Auto-sync: 2026-06-28 13:45:03 (part 2)

results/h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_k1grid_tight_safe_ray_margin0p20_summary.json

@@ -0,0 +1,313 @@
+{
+  "run_root": "/scratch/knguy52/dovla/experiments/dovla_h16_policy_ckpt_runs",
+  "objective": "near_miss_policy_bc5",
+  "out_name": "policy_rollout_retrieval_residual_k1grid_tight_safe_ray_margin0p20.json",
+  "num_completed": 3,
+  "baseline_h4_policy_success": 0.2967,
+  "baseline_h16_rank_checkpoint_success": 0.29739130434782607,
+  "mean_success": 0.34840579710144925,
+  "std_success": 0.01458521231931493,
+  "mean_progress": 0.5659862564105076,
+  "mean_action_mse_to_best": 0.4006169540839999,
+  "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_k1grid_tight_safe_ray_margin0p20.json",
+      "num_groups": 575,
+      "selection_mode": "retrieval_residual",
+      "num_candidates": 48,
+      "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": 1,
+      "retrieval_metric": "raw",
+      "retrieval_type_min_success": 0.0,
+      "retrieval_residual_scale": 0.4,
+      "retrieval_residual_scales": [
+        0.3,
+        0.4,
+        0.5
+      ],
+      "retrieval_residual_anchor": "expert",
+      "retrieval_residual_reduce": "none",
+      "policy_rollout_success_rate": 0.3408695652173913,
+      "policy_rollout_progress": 0.5540020645681121,
+      "oracle_success_rate": 0.8573913043478261,
+      "action_mse_to_best": 0.3867582200864411,
+      "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": 0.3032734078140213,
+          "expert_success_rate": 0.9375,
+          "num_groups": 208,
+          "oracle_success_rate": 0.9471153846153846,
+          "policy_expert_regret": 0.9612491831111794,
+          "policy_oracle_regret": 0.9746720505376848,
+          "policy_rollout_progress": 0.6160174607155988,
+          "policy_rollout_success_rate": 0.34134615384615385,
+          "restore_max_error": 4.76837158203125e-07
+        },
+        "PullCube-v1": {
+          "action_mse_to_best": 0.5990147828668743,
+          "expert_success_rate": 0.19480519480519481,
+          "num_groups": 77,
+          "oracle_success_rate": 0.36363636363636365,
+          "policy_expert_regret": 0.1972016387274368,
+          "policy_oracle_regret": 0.3692050660543095,
+          "policy_rollout_progress": 0.34430154372768645,
+          "policy_rollout_success_rate": 0.24675324675324675,
+          "restore_max_error": 4.76837158203125e-07
+        },
+        "PushCube-v1": {
+          "action_mse_to_best": 0.36807565609032805,
+          "expert_success_rate": 0.8617021276595744,
+          "num_groups": 94,
+          "oracle_success_rate": 0.9893617021276596,
+          "policy_expert_regret": 0.5719942746961371,
+          "policy_oracle_regret": 0.6684110583143031,
+          "policy_rollout_progress": 0.6826527714729309,
+          "policy_rollout_success_rate": 0.648936170212766,
+          "restore_max_error": 4.76837158203125e-07
+        },
+        "StackCube-v1": {
+          "action_mse_to_best": 0.4805169538547746,
+          "expert_success_rate": 0.6767676767676768,
+          "num_groups": 99,
+          "oracle_success_rate": 0.8585858585858586,
+          "policy_expert_regret": 0.9659991681274741,
+          "policy_oracle_regret": 1.1401587626849763,
+          "policy_rollout_progress": 0.44842775196138057,
+          "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_k1grid_tight_safe_ray_margin0p20.json",
+      "num_groups": 575,
+      "selection_mode": "retrieval_residual",
+      "num_candidates": 48,
+      "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": 1,
+      "retrieval_metric": "raw",
+      "retrieval_type_min_success": 0.0,
+      "retrieval_residual_scale": 0.4,
+      "retrieval_residual_scales": [
+        0.3,
+        0.4,
+        0.5
+      ],
+      "retrieval_residual_anchor": "expert",
+      "retrieval_residual_reduce": "none",
+      "policy_rollout_success_rate": 0.3391304347826087,
+      "policy_rollout_progress": 0.5638537720331679,
+      "oracle_success_rate": 0.8695652173913043,
+      "action_mse_to_best": 0.39135909198742846,
+      "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": 0.320721254258862,
+          "expert_success_rate": 0.9402173913043478,
+          "num_groups": 184,
+          "oracle_success_rate": 0.9456521739130435,
+          "policy_expert_regret": 1.0790188238834557,
+          "policy_oracle_regret": 1.079042828358386,
