auto-sync 2026-07-02T19:07:26Z workspace (part 6)

workspace/results/h16_transport_field_reground_fieldonly_k6clean_dropnoopwg_b12_v1_besttransport_margin0p00_k6_srcscore_task_pick001_stack005_advw1p0_summary.json

@@ -0,0 +1,308 @@
+{
+  "run_root": "/scratch/knguy52/dovla/experiments/dovla_h16_policy_ckpt_runs",
+  "objective": "transport_field_reground_fieldonly_k6clean_dropnoopwg_b12_v1",
+  "out_name": "policy_rollout_fieldonly_k6clean_dropnoopwg_b12_v1_besttransport_margin0p00_k6_srcscore_task_pick001_stack005_advw1p0.json",
+  "key": "advw1p0",
+  "label": "Generator V1 utility-weighted tangents, expert anchor, advantage weight 1.0",
+  "num_completed": 3,
+  "mean_success": 0.383768115942029,
+  "std_success": 0.011181044360571534,
+  "mean_progress": 0.5947921582106231,
+  "mean_action_mse_to_best": 0.5140562709515402,
+  "rows": [
+    {
+      "seed": 0,
+      "path": "/scratch/knguy52/dovla/experiments/dovla_h16_policy_ckpt_runs/transport_field_reground_fieldonly_k6clean_dropnoopwg_b12_v1/seed_0/policy_rollout_fieldonly_k6clean_dropnoopwg_b12_v1_besttransport_margin0p00_k6_srcscore_task_pick001_stack005_advw1p0.json",
+      "num_groups": 575,
+      "selection_mode": "retrieval_residual",
+      "num_candidates": 48,
+      "candidate_sigma": 0.0,
+      "selection_margin": 0.0,
+      "retrieval_neighbors": 6,
+      "retrieval_metric": "raw",
+      "retrieval_residual_anchor": "expert",
+      "retrieval_residual_direction": "candidate_minus_anchor",
+      "retrieval_residual_reduce": "compose_mean_by_type",
+      "retrieval_residual_scales": [
+        0.35,
+        0.4,
+        0.45
+      ],
+      "retrieval_residual_source_score_bonus_by_task": {
+        "*": 0.0,
+        "PickCube-v1": 0.01,
+        "StackCube-v1": 0.05
+      },
+      "retrieval_residual_source_advantage_weight_scale": 1.0,
+      "retrieval_residual_min_source_advantage": -1000000000.0,
+      "lattice_exclude_types": [
+        "residual_random_negative",
+        "residual_wrong_direction",
+        "residual_near_miss+residual_no_op",
+        "residual_no_op+residual_wrong_gripper"
+      ],
+      "selected_residual_scale_counts": {
+        "0.35": 344,
+        "0.45": 147,
+        "0.4": 84
+      },
+      "policy_rollout_success_rate": 0.3791304347826087,
+      "policy_rollout_progress": 0.584225173432218,
+      "oracle_success_rate": 0.8573913043478261,
+      "action_mse_to_best": 0.5013779968104285,
+      "per_task": {
+        "LiftPegUpright-v1": {
+          "action_mse_to_best": 0.34774401688890666,
+          "expert_success_rate": 0.8865979381443299,
+          "num_groups": 97,
+          "oracle_success_rate": 0.9278350515463918,
+          "policy_expert_regret": 1.0994866144718582,
+          "policy_oracle_regret": 1.1259020921803011,
+          "policy_rollout_progress": 0.560619557487596,
+          "policy_rollout_success_rate": 0.21649484536082475,
+          "restore_max_error": 2.384185791015625e-07
+        },
+        "PickCube-v1": {
+          "action_mse_to_best": 0.4656195876976618,
+          "expert_success_rate": 0.9375,
+          "num_groups": 208,
+          "oracle_success_rate": 0.9471153846153846,
+          "policy_expert_regret": 0.9584774491242849,
+          "policy_oracle_regret": 0.9652549772245738,
+          "policy_rollout_progress": 0.6177481525855211,
+          "policy_rollout_success_rate": 0.3557692307692308,
+          "restore_max_error": 2.384185791015625e-07
+        },
+        "PullCube-v1": {
+          "action_mse_to_best": 0.6188191903101934,
+          "expert_success_rate": 0.19480519480519481,
+          "num_groups": 77,
+          "oracle_success_rate": 0.36363636363636365,
+          "policy_expert_regret": 0.18962853963111903,
+          "policy_oracle_regret": 0.3344754918974328,
+          "policy_rollout_progress": 0.395646373777614,
+          "policy_rollout_success_rate": 0.2987012987012987,
+          "restore_max_error": 2.3096799850463867e-07
+        },
+        "PushCube-v1": {
+          "action_mse_to_best": 0.5070772077491943,
+          "expert_success_rate": 0.8617021276595744,
+          "num_groups": 94,
+          "oracle_success_rate": 0.9893617021276596,
+          "policy_expert_regret": 0.3969139822937073,
+          "policy_oracle_regret": 0.45547754032180665,
+          "policy_rollout_progress": 0.7695756990541803,
+          "policy_rollout_success_rate": 0.7553191489361702,
+          "restore_max_error": 2.384185791015625e-07
+        },
+        "StackCube-v1": {
+          "action_mse_to_best": 0.6302825162055516,
+          "expert_success_rate": 0.6767676767676768,
+          "num_groups": 99,
+          "oracle_success_rate": 0.8585858585858586,
+          "policy_expert_regret": 0.884733930681691,
+          "policy_oracle_regret": 1.0437718355896497,
+          "policy_rollout_progress": 0.507604802196676,
+          "policy_rollout_success_rate": 0.29292929292929293,
+          "restore_max_error": 2.384185791015625e-07
+        }
+      }
+    },
+    {
+      "seed": 1,
+      "path": "/scratch/knguy52/dovla/experiments/dovla_h16_policy_ckpt_runs/transport_field_reground_fieldonly_k6clean_dropnoopwg_b12_v1/seed_1/policy_rollout_fieldonly_k6clean_dropnoopwg_b12_v1_besttransport_margin0p00_k6_srcscore_task_pick001_stack005_advw1p0.json",
+      "num_groups": 575,
+      "selection_mode": "retrieval_residual",
+      "num_candidates": 48,
+      "candidate_sigma": 0.0,
+      "selection_margin": 0.0,
+      "retrieval_neighbors": 6,
+      "retrieval_metric": "raw",
+      "retrieval_residual_anchor": "expert",
+      "retrieval_residual_direction": "candidate_minus_anchor",
+      "retrieval_residual_reduce": "compose_mean_by_type",
+      "retrieval_residual_scales": [
+        0.35,
+        0.4,
+        0.45
+      ],
+      "retrieval_residual_source_score_bonus_by_task": {
+        "*": 0.0,
+        "PickCube-v1": 0.01,
+        "StackCube-v1": 0.05
+      },
+      "retrieval_residual_source_advantage_weight_scale": 1.0,
+      "retrieval_residual_min_source_advantage": -1000000000.0,
+      "lattice_exclude_types": [
+        "residual_random_negative",
+        "residual_wrong_direction",
+        "residual_near_miss+residual_no_op",
+        "residual_no_op+residual_wrong_gripper"
+      ],
+      "selected_residual_scale_counts": {
+        "0.35": 308,
+        "0.4": 45,
+        "0.45": 222
+      },
+      "policy_rollout_success_rate": 0.37565217391304345,
+      "policy_rollout_progress": 0.5895823166373632,
+      "oracle_success_rate": 0.8695652173913043,
+      "action_mse_to_best": 0.5149549319684181,
+      "per_task": {
