployment="yes",
+        same_state_proposals="no",
+        expert_proposal="no",
+        story_role="negative student of transported tangent teacher",
+        pending_job="14862455/14862456/14862457/14862458",
+    ),
     ResultSpec(
         key="retrieval_residual",
         label="Train-state counterfactual residual retrieval",
@@ -245,6 +255,36 @@ SPECS = [
         story_role="typed tangent scale fine sweep",
         pending_job="14859503/14859504",
     ),
+    ResultSpec(
+        key="retrieval_residual_scale035_safe_margin020",
+        label="Train-state residual retrieval, scale 0.35, safe residuals, advantage margin 0.20",
+        path="h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_scale0p35_safe_types_margin0p20_summary.json",
+        clean_deployment="yes",
+        same_state_proposals="no",
+        expert_proposal="no",
+        story_role="counterfactual advantage abstention",
+        pending_job="14862714/14862715",
+    ),
+    ResultSpec(
+        key="retrieval_residual_scale050_safe_margin020",
+        label="Train-state residual retrieval, scale 0.50, safe residuals, advantage margin 0.20",
+        path="h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_scale0p50_safe_types_margin0p20_summary.json",
+        clean_deployment="yes",
+        same_state_proposals="no",
+        expert_proposal="no",
+        story_role="counterfactual advantage abstention scale tie",
+        pending_job="14862802/14862803",
+    ),
+    ResultSpec(
+        key="retrieval_residual_policy_anchor_scale035_safe",
+        label="Policy-relative train-state residual retrieval, scale 0.35, safe non-expert residuals",
+        path="h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_policy_anchor_scale0p35_safe_noexpert_summary.json",
+        clean_deployment="yes",
+        same_state_proposals="no",
+        expert_proposal="no",
+        story_role="policy-relative tangent anchor diagnostic",
+        pending_job="14862605/14862606",
+    ),
     ResultSpec(
         key="retrieval_residual_scale030_safe_types",
         label="Train-state residual retrieval, scale 0.30, policy/no-op/wrong-gripper residuals",
@@ -365,6 +405,26 @@ SPECS = [
         story_role="train-split residual family reliability prior",
         pending_job="14859297/14859298",
     ),
+    ResultSpec(
+        key="retrieval_residual_scale035_type_success010",
+        label="Train-state residual retrieval, scale 0.35, train family success >= 0.10",
+        path="h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_scale0p35_type_success010_summary.json",
+        clean_deployment="yes",
+        same_state_proposals="no",
+        expert_proposal="no",
+        story_role="repaired train-split reliability-prior diagnostic",
+        pending_job="14862609/14862610",
+    ),
+    ResultSpec(
+        key="retrieval_residual_scale035_type_success025",
+        label="Train-state residual retrieval, scale 0.35, train family success >= 0.25",
+        path="h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_scale0p35_type_success025_summary.json",
+        clean_deployment="yes",
+        same_state_proposals="no",
+        expert_proposal="no",
+        story_role="repaired train-split reliability-prior diagnostic",
+        pending_job="14862611/14862612",
+    ),
     ResultSpec(
         key="retrieval_residual_scale075",
         label="Train-state residual retrieval, scale 0.75",