+          "policy_rollout_progress": 0.5716178079500146,
+          "policy_rollout_success_rate": 0.28804347826086957,
+          "restore_max_error": 4.76837158203125e-07
+        },
+        "PullCube-v1": {
+          "action_mse_to_best": 0.5872988796939975,
+          "expert_success_rate": 0.25,
+          "num_groups": 76,
+          "oracle_success_rate": 0.40789473684210525,
+          "policy_expert_regret": 0.30238124975738556,
+          "policy_oracle_regret": 0.5495397285068113,
+          "policy_rollout_progress": 0.2761718237429465,
+          "policy_rollout_success_rate": 0.17105263157894737,
+          "restore_max_error": 4.76837158203125e-07
+        },
+        "PushCube-v1": {
+          "action_mse_to_best": 0.36165393194234047,
+          "expert_success_rate": 0.8198198198198198,
+          "num_groups": 111,
+          "oracle_success_rate": 1.0,
+          "policy_expert_regret": 0.33009791642695935,
+          "policy_oracle_regret": 0.3798210240967639,
+          "policy_rollout_progress": 0.8183771741014343,
+          "policy_rollout_success_rate": 0.8018018018018018,
+          "restore_max_error": 4.76837158203125e-07
+        },
+        "StackCube-v1": {
+          "action_mse_to_best": 0.4884717973194771,
+          "expert_success_rate": 0.6923076923076923,
+          "num_groups": 91,
+          "oracle_success_rate": 0.8571428571428571,
+          "policy_expert_regret": 1.1291170658973546,
+          "policy_oracle_regret": 1.3012648200268273,
+          "policy_rollout_progress": 0.38934050713266644,
+          "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_k1grid_tight_safe_ray_margin0p20.json",
+      "num_groups": 575,
+      "selection_mode": "retrieval_residual",
+      "num_candidates": 48,
+      "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": 1,
+      "retrieval_metric": "raw",
+      "retrieval_type_min_success": 0.0,
+      "retrieval_residual_scale": 0.4,
+      "retrieval_residual_scales": [
+        0.3,
+        0.4,
+        0.5
+      ],
+      "retrieval_residual_anchor": "expert",
+      "retrieval_residual_reduce": "none",
+      "policy_rollout_success_rate": 0.3652173913043478,
+      "policy_rollout_progress": 0.5801029326302428,
+      "oracle_success_rate": 0.8765217391304347,
+      "action_mse_to_best": 0.42373355017813,
+      "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": 0.3542889782741743,
+          "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": 0.3389337710245992,
+          "expert_success_rate": 0.9444444444444444,
+          "num_groups": 198,
+          "oracle_success_rate": 0.9595959595959596,
+          "policy_expert_regret": 0.9908797259000365,
+          "policy_oracle_regret": 0.9979631764287449,
+          "policy_rollout_progress": 0.6285531538435154,
+          "policy_rollout_success_rate": 0.3282828282828283,
+          "restore_max_error": 4.76837158203125e-07
+        },
+        "PullCube-v1": {
+          "action_mse_to_best": 0.6428048797779613,
+          "expert_success_rate": 0.24444444444444444,
+          "num_groups": 90,
+          "oracle_success_rate": 0.4666666666666667,
+          "policy_expert_regret": 0.2969050312897606,
+          "policy_oracle_regret": 0.5031054516702271,
+          "policy_rollout_progress": 0.328219331545693,
+          "policy_rollout_success_rate": 0.2222222222222222,
+          "restore_max_error": 4.0978193283081055e-07
+        },
+        "PushCube-v1": {
+          "action_mse_to_best": 0.3917866135292714,
+          "expert_success_rate": 0.8514851485148515,
+          "num_groups": 101,
+          "oracle_success_rate": 1.0,
+          "policy_expert_regret": 0.38038931005071885,
+          "policy_oracle_regret": 0.4174000021844807,
+          "policy_rollout_progress": 0.8004217799937371,
+          "policy_rollout_success_rate": 0.7821782178217822,
+          "restore_max_error": 4.76837158203125e-07
+        },
+        "StackCube-v1": {
+          "action_mse_to_best": 0.5011475069862273,
+          "expert_success_rate": 0.7666666666666667,
+          "num_groups": 90,
+          "oracle_success_rate": 0.9111111111111111,
+          "policy_expert_regret": 1.157637205057674,
+          "policy_oracle_regret": 1.3196244907048014,
+          "policy_rollout_progress": 0.4068348854780197,
+          "policy_rollout_success_rate": 0.15555555555555556,
+          "restore_max_error": 4.76837158203125e-07
+        }
+      }
+    }
+  ]
+}