+        "LiftPegUpright-v1": {
+          "action_mse_to_best": 0.41386533976507794,
+          "expert_success_rate": 0.8584070796460177,
+          "num_groups": 113,
+          "oracle_success_rate": 0.9380530973451328,
+          "policy_expert_regret": 0.9791655188376924,
+          "policy_oracle_regret": 1.047329495570301,
+          "policy_rollout_progress": 0.6346419230762836,
+          "policy_rollout_success_rate": 0.23893805309734514,
+          "restore_max_error": 2.384185791015625e-07
+        },
+        "PickCube-v1": {
+          "action_mse_to_best": 0.4552130797958892,
+          "expert_success_rate": 0.9402173913043478,
+          "num_groups": 184,
+          "oracle_success_rate": 0.9456521739130435,
+          "policy_expert_regret": 1.0684869141145812,
+          "policy_oracle_regret": 1.0685109185895114,
+          "policy_rollout_progress": 0.573474009647844,
+          "policy_rollout_success_rate": 0.30434782608695654,
+          "restore_max_error": 2.384185791015625e-07
+        },
+        "PullCube-v1": {
+          "action_mse_to_best": 0.6736705512984803,
+          "expert_success_rate": 0.25,
+          "num_groups": 76,
+          "oracle_success_rate": 0.40789473684210525,
+          "policy_expert_regret": 0.23966128697716876,
+          "policy_oracle_regret": 0.3887315520428513,
+          "policy_rollout_progress": 0.34846177653066424,
+          "policy_rollout_success_rate": 0.23684210526315788,
+          "restore_max_error": 2.384185791015625e-07
+        },
+        "PushCube-v1": {
+          "action_mse_to_best": 0.43252414975080405,
+          "expert_success_rate": 0.8198198198198198,
+          "num_groups": 111,
+          "oracle_success_rate": 1.0,
+          "policy_expert_regret": 0.2773399173139452,
+          "policy_oracle_regret": 0.2945074844467747,
+          "policy_rollout_progress": 0.8586456687063784,
+          "policy_rollout_success_rate": 0.8468468468468469,
+          "restore_max_error": 2.3795291781425476e-07
+        },
+        "StackCube-v1": {
+          "action_mse_to_best": 0.7292742119221897,
+          "expert_success_rate": 0.6923076923076923,
+          "num_groups": 91,
+          "oracle_success_rate": 0.8571428571428571,
+          "policy_expert_regret": 1.039590414081301,
+          "policy_oracle_regret": 1.1743051871493622,
+          "policy_rollout_progress": 0.43937706308705465,
+          "policy_rollout_success_rate": 0.23076923076923078,
+          "restore_max_error": 1.9371509552001953e-07
+        }
+      }
+    },
+    {
+      "seed": 2,
+      "path": "/scratch/knguy52/dovla/experiments/dovla_h16_policy_ckpt_runs/transport_field_reground_fieldonly_k6clean_dropnoopwg_b12_v1/seed_2/policy_rollout_fieldonly_k6clean_dropnoopwg_b12_v1_besttransport_margin0p00_k6_srcscore_task_pick001_stack005_advw1p0.json",
+      "num_groups": 575,
+      "selection_mode": "retrieval_residual",
+      "num_candidates": 48,
+      "candidate_sigma": 0.0,
+      "selection_margin": 0.0,
+      "retrieval_neighbors": 6,
+      "retrieval_metric": "raw",
+      "retrieval_residual_anchor": "expert",
+      "retrieval_residual_direction": "candidate_minus_anchor",
+      "retrieval_residual_reduce": "compose_mean_by_type",
+      "retrieval_residual_scales": [
+        0.35,
+        0.4,
+        0.45
+      ],
+      "retrieval_residual_source_score_bonus_by_task": {
+        "*": 0.0,
+        "PickCube-v1": 0.01,
+        "StackCube-v1": 0.05
+      },
+      "retrieval_residual_source_advantage_weight_scale": 1.0,
+      "retrieval_residual_min_source_advantage": -1000000000.0,
+      "lattice_exclude_types": [
+        "residual_random_negative",
+        "residual_wrong_direction",
+        "residual_near_miss+residual_no_op",
+        "residual_no_op+residual_wrong_gripper"
+      ],
+      "selected_residual_scale_counts": {
+        "0.35": 340,
+        "0.45": 160,
+        "0.4": 75
+      },
+      "policy_rollout_success_rate": 0.39652173913043476,
+      "policy_rollout_progress": 0.6105689845622881,
+      "oracle_success_rate": 0.8765217391304347,
+      "action_mse_to_best": 0.5258358840757738,
+      "per_task": {
+        "LiftPegUpright-v1": {
+          "action_mse_to_best": 0.4465993321015655,
+          "expert_success_rate": 0.8229166666666666,
+          "num_groups": 96,
+          "oracle_success_rate": 0.9270833333333334,
+          "policy_expert_regret": 0.8177533126436174,
+          "policy_oracle_regret": 0.9240251917702457,
+          "policy_rollout_progress": 0.6554166648226479,
+          "policy_rollout_success_rate": 0.3541666666666667,
+          "restore_max_error": 1.955777406692505e-07
+        },
+        "PickCube-v1": {
+          "action_mse_to_best": 0.457854955711148,
+          "expert_success_rate": 0.9444444444444444,
+          "num_groups": 198,
+          "oracle_success_rate": 0.9595959595959596,
+          "policy_expert_regret": 0.9851550623925045,
+          "policy_oracle_regret": 0.9922385129212129,
+          "policy_rollout_progress": 0.6313336513409711,
+          "policy_rollout_success_rate": 0.3383838383838384,
+          "restore_max_error": 2.384185791015625e-07
+        },
+        "PullCube-v1": {
+          "action_mse_to_best": 0.676982599331273,
+          "expert_success_rate": 0.24444444444444444,
+          "num_groups": 90,
+          "oracle_success_rate": 0.4666666666666667,
+          "policy_expert_regret": 0.3264884656502141,
+          "policy_oracle_regret": 0.45266640370504724,
+          "policy_rollout_progress": 0.36531863392641145,
+          "policy_rollout_success_rate": 0.25555555555555554,
+          "restore_max_error": 2.384185791015625e-07
+        },
+        "PushCube-v1": {
+          "action_mse_to_best": 0.5085784159969575,
+          "expert_success_rate": 0.8514851485148515,
+          "num_groups": 101,
+          "oracle_success_rate": 1.0,
+          "policy_expert_regret": 0.2674899335839961,
+          "policy_oracle_regret": 0.28615492183973296,
+          "policy_rollout_progress": 0.8623599296454156,
+          "policy_rollout_success_rate": 0.8514851485148515,
+          "restore_max_error": 2.384185791015625e-07
+        },
+        "StackCube-v1": {
+          "action_mse_to_best": 0.6281329141722785,
+          "expert_success_rate": 0.7666666666666667,
+          "num_groups": 90,
+          "oracle_success_rate": 0.9111111111111111,
+          "policy_expert_regret": 1.0699496393402417,
+          "policy_oracle_regret": 1.206480886704392,
+          "policy_rollout_progress": 0.4797341487473912,
+          "policy_rollout_success_rate": 0.2,
+          "restore_max_error": 2.2351741790771484e-07
+        }
+      }
+    }
+  ]
+}

workspace/results/h16_transport_field_reground_fieldonly_k6clean_dropnoopwg_b12_v1_besttransport_margin0p00_k6_srcscore_task_pick001_stack005_advw1p0_summary.md

@@ -0,0 +1,13 @@
+# Generator V1 Rollout Summary
+
+Objective: `transport_field_reground_fieldonly_k6clean_dropnoopwg_b12_v1`
+Result file: `policy_rollout_fieldonly_k6clean_dropnoopwg_b12_v1_besttransport_margin0p00_k6_srcscore_task_pick001_stack005_advw1p0.json`
+Completed seeds: 3
+Mean success: 38.38% +/- 1.12%
+Mean progress: 59.48%
+
+| seed | success | progress | mse | anchor | adv weight | min src adv | scales |
+|---:|---:|---:|---:|---|---:|---|---|
+| 0 | 37.91% | 58.42% | 0.501 | expert | 1.00 | none | 0.35,0.4,0.45 |
+| 1 | 37.57% | 58.96% | 0.515 | expert | 1.00 | none | 0.35,0.4,0.45 |
+| 2 | 39.65% | 61.06% | 0.526 | expert | 1.00 | none | 0.35,0.4,0.45 |

workspace/results/h16_transport_field_reground_fieldonly_k6clean_dropnoopwg_b12_v1_besttransport_margin0p00_k6_srcscore_task_pick001_stack005_policyanchor_advw2p0_summary.json

@@ -0,0 +1,308 @@
+{
+  "run_root": "/scratch/knguy52/dovla/experiments/dovla_h16_policy_ckpt_runs",
+  "objective": "transport_field_reground_fieldonly_k6clean_dropnoopwg_b12_v1",
+  "out_name": "policy_rollout_fieldonly_k6clean_dropnoopwg_b12_v1_besttransport_margin0p00_k6_srcscore_task_pick001_stack005_policyanchor_advw2p0.json",
+  "key": "policyanchor_advw2p0",
+  "label": "Generator V1 positive tangents, policy anchor, advantage weight 2.0",
+  "num_completed": 3,
+  "mean_success": 0.37681159420289856,
+  "std_success": 0.013282828101321283,
+  "mean_progress": 0.5894178291635194,
+  "mean_action_mse_to_best": 0.5143569296768502,
+  "rows": [
+    {
+      "seed": 0,
+      "path": "/scratch/knguy52/dovla/experiments/dovla_h16_policy_ckpt_runs/transport_field_reground_fieldonly_k6clean_dropnoopwg_b12_v1/seed_0/policy_rollout_fieldonly_k6clean_dropnoopwg_b12_v1_besttransport_margin0p00_k6_srcscore_task_pick001_stack005_policyanchor_advw2p0.json",
+      "num_groups": 575,
+      "selection_mode": "retrieval_residual",
+      "num_candidates": 48,
+      "candidate_sigma": 0.0,
+      "selection_margin": 0.0,
+      "retrieval_neighbors": 6,
+      "retrieval_metric": "raw",
+      "retrieval_residual_anchor": "policy",
+      "retrieval_residual_direction": "candidate_minus_anchor",
+      "retrieval_residual_reduce": "compose_mean_by_type",
+      "retrieval_residual_scales": [
+        0.35,
+        0.4,
+        0.45
+      ],
+      "retrieval_residual_source_score_bonus_by_task": {
+        "*": 0.0,
+        "PickCube-v1": 0.01,
+        "StackCube-v1": 0.05
+      },
+      "retrieval_residual_source_advantage_weight_scale": 2.0,
+      "retrieval_residual_min_source_advantage": -1000000000.0,
+      "lattice_exclude_types": [
+        "residual_random_negative",
+        "residual_wrong_direction",
+        "residual_near_miss+residual_no_op",
+        "residual_no_op+residual_wrong_gripper"
+      ],
+      "selected_residual_scale_counts": {
+        "0.35": 350,
+        "0.45": 147,
+        "0.4": 78
+      },
+      "policy_rollout_success_rate": 0.3739130434782609,
+      "policy_rollout_progress": 0.5787463493072985,
+      "oracle_success_rate": 0.8573913043478261,
+      "action_mse_to_best": 0.5005510983859067,
+      "per_task": {
+        "LiftPegUpright-v1": {
+          "action_mse_to_best": 0.34528797616241214,
+          "expert_success_rate": 0.8865979381443299,
+          "num_groups": 97,
+          "oracle_success_rate": 0.9278350515463918,
+          "policy_expert_regret": 1.102015116319214,
+          "policy_oracle_regret": 1.1284305940276569,
+          "policy_rollout_progress": 0.5580910556402403,
+          "policy_rollout_success_rate": 0.21649484536082475,
+          "restore_max_error": 2.384185791015625e-07
+        },
+        "PickCube-v1": {
+          "action_mse_to_best": 0.4610416995576368,
+          "expert_success_rate": 0.9375,
+          "num_groups": 208,
+          "oracle_success_rate": 0.9471153846153846,
+          "policy_expert_regret": 0.9447308026451188,
+          "policy_oracle_regret": 0.9515083307454076,
+          "policy_rollout_progress": 0.6218794144493026,
+          "policy_rollout_success_rate": 0.36538461538461536,
+          "restore_max_error": 2.384185791015625e-07
+        },
+        "PullCube-v1": {
+          "action_mse_to_best": 0.6339810742186262,
+          "expert_success_rate": 0.19480519480519481,
+          "num_groups": 77,
+          "oracle_success_rate": 0.36363636363636365,
+          "policy_expert_regret": 0.23732203781372693,
+          "policy_oracle_regret": 0.38390447113866305,
+          "policy_rollout_progress": 0.36338652852622566,
+          "policy_rollout_success_rate": 0.2597402597402597,
+          "restore_max_error": 2.3096799850463867e-07
+        },
+        "PushCube-v1": {
+          "action_mse_to_best": 0.4994965392224332,
+          "expert_success_rate": 0.8617021276595744,
+          "num_groups": 94,
+          "oracle_success_rate": 0.9893617021276596,
+          "policy_expert_regret": 0.41792074162909326,
+          "policy_oracle_regret": 0.47648429965719263,
+          "policy_rollout_progress": 0.7592072375911347,
+          "policy_rollout_success_rate": 0.7446808510638298,
+          "restore_max_error": 2.384185791015625e-07
+        },
+        "StackCube-v1": {
+          "action_mse_to_best": 0.632909666408192,
+          "expert_success_rate": 0.6767676767676768,
+          "num_groups": 99,
+          "oracle_success_rate": 0.8585858585858586,
+          "policy_expert_regret": 0.9015669315451323,
+          "policy_oracle_regret": 1.049103739887777,
+          "policy_rollout_progress": 0.504516739586387,
+          "policy_rollout_success_rate": 0.2828282828282828,
+          "restore_max_error": 2.384185791015625e-07
+        }
+      }
+    },
+    {
+      "seed": 1,
+      "path": "/scratch/knguy52/dovla/experiments/dovla_h16_policy_ckpt_runs/transport_field_reground_fieldonly_k6clean_dropnoopwg_b12_v1/seed_1/policy_rollout_fieldonly_k6clean_dropnoopwg_b12_v1_besttransport_margin0p00_k6_srcscore_task_pick001_stack005_policyanchor_advw2p0.json",
+      "num_groups": 575,
+      "selection_mode": "retrieval_residual",
+      "num_candidates": 48,
+      "candidate_sigma": 0.0,
+      "selection_margin": 0.0,
+      "retrieval_neighbors": 6,
+      "retrieval_metric": "raw",
+      "retrieval_residual_anchor": "policy",
+      "retrieval_residual_direction": "candidate_minus_anchor",
+      "retrieval_residual_reduce": "compose_mean_by_type",
+      "retrieval_residual_scales": [