results/h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_k1grid_tight_safe_ray_margin0p20_summary.md

@@ -0,0 +1,19 @@
+# h=16 Best-Policy Checkpoint Rollout
+
+Run root: `/scratch/knguy52/dovla/experiments/dovla_h16_policy_ckpt_runs`
+Objective: `near_miss_policy_bc5`
+Result file: `policy_rollout_retrieval_residual_k1grid_tight_safe_ray_margin0p20.json`
+Completed seeds: 3
+Baseline h=4 policy success: 29.67%
+Baseline h=16 rank-checkpoint success: 29.74%
+
+Mean success: 34.84% +/- 1.46%
+Gain vs h=16 rank checkpoint: +5.10%
+Mean progress: 56.60%
+Mean action MSE to best: 0.401
+
+| seed | mode | k | policy cand | retrieval K | retrieval metric | residual anchor | residual reduce | min type success | residual scale | residual scales | margin | sigma | opt steps | trust | success | progress | oracle | action MSE |
+|---:|---|---:|---|---:|---|---|---|---:|---:|---|---:|---:|---:|---:|---:|---:|---:|---:|
+| 0 | retrieval_residual | 48 | no | 1 | raw | expert | none | 0.00 | 0.40 | 0.30,0.40,0.50 | 0.200 | 0.00 | 0 | 0.00 | 34.09% | 55.40% | 85.74% | 0.387 |
+| 1 | retrieval_residual | 48 | no | 1 | raw | expert | none | 0.00 | 0.40 | 0.30,0.40,0.50 | 0.200 | 0.00 | 0 | 0.00 | 33.91% | 56.39% | 86.96% | 0.391 |
+| 2 | retrieval_residual | 48 | no | 1 | raw | expert | none | 0.00 | 0.40 | 0.30,0.40,0.50 | 0.200 | 0.00 | 0 | 0.00 | 36.52% | 58.01% | 87.65% | 0.424 |

results/paper_analysis.json

@@ -1,5 +1,5 @@
 {
-  "generated_utc": "2026-06-28T17:38:38+00:00",
+  "generated_utc": "2026-06-28T17:49:47+00:00",
   "mechanism_gap": {
     "best_clean_vs_direct_same_ckpt": 0.06724637681159418,
     "best_clean_vs_h16": 0.05275362318840582,

results/paper_analysis.md

@@ -1,6 +1,6 @@
 # Paper Analysis
 
-Generated: `2026-06-28T17:38:38+00:00`
+Generated: `2026-06-28T17:49:47+00:00`
 
 ## Main Seed Statistics
 

results/paper_story_memo.md

@@ -91,8 +91,9 @@ test-time search. The cleaner novelty is:
 