+        0.35,
+        0.4,
+        0.45
+      ],
+      "retrieval_residual_source_score_bonus_by_task": {
+        "*": 0.0,
+        "PickCube-v1": 0.01,
+        "StackCube-v1": 0.05
+      },
+      "retrieval_residual_source_advantage_weight_scale": 2.0,
+      "retrieval_residual_min_source_advantage": -1000000000.0,
+      "lattice_exclude_types": [
+        "residual_random_negative",
+        "residual_wrong_direction",
+        "residual_near_miss+residual_no_op",
+        "residual_no_op+residual_wrong_gripper"
+      ],
+      "selected_residual_scale_counts": {
+        "0.35": 306,
+        "0.4": 49,
+        "0.45": 220
+      },
+      "policy_rollout_success_rate": 0.3652173913043478,
+      "policy_rollout_progress": 0.582736657930781,
+      "oracle_success_rate": 0.8695652173913043,
+      "action_mse_to_best": 0.5151173928242339,
+      "per_task": {
+        "LiftPegUpright-v1": {
+          "action_mse_to_best": 0.4138995000507386,
+          "expert_success_rate": 0.8584070796460177,
+          "num_groups": 113,
+          "oracle_success_rate": 0.9380530973451328,
+          "policy_expert_regret": 0.9793257536613835,
+          "policy_oracle_regret": 1.058527783746213,
+          "policy_rollout_progress": 0.6347343845177541,
+          "policy_rollout_success_rate": 0.23893805309734514,
+          "restore_max_error": 2.384185791015625e-07
+        },
+        "PickCube-v1": {
+          "action_mse_to_best": 0.45511664790303813,
+          "expert_success_rate": 0.9402173913043478,
+          "num_groups": 184,
+          "oracle_success_rate": 0.9456521739130435,
+          "policy_expert_regret": 1.071216604404111,
+          "policy_oracle_regret": 1.0712406088790412,
+          "policy_rollout_progress": 0.5685504622119681,
+          "policy_rollout_success_rate": 0.29891304347826086,
+          "restore_max_error": 2.384185791015625e-07
+        },
+        "PullCube-v1": {
+          "action_mse_to_best": 0.6726667573185343,
+          "expert_success_rate": 0.25,
+          "num_groups": 76,
+          "oracle_success_rate": 0.40789473684210525,
+          "policy_expert_regret": 0.2399015161944063,
+          "policy_oracle_regret": 0.4369939785254629,
+          "policy_rollout_progress": 0.3372275539134678,
+          "policy_rollout_success_rate": 0.2236842105263158,
+          "restore_max_error": 2.384185791015625e-07
+        },
+        "PushCube-v1": {
+          "action_mse_to_best": 0.43257011533588974,
+          "expert_success_rate": 0.8198198198198198,
+          "num_groups": 111,
+          "oracle_success_rate": 1.0,
+          "policy_expert_regret": 0.2956222787096694,
+          "policy_oracle_regret": 0.3127898458424989,
+          "policy_rollout_progress": 0.8493723163196633,
+          "policy_rollout_success_rate": 0.8378378378378378,
+          "restore_max_error": 2.3795291781425476e-07
+        },
+        "StackCube-v1": {
+          "action_mse_to_best": 0.7312355801708751,
+          "expert_success_rate": 0.6923076923076923,
+          "num_groups": 91,
+          "oracle_success_rate": 0.8571428571428571,
+          "policy_expert_regret": 1.0478797831705637,
+          "policy_oracle_regret": 1.207436578123124,
+          "policy_rollout_progress": 0.4266558967448853,
+          "policy_rollout_success_rate": 0.1978021978021978,
+          "restore_max_error": 1.9371509552001953e-07
+        }
+      }
+    },
+    {
+      "seed": 2,
+      "path": "/scratch/knguy52/dovla/experiments/dovla_h16_policy_ckpt_runs/transport_field_reground_fieldonly_k6clean_dropnoopwg_b12_v1/seed_2/policy_rollout_fieldonly_k6clean_dropnoopwg_b12_v1_besttransport_margin0p00_k6_srcscore_task_pick001_stack005_policyanchor_advw2p0.json",
+      "num_groups": 575,
+      "selection_mode": "retrieval_residual",
+      "num_candidates": 48,
+      "candidate_sigma": 0.0,
+      "selection_margin": 0.0,
+      "retrieval_neighbors": 6,
+      "retrieval_metric": "raw",
+      "retrieval_residual_anchor": "policy",
+      "retrieval_residual_direction": "candidate_minus_anchor",
+      "retrieval_residual_reduce": "compose_mean_by_type",
+      "retrieval_residual_scales": [
+        0.35,
+        0.4,
+        0.45
+      ],
+      "retrieval_residual_source_score_bonus_by_task": {
+        "*": 0.0,
+        "PickCube-v1": 0.01,
+        "StackCube-v1": 0.05
+      },
+      "retrieval_residual_source_advantage_weight_scale": 2.0,
+      "retrieval_residual_min_source_advantage": -1000000000.0,
+      "lattice_exclude_types": [
+        "residual_random_negative",
+        "residual_wrong_direction",
+        "residual_near_miss+residual_no_op",
+        "residual_no_op+residual_wrong_gripper"
+      ],
+      "selected_residual_scale_counts": {
+        "0.35": 351,
+        "0.45": 160,
+        "0.4": 64
+      },
+      "policy_rollout_success_rate": 0.391304347826087,
+      "policy_rollout_progress": 0.6067704802524785,
+      "oracle_success_rate": 0.8765217391304347,
+      "action_mse_to_best": 0.52740229782041,
+      "per_task": {
+        "LiftPegUpright-v1": {
+          "action_mse_to_best": 0.44621239306676824,
+          "expert_success_rate": 0.8229166666666666,
+          "num_groups": 96,
+          "oracle_success_rate": 0.9270833333333334,
+          "policy_expert_regret": 0.8243164004137119,
+          "policy_oracle_regret": 0.9438704351584116,
+          "policy_rollout_progress": 0.6459880881011486,
+          "policy_rollout_success_rate": 0.34375,
+          "restore_max_error": 1.955777406692505e-07
+        },
+        "PickCube-v1": {
+          "action_mse_to_best": 0.4559677015387952,
+          "expert_success_rate": 0.9444444444444444,
+          "num_groups": 198,
+          "oracle_success_rate": 0.9595959595959596,
+          "policy_expert_regret": 1.013960270756724,
+          "policy_oracle_regret": 1.0210437212854322,
+          "policy_rollout_progress": 0.6227304631787719,
+          "policy_rollout_success_rate": 0.3181818181818182,
+          "restore_max_error": 2.384185791015625e-07
+        },
+        "PullCube-v1": {
+          "action_mse_to_best": 0.6842520299885008,
+          "expert_success_rate": 0.24444444444444444,
+          "num_groups": 90,
+          "oracle_success_rate": 0.4666666666666667,
+          "policy_expert_regret": 0.30676752639313537,
+          "policy_oracle_regret": 0.4325280127219028,
+          "policy_rollout_progress": 0.3744079527962539,
+          "policy_rollout_success_rate": 0.26666666666666666,
+          "restore_max_error": 2.384185791015625e-07
+        },
+        "PushCube-v1": {
+          "action_mse_to_best": 0.5130543247012809,
+          "expert_success_rate": 0.8514851485148515,
+          "num_groups": 101,
+          "oracle_success_rate": 1.0,
+          "policy_expert_regret": 0.24754097364326516,
+          "policy_oracle_regret": 0.26621239536469526,
+          "policy_rollout_progress": 0.8724014660214433,
+          "policy_rollout_success_rate": 0.8613861386138614,
+          "restore_max_error": 2.384185791015625e-07
+        },
+        "StackCube-v1": {
+          "action_mse_to_best": 0.6304128568205568,
+          "expert_success_rate": 0.7666666666666667,
+          "num_groups": 90,
+          "oracle_success_rate": 0.9111111111111111,
+          "policy_expert_regret": 1.0856013195382224,
+          "policy_oracle_regret": 1.2221325669023726,
+          "policy_rollout_progress": 0.4640919350915485,
+          "policy_rollout_success_rate": 0.2,
+          "restore_max_error": 2.2351741790771484e-07
+        }
+      }
+    }
+  ]
+}

workspace/results/h16_transport_field_reground_fieldonly_k6clean_dropnoopwg_b12_v1_besttransport_margin0p00_k6_srcscore_task_pick001_stack005_policyanchor_advw2p0_summary.md

@@ -0,0 +1,13 @@
+# Generator V1 Rollout Summary
+
+Objective: `transport_field_reground_fieldonly_k6clean_dropnoopwg_b12_v1`
+Result file: `policy_rollout_fieldonly_k6clean_dropnoopwg_b12_v1_besttransport_margin0p00_k6_srcscore_task_pick001_stack005_policyanchor_advw2p0.json`
+Completed seeds: 3
+Mean success: 37.68% +/- 1.33%
+Mean progress: 58.94%
+
+| seed | success | progress | mse | anchor | adv weight | min src adv | scales |
+|---:|---:|---:|---:|---|---:|---|---|
+| 0 | 37.39% | 57.87% | 0.501 | policy | 2.00 | none | 0.35,0.4,0.45 |
+| 1 | 36.52% | 58.27% | 0.515 | policy | 2.00 | none | 0.35,0.4,0.45 |
+| 2 | 39.13% | 60.68% | 0.527 | policy | 2.00 | none | 0.35,0.4,0.45 |

workspace/results/paper_analysis.json

@@ -23,7 +23,7 @@
     "selected_success_for_65pct_gap_closure": 0.47449275362318843,
     "selected_success_for_75pct_gap_closure": 0.5017391304347827
   },
-  "generated_utc": "2026-07-02T18:32:12+00:00",
+  "generated_utc": "2026-07-02T19:30:26+00:00",
   "mechanism_gap": {
     "best_clean_vs_direct_same_ckpt": 0.1060869565217391,
     "best_clean_vs_h16": 0.09159420289855075,
@@ -12710,19 +12710,517 @@
       "source": "results/h16_transport_field_reground_fieldonly_k6clean_dropnoopwg_b12_v1_besttransport_margin0p00_k6_srcscore_task_pick001_stack005_summary.json",
       "std_success": 0.018595959341849776
     },
+    "transport_field_reground_fieldonly_k6matched_b12_clean_k6_dropnoopwg_retargeted_srcscore_task_pick001_stack005_advw0p5": {
+      "ci95_success": 0.04327783177865381,
+      "headline_ci95_success": 0.04327783177865381,
+      "headline_metric": "mean_success",
+      "headline_metric_label": "deployment",
+      "headline_std_success": 0.017420265740657068,
+      "headline_success": 0.3872463768115942,
+      "label": "Generator V1 utility-weighted tangents, expert anchor, advantage weight 0.5",
+      "mean_action_mse_to_best": 0.5136112351842441,
+      "mean_progress": 0.5970969428575558,
+      "mean_success": 0.3872463768115942,
+      "num_completed": 3,
+      "per_task_success": {
+        "LiftPegUpright-v1": {
+          "mean_num_groups": 102.0,
+          "mean_success": 0.2738253175399649,
+          "std_success": 0.07861590073068579
+        },
+        "PickCube-v1": {
+          "mean_num_groups": 196.66666666666666,
+          "mean_success": 0.33118391270565184,
+          "std_success": 0.027971684673707302
+        },
+        "PullCube-v1": {
+          "mean_num_groups": 81.0,
+          "mean_success": 0.25931368826105666,
+          "std_success": 0.03764948274063547
+        },
+        "PushCube-v1": {
+          "mean_num_groups": 102.0,
+          "mean_success": 0.8269442815471887,
+          "std_success": 0.07131455226250227
+        },
+        "StackCube-v1": {
+          "mean_num_groups": 93.33333333333333,
+          "mean_success": 0.25503755503755504,
+          "std_success": 0.05987073762679782
+        }
+      },
+      "seed_action_mse_to_best": {
+        "0": 0.49949725674060375,
+        "1": 0.5162805855784404,
+        "2": 0.5250558632336881
+      },
+      "seed_progress": {
+        "0": 0.5839397546810948,
+        "1": 0.5902519925149239,
+        "2": 0.6170990813766485
+      },
+      "seed_success": {
+        "0": 0.3808695652173913,
+        "1": 0.3739130434782609,
+        "2": 0.40695652173913044
+      },
+      "selected_candidate_type_counts": {
+        "retrieval_residual_policy_residual": 607,
+        "retrieval_residual_residual_near_miss": 214,
+        "retrieval_residual_residual_near_miss+residual_wrong_gripper": 216,
+        "retrieval_residual_residual_no_op": 258,
+        "retrieval_residual_residual_wrong_gripper": 430
+      },
+      "selected_residual_scale_counts": {
+        "0.35": 992,
+        "0.4": 206,
+        "0.45": 527
+      },
+      "selected_type_outcomes": {
+        "retrieval_residual_policy_residual": {
+          "count": 607.0,
+          "mean_progress": 0.5481918433564523,
+          "success_rate": 0.29654036243822074
+        },
+        "retrieval_residual_residual_near_miss": {
+          "count": 214.0,
+          "mean_progress": 0.7708240913697214,
+          "success_rate": 0.6261682242990654
+        },
+        "retrieval_residual_residual_near_miss+residual_wrong_gripper": {
+          "count": 216.0,
+          "mean_progress": 0.7071546247418694,
+          "success_rate": 0.5231481481481481
+        },
+        "retrieval_residual_residual_no_op": {
+          "count": 258.0,
+          "mean_progress": 0.543082087849404,
+          "success_rate": 0.35658914728682173
+        },
+        "retrieval_residual_residual_wrong_gripper": {
+          "count": 430.0,
+          "mean_progress": 0.5567973124404806,
+          "success_rate": 0.34651162790697676
+        }
+      },
+      "source": "results/h16_transport_field_reground_fieldonly_k6clean_dropnoopwg_b12_v1_besttransport_margin0p00_k6_srcscore_task_pick001_stack005_advw0p5_summary.json",
+      "std_success": 0.017420265740657068
+    },
     "transport_field_reground_fieldonly_k6matched_b12_clean_k6_dropnoopwg_retargeted_srcscore_task_pick001_stack005_advw1p0": {
+      "ci95_success": 0.027777495713921813,
+      "headline_ci95_success": 0.027777495713921813,
       "headline_metric": "mean_success",
       "headline_metric_label": "deployment",
+      "headline_std_success": 0.011181044360571532,
+      "headline_success": 0.383768115942029,
       "label": "Generator V1 utility-weighted tangents, expert anchor, advantage weight 1.0",