 ## Job Status
 
-Last checked: `2026-06-28 17:36 UTC`. Ray-search jobs are queued, pending on
-GPU priority.
+Last checked: `2026-06-28 17:40 UTC`. K1 tight ray-search job
+`14868993_[0-2]` has started running; the K2/K4 ray-search jobs are still queued,
+pending on GPU priority.
 
 - `14858328`-`14858333`: completed train-split `field_selected_noexpert_bc5`;
   direct rollout is 26.84%, field-guided best is 27.65%.
@@ -143,12 +144,13 @@ GPU priority.
   and `0.45` reach 34.72% and 34.84%; K4 median and K8 mean at scale `0.40`
   both reach 34.67%. The follow-up confirms K2 raw residual transport remains
   the best clean row.
-- `14868993`/`14868995`/`14868997`/`14868999`: pending counterfactual tangent
-  ray-search rollouts for K1/K2/K4 safe residual retrieval with scale grids.
-  Summary jobs `14868994`/`14868996`/`14868998`/`14869000` and table rebuild
-  `14869605` are dependency-gated on those rollouts. Updated rebuild job
-  `14869860` is also dependency-gated and will regenerate both paper table and
-  paired analysis outputs with the current script.
+- `14868993_[0-2]`: running the K1 tight counterfactual tangent ray-search
+  rollout. Jobs `14868995`/`14868997`/`14868999` are still pending K2/K4 safe
+  residual ray-search rollouts with scale grids. Summary jobs `14868994`/
+  `14868996`/`14868998`/`14869000` and table rebuild `14869605` are
+  dependency-gated on those rollouts. Updated rebuild job `14869860` is also
+  dependency-gated and will regenerate both paper table and paired analysis
+  outputs with the current script.
 - `14869627`: completed CPU Apptainer smoke for the new residual scale-grid
   selector. It selected index `3` on a two-residual/two-scale toy case and
   returned the expected action `0.20`, validating the candidate expansion and

results/paper_table_status.json

@@ -527,14 +527,14 @@
       "story_role": "counterfactual tangent ray-search diagnostic",
       "fallback_success": null,
       "pending_job": "14868993/14868994",
-      "path_exists": false,
-      "status": "pending",
-      "success": null,
-      "std_success": null,
+      "path_exists": true,
+      "status": "complete",
+      "success": 0.34840579710144925,
+      "std_success": 0.01458521231931493,
       "completed_seeds": null,
-      "num_completed": null,
+      "num_completed": 3,
       "best_config": null,
-      "gain_vs_h16_policy": null
+      "gain_vs_h16_policy": 0.051014492753623186
     },
     {
       "key": "retrieval_residual_k2grid_tight_safe_ray_margin020",
@@ -546,14 +546,14 @@
       "story_role": "counterfactual tangent ray-search diagnostic",
       "fallback_success": null,
       "pending_job": "14868995/14868996",
-      "path_exists": false,
-      "status": "pending",
-      "success": null,
-      "std_success": null,
+      "path_exists": true,
+      "status": "complete",
+      "success": 0.34840579710144925,
+      "std_success": 0.013053136520808913,
       "completed_seeds": null,
-      "num_completed": null,
+      "num_completed": 3,
       "best_config": null,
-      "gain_vs_h16_policy": null
+      "gain_vs_h16_policy": 0.051014492753623186
     },
     {
       "key": "retrieval_residual_k2grid_broad_safe_ray_margin020",
@@ -565,14 +565,14 @@
       "story_role": "counterfactual tangent ray-search diagnostic",
       "fallback_success": null,
       "pending_job": "14868997/14868998",
-      "path_exists": false,
-      "status": "pending",
-      "success": null,
-      "std_success": null,
+      "path_exists": true,
+      "status": "complete",
+      "success": 0.34956521739130436,
+      "std_success": 0.015061311370164159,
       "completed_seeds": null,
-      "num_completed": null,
+      "num_completed": 3,
       "best_config": null,
-      "gain_vs_h16_policy": null
+      "gain_vs_h16_policy": 0.05217391304347829
     },
     {
       "key": "retrieval_residual_k4grid_tight_safe_ray_margin020",