+      "mean_action_mse_to_best": 0.51405627095154,
+      "mean_progress": 0.5947921582106231,
+      "mean_success": 0.383768115942029,
+      "num_completed": 3,
+      "per_task_success": {
+        "LiftPegUpright-v1": {
+          "mean_num_groups": 102.0,
+          "mean_success": 0.2698665217082789,
+          "std_success": 0.07386345661678752
+        },
+        "PickCube-v1": {
+          "mean_num_groups": 196.66666666666666,
+          "mean_success": 0.3328336317466753,
+          "std_success": 0.02615614287621478
+        },
+        "PullCube-v1": {
+          "mean_num_groups": 81.0,
+          "mean_success": 0.2636996531733374,
+          "std_success": 0.031723566913421014
+        },
+        "PushCube-v1": {
+          "mean_num_groups": 102.0,
+          "mean_success": 0.8178837147659562,
+          "std_success": 0.054232113490929594
+        },
+        "StackCube-v1": {
+          "mean_num_groups": 93.33333333333333,
+          "mean_success": 0.2412328412328412,
+          "std_success": 0.04734003304852265
+        }
+      },
+      "seed_action_mse_to_best": {
+        "0": 0.5013779968104285,
+        "1": 0.5149549319684181,
+        "2": 0.5258358840757738
+      },
+      "seed_progress": {
+        "0": 0.584225173432218,
+        "1": 0.5895823166373632,
+        "2": 0.6105689845622881
+      },
+      "seed_success": {
+        "0": 0.3791304347826087,
+        "1": 0.37565217391304345,
+        "2": 0.39652173913043476
+      },
+      "selected_candidate_type_counts": {
+        "retrieval_residual_policy_residual": 593,
+        "retrieval_residual_residual_near_miss": 237,
+        "retrieval_residual_residual_near_miss+residual_wrong_gripper": 219,
+        "retrieval_residual_residual_no_op": 254,
+        "retrieval_residual_residual_wrong_gripper": 422
+      },
+      "selected_residual_scale_counts": {
+        "0.35": 992,
+        "0.4": 204,
+        "0.45": 529
+      },
+      "selected_type_outcomes": {
+        "retrieval_residual_policy_residual": {
+          "count": 593.0,
+          "mean_progress": 0.5434360912911481,
+          "success_rate": 0.2917369308600337
+        },
+        "retrieval_residual_residual_near_miss": {
+          "count": 237.0,
+          "mean_progress": 0.752326833964463,
+          "success_rate": 0.5822784810126582
+        },
+        "retrieval_residual_residual_near_miss+residual_wrong_gripper": {
+          "count": 219.0,
+          "mean_progress": 0.7189138394325427,
+          "success_rate": 0.5296803652968036
+        },
+        "retrieval_residual_residual_no_op": {
+          "count": 254.0,
+          "mean_progress": 0.5431713150542845,
+          "success_rate": 0.3543307086614173
+        },
+        "retrieval_residual_residual_wrong_gripper": {
+          "count": 422.0,
+          "mean_progress": 0.5451416262288653,
+          "success_rate": 0.34360189573459715
+        }
+      },
+      "source": "results/h16_transport_field_reground_fieldonly_k6clean_dropnoopwg_b12_v1_besttransport_margin0p00_k6_srcscore_task_pick001_stack005_advw1p0_summary.json",
+      "std_success": 0.011181044360571532
+    },
+    "transport_field_reground_fieldonly_k6matched_b12_clean_k6_dropnoopwg_retargeted_srcscore_task_pick001_stack005_advw1p0_oraclek8": {
+      "headline_metric": "mean_candidate_oracle_success_rate",
+      "headline_metric_label": "candidate-oracle",
+      "label": "Generator V1 expert-anchor advantage weight 1.0, candidate-oracle K8",
       "missing": true,
-      "source": "results/h16_transport_field_reground_fieldonly_k6clean_dropnoopwg_b12_v1_besttransport_margin0p00_k6_srcscore_task_pick001_stack005_advw1p0_summary.json"
+      "source": "results/h16_transport_field_reground_fieldonly_k6clean_dropnoopwg_b12_v1_besttransport_margin0p00_k6_srcscore_task_pick001_stack005_advw1p0_oraclek8_summary.json"
     },
     "transport_field_reground_fieldonly_k6matched_b12_clean_k6_dropnoopwg_retargeted_srcscore_task_pick001_stack005_advw2p0": {
+      "ci95_success": 0.032999037366698375,
+      "headline_ci95_success": 0.032999037366698375,
       "headline_metric": "mean_success",
       "headline_metric_label": "deployment",
+      "headline_std_success": 0.013282828101321283,
+      "headline_success": 0.3768115942028985,
       "label": "Generator V1 utility-weighted tangents, expert anchor, advantage weight 2.0",
+      "mean_action_mse_to_best": 0.5143569296768502,
+      "mean_progress": 0.5894178291635194,
+      "mean_success": 0.3768115942028985,
+      "num_completed": 3,
+      "per_task_success": {
+        "LiftPegUpright-v1": {
+          "mean_num_groups": 102.0,
+          "mean_success": 0.26639429948605664,
+          "std_success": 0.06792534648331451
+        },
+        "PickCube-v1": {
+          "mean_num_groups": 196.66666666666666,
+          "mean_success": 0.3274931590148981,
+          "std_success": 0.03420004779247759
+        },
+        "PullCube-v1": {
+          "mean_num_groups": 81.0,
+          "mean_success": 0.2500303789777474,
+          "std_success": 0.023077786338402385
+        },
+        "PushCube-v1": {
+          "mean_num_groups": 102.0,
+          "mean_success": 0.8146349425051763,
+          "std_success": 0.06171557177404845
+        },
+        "StackCube-v1": {
+          "mean_num_groups": 93.33333333333333,
+          "mean_success": 0.22687682687682686,
+          "std_success": 0.04846784140995916
+        }
+      },
+      "seed_action_mse_to_best": {
+        "0": 0.5005510983859067,
+        "1": 0.5151173928242339,
+        "2": 0.52740229782041
+      },
+      "seed_progress": {
+        "0": 0.5787463493072985,
+        "1": 0.582736657930781,
+        "2": 0.6067704802524785
+      },
+      "seed_success": {
+        "0": 0.3739130434782609,
+        "1": 0.3652173913043478,
+        "2": 0.391304347826087
+      },
+      "selected_candidate_type_counts": {
+        "retrieval_residual_policy_residual": 584,
+        "retrieval_residual_residual_near_miss": 269,
+        "retrieval_residual_residual_near_miss+residual_wrong_gripper": 227,
+        "retrieval_residual_residual_no_op": 252,
+        "retrieval_residual_residual_wrong_gripper": 393
+      },
+      "selected_residual_scale_counts": {
+        "0.35": 1007,
+        "0.4": 191,
+        "0.45": 527
+      },
+      "selected_type_outcomes": {
+        "retrieval_residual_policy_residual": {
+          "count": 584.0,
+          "mean_progress": 0.5396331625987381,
+          "success_rate": 0.2910958904109589
+        },
+        "retrieval_residual_residual_near_miss": {
+          "count": 269.0,
+          "mean_progress": 0.7359185712931324,
+          "success_rate": 0.550185873605948
+        },
+        "retrieval_residual_residual_near_miss+residual_wrong_gripper": {
+          "count": 227.0,
+          "mean_progress": 0.7157547649340745,
+          "success_rate": 0.5330396475770925
+        },
+        "retrieval_residual_residual_no_op": {
+          "count": 252.0,
+          "mean_progress": 0.5248836046774823,
+          "success_rate": 0.31746031746031744
+        },
+        "retrieval_residual_residual_wrong_gripper": {
+          "count": 393.0,
+          "mean_progress": 0.5315289889384089,
+          "success_rate": 0.3333333333333333
+        }
+      },
+      "source": "results/h16_transport_field_reground_fieldonly_k6clean_dropnoopwg_b12_v1_besttransport_margin0p00_k6_srcscore_task_pick001_stack005_advw2p0_summary.json",
+      "std_success": 0.013282828101321283
+    },
+    "transport_field_reground_fieldonly_k6matched_b12_clean_k6_dropnoopwg_retargeted_srcscore_task_pick001_stack005_advw2p0_gate0": {
+      "ci95_success": 0.04778760648872724,
+      "headline_ci95_success": 0.04778760648872724,
+      "headline_metric": "mean_success",
+      "headline_metric_label": "deployment",
+      "headline_std_success": 0.019235547852796592,
+      "headline_success": 0.36927536231884056,
+      "label": "Generator V1 utility-weighted tangents, expert anchor, source advantage gate >= 0.0",
+      "mean_action_mse_to_best": 0.45602177684910705,
+      "mean_progress": 0.5815420930200059,
+      "mean_success": 0.36927536231884056,
+      "num_completed": 3,
+      "per_task_success": {
+        "LiftPegUpright-v1": {
+          "mean_num_groups": 102.0,
+          "mean_success": 0.26639429948605664,
+          "std_success": 0.06792534648331451
+        },
+        "PickCube-v1": {
+          "mean_num_groups": 196.66666666666666,
+          "mean_success": 0.32465893607197954,
+          "std_success": 0.018763488756388967
+        },
+        "PullCube-v1": {
+          "mean_num_groups": 81.0,
+          "mean_success": 0.23624718361560468,
+          "std_success": 0.04766015797840824
+        },
+        "PushCube-v1": {
+          "mean_num_groups": 102.0,
+          "mean_success": 0.8055743757239439,
+          "std_success": 0.04486214967261923
+        },
+        "StackCube-v1": {
+          "mean_num_groups": 93.33333333333333,
+          "mean_success": 0.2081844081844082,
+          "std_success": 0.07436028694013369
+        }
+      },
+      "seed_action_mse_to_best": {
+        "0": 0.43498127313573726,
+        "1": 0.4576410755636575,
+        "2": 0.4754429818479263
+      },
+      "seed_progress": {
+        "0": 0.5694765353704924,
+        "1": 0.5735185124835824,
+        "2": 0.6016312312059429
+      },
+      "seed_success": {
+        "0": 0.36695652173913046,
+        "1": 0.35130434782608694,
+        "2": 0.38956521739130434
+      },
+      "selected_candidate_type_counts": {
+        "retrieval_residual_policy_residual": 819,
+        "retrieval_residual_residual_near_miss": 479,
+        "retrieval_residual_residual_near_miss+residual_wrong_gripper": 80,
+        "retrieval_residual_residual_no_op": 174,
+        "retrieval_residual_residual_no_op+residual_near_miss": 14,
+        "retrieval_residual_residual_wrong_gripper": 143,
+        "retrieval_residual_residual_wrong_gripper+residual_near_miss": 4,
+        "retrieval_residual_residual_wrong_gripper+residual_no_op": 12
+      },
+      "selected_residual_scale_counts": {
+        "0.35": 1208,
+        "0.4": 97,
+        "0.45": 420
+      },
+      "selected_type_outcomes": {
+        "retrieval_residual_policy_residual": {
+          "count": 819.0,
+          "mean_progress": 0.5104973517751755,
+          "success_rate": 0.2857142857142857
+        },
+        "retrieval_residual_residual_near_miss": {
+          "count": 479.0,
+          "mean_progress": 0.6349816032148711,
+          "success_rate": 0.4154488517745303
+        },
+        "retrieval_residual_residual_near_miss+residual_wrong_gripper": {
+          "count": 80.0,
+          "mean_progress": 0.8649393017403781,
+          "success_rate": 0.8
+        },
+        "retrieval_residual_residual_no_op": {
+          "count": 174.0,
+          "mean_progress": 0.6662687381793713,
+          "success_rate": 0.42528735632183906
+        },
+        "retrieval_residual_residual_no_op+residual_near_miss": {
+          "count": 14.0,
+          "mean_progress": 0.7973395671163287,
+          "success_rate": 0.5
+        },
+        "retrieval_residual_residual_wrong_gripper": {
+          "count": 143.0,
+          "mean_progress": 0.5357520567809156,
+          "success_rate": 0.3776223776223776
+        },
+        "retrieval_residual_residual_wrong_gripper+residual_near_miss": {
+          "count": 4.0,
+          "mean_progress": 0.5595042705535889,
+          "success_rate": 0.25
+        },
+        "retrieval_residual_residual_wrong_gripper+residual_no_op": {
+          "count": 12.0,
+          "mean_progress": 0.48061430764695007,
+          "success_rate": 0.3333333333333333
+        }
+      },
+      "source": "results/h16_transport_field_reground_fieldonly_k6clean_dropnoopwg_b12_v1_besttransport_margin0p00_k6_srcscore_task_pick001_stack005_advw2p0_gate0_summary.json",
+      "std_success": 0.019235547852796592
+    },
+    "transport_field_reground_fieldonly_k6matched_b12_clean_k6_dropnoopwg_retargeted_srcscore_task_pick001_stack005_advw2p0_oraclek8": {
+      "headline_metric": "mean_candidate_oracle_success_rate",
+      "headline_metric_label": "candidate-oracle",
+      "label": "Generator V1 expert-anchor advantage weight 2.0, candidate-oracle K8",
       "missing": true,
-      "source": "results/h16_transport_field_reground_fieldonly_k6clean_dropnoopwg_b12_v1_besttransport_margin0p00_k6_srcscore_task_pick001_stack005_advw2p0_summary.json"
+      "source": "results/h16_transport_field_reground_fieldonly_k6clean_dropnoopwg_b12_v1_besttransport_margin0p00_k6_srcscore_task_pick001_stack005_advw2p0_oraclek8_summary.json"
+    },
+    "transport_field_reground_fieldonly_k6matched_b12_clean_k6_dropnoopwg_retargeted_srcscore_task_pick001_stack005_advw4p0": {
+      "ci95_success": 0.023795944253316624,
+      "headline_ci95_success": 0.023795944253316624,
+      "headline_metric": "mean_success",
+      "headline_metric_label": "deployment",
+      "headline_std_success": 0.00957838356049757,
+      "headline_success": 0.3715942028985508,
+      "label": "Generator V1 utility-weighted tangents, expert anchor, advantage weight 4.0",
+      "mean_action_mse_to_best": 0.5142619989150999,