results/paper_table_status.md

@@ -30,9 +30,9 @@ Baseline h=16 policy: 29.74%
 | retrieval_residual_scale035_safe_margin020 | Train-state residual retrieval, scale 0.35, safe residuals, advantage margin 0.20 | complete | 34.84% | +5.10 pp | yes | no | no | counterfactual advantage abstention |
 | retrieval_residual_scale050_safe_margin020 | Train-state residual retrieval, scale 0.50, safe residuals, advantage margin 0.20 | complete | 34.84% | +5.10 pp | yes | no | no | counterfactual advantage abstention scale tie |
 | retrieval_residual_knn2_scale040_safe_margin020 | K2 train-state residual retrieval, scale 0.40, safe residuals, advantage margin 0.20 | complete | 35.01% | +5.28 pp | yes | no | no | best clean counterfactual advantage abstention |
-| retrieval_residual_k1grid_tight_safe_ray_margin020 | K1 train-state residual ray search, safe residuals, scales 0.30/0.40/0.50, advantage margin 0.20 | pending 14868993/14868994 | pending | pending | yes | no | no | counterfactual tangent ray-search diagnostic |
-| retrieval_residual_k2grid_tight_safe_ray_margin020 | K2 train-state residual ray search, safe residuals, scales 0.30/0.40/0.50, advantage margin 0.20 | pending 14868995/14868996 | pending | pending | yes | no | no | counterfactual tangent ray-search diagnostic |
-| retrieval_residual_k2grid_broad_safe_ray_margin020 | K2 train-state residual ray search, safe residuals, scales 0.20/0.35/0.50/0.65, advantage margin 0.20 | pending 14868997/14868998 | pending | pending | yes | no | no | counterfactual tangent ray-search diagnostic |
+| retrieval_residual_k1grid_tight_safe_ray_margin020 | K1 train-state residual ray search, safe residuals, scales 0.30/0.40/0.50, advantage margin 0.20 | complete | 34.84% | +5.10 pp | yes | no | no | counterfactual tangent ray-search diagnostic |
+| retrieval_residual_k2grid_tight_safe_ray_margin020 | K2 train-state residual ray search, safe residuals, scales 0.30/0.40/0.50, advantage margin 0.20 | complete | 34.84% | +5.10 pp | yes | no | no | counterfactual tangent ray-search diagnostic |
+| retrieval_residual_k2grid_broad_safe_ray_margin020 | K2 train-state residual ray search, safe residuals, scales 0.20/0.35/0.50/0.65, advantage margin 0.20 | complete | 34.96% | +5.22 pp | yes | no | no | counterfactual tangent ray-search diagnostic |
 | retrieval_residual_k4grid_tight_safe_ray_margin020 | K4 train-state residual ray search, safe residuals, scales 0.30/0.40/0.50, advantage margin 0.20 | pending 14868999/14869000 | pending | pending | yes | no | no | counterfactual tangent ray-search diagnostic |
 | retrieval_residual_k4_scale040_safe_margin020_mean_by_type | K4 train-state residual retrieval, scale 0.40, safe residuals, mean-by-type tangent consensus | complete | 34.96% | +5.22 pp | yes | no | no | counterfactual tangent consensus near-tie ablation |
 | retrieval_residual_policy_anchor_scale035_safe | Policy-relative train-state residual retrieval, scale 0.35, safe non-expert residuals | complete | 33.74% | +4.00 pp | yes | no | no | policy-relative tangent anchor diagnostic |