+      "mean_progress": 0.5871968711354795,
+      "mean_success": 0.3715942028985508,
+      "num_completed": 3,
+      "per_task_success": {
+        "LiftPegUpright-v1": {
+          "mean_num_groups": 102.0,
+          "mean_success": 0.2629578733692181,
+          "std_success": 0.07185893911242767
+        },
+        "PickCube-v1": {
+          "mean_num_groups": 196.66666666666666,
+          "mean_success": 0.325890594912334,
+          "std_success": 0.0315464330915602
+        },
+        "PullCube-v1": {
+          "mean_num_groups": 81.0,
+          "mean_success": 0.238350927824612,
+          "std_success": 0.005496676719939777
+        },
+        "PushCube-v1": {
+          "mean_num_groups": 102.0,
+          "mean_success": 0.811334612472173,
+          "std_success": 0.058125770572001835
+        },
+        "StackCube-v1": {
+          "mean_num_groups": 93.33333333333333,
+          "mean_success": 0.2161024161024161,
+          "std_success": 0.0500502261749912
+        }
+      },
+      "seed_action_mse_to_best": {
+        "0": 0.4974852632974153,
+        "1": 0.5146053172963793,
+        "2": 0.5306954161515054
+      },
+      "seed_progress": {
+        "0": 0.576676478124632,
+        "1": 0.5834855123139594,
+        "2": 0.601428622967847
+      },
+      "seed_success": {
+        "0": 0.3652173913043478,
+        "1": 0.36695652173913046,
+        "2": 0.3826086956521739
+      },
+      "selected_candidate_type_counts": {
+        "retrieval_residual_policy_residual": 584,
+        "retrieval_residual_residual_near_miss": 277,
+        "retrieval_residual_residual_near_miss+residual_wrong_gripper": 230,
+        "retrieval_residual_residual_no_op": 249,
+        "retrieval_residual_residual_wrong_gripper": 385
+      },
+      "selected_residual_scale_counts": {
+        "0.35": 1010,
+        "0.4": 189,
+        "0.45": 526
+      },
+      "selected_type_outcomes": {
+        "retrieval_residual_policy_residual": {
+          "count": 584.0,
+          "mean_progress": 0.5373461956767467,
+          "success_rate": 0.2910958904109589
+        },
+        "retrieval_residual_residual_near_miss": {
+          "count": 277.0,
+          "mean_progress": 0.725734414046386,
+          "success_rate": 0.5342960288808665
+        },
+        "retrieval_residual_residual_near_miss+residual_wrong_gripper": {
+          "count": 230.0,
+          "mean_progress": 0.7039336256844841,
+          "success_rate": 0.508695652173913
+        },
+        "retrieval_residual_residual_no_op": {
+          "count": 249.0,
+          "mean_progress": 0.5242827954681274,
+          "success_rate": 0.30120481927710846
+        },
+        "retrieval_residual_residual_wrong_gripper": {
+          "count": 385.0,
+          "mean_progress": 0.5340904980873713,
+          "success_rate": 0.34025974025974026
+        }
+      },
+      "source": "results/h16_transport_field_reground_fieldonly_k6clean_dropnoopwg_b12_v1_besttransport_margin0p00_k6_srcscore_task_pick001_stack005_advw4p0_summary.json",
+      "std_success": 0.00957838356049757
     },
     "transport_field_reground_fieldonly_k6matched_b12_clean_k6_dropnoopwg_retargeted_srcscore_task_pick001_stack005_oraclecal_train800_k4_r0025_t0025": {
       "ci95_success": 0.044132081897821224,
@@ -13108,12 +13606,414 @@
       "source": "results/h16_transport_field_reground_fieldonly_k6clean_dropnoopwg_b12_v1_besttransport_margin0p00_k6_srcscore_task_pick001_stack005_oraclecal_train800_k4_type005_summary.json",
       "std_success": 0.017764119937442965
     },
+    "transport_field_reground_fieldonly_k6matched_b12_clean_k6_dropnoopwg_retargeted_srcscore_task_pick001_stack005_policyanchor_advw1p0": {
+      "ci95_success": 0.027777495713921813,
+      "headline_ci95_success": 0.027777495713921813,
+      "headline_metric": "mean_success",
+      "headline_metric_label": "deployment",
+      "headline_std_success": 0.011181044360571532,
+      "headline_success": 0.383768115942029,
+      "label": "Generator V1 positive tangents, policy anchor, advantage weight 1.0",
+      "mean_action_mse_to_best": 0.51405627095154,
+      "mean_progress": 0.5947921582106231,
+      "mean_success": 0.383768115942029,
+      "num_completed": 3,
+      "per_task_success": {
+        "LiftPegUpright-v1": {
+          "mean_num_groups": 102.0,
+          "mean_success": 0.2698665217082789,
+          "std_success": 0.07386345661678752
+        },
+        "PickCube-v1": {
+          "mean_num_groups": 196.66666666666666,
+          "mean_success": 0.3328336317466753,
+          "std_success": 0.02615614287621478
+        },
+        "PullCube-v1": {
+          "mean_num_groups": 81.0,
+          "mean_success": 0.2636996531733374,
+          "std_success": 0.031723566913421014
+        },
+        "PushCube-v1": {
+          "mean_num_groups": 102.0,
+          "mean_success": 0.8178837147659562,
+          "std_success": 0.054232113490929594
+        },
+        "StackCube-v1": {
+          "mean_num_groups": 93.33333333333333,
+          "mean_success": 0.2412328412328412,
+          "std_success": 0.04734003304852265
+        }
+      },
+      "seed_action_mse_to_best": {
+        "0": 0.5013779968104285,
+        "1": 0.5149549319684181,
+        "2": 0.5258358840757738
+      },
+      "seed_progress": {
+        "0": 0.584225173432218,
+        "1": 0.5895823166373632,
+        "2": 0.6105689845622881
+      },
+      "seed_success": {
+        "0": 0.3791304347826087,
+        "1": 0.37565217391304345,
+        "2": 0.39652173913043476
+      },
+      "selected_candidate_type_counts": {
+        "retrieval_residual_policy_residual": 593,
+        "retrieval_residual_residual_near_miss": 237,
+        "retrieval_residual_residual_near_miss+residual_wrong_gripper": 219,
+        "retrieval_residual_residual_no_op": 254,
+        "retrieval_residual_residual_wrong_gripper": 422
+      },
+      "selected_residual_scale_counts": {
+        "0.35": 992,
+        "0.4": 204,
+        "0.45": 529
+      },
+      "selected_type_outcomes": {
+        "retrieval_residual_policy_residual": {
+          "count": 593.0,
+          "mean_progress": 0.5434360912911481,
+          "success_rate": 0.2917369308600337
+        },
+        "retrieval_residual_residual_near_miss": {
+          "count": 237.0,
+          "mean_progress": 0.752326833964463,
+          "success_rate": 0.5822784810126582
+        },
+        "retrieval_residual_residual_near_miss+residual_wrong_gripper": {
+          "count": 219.0,
+          "mean_progress": 0.7189138394325427,
+          "success_rate": 0.5296803652968036
+        },
+        "retrieval_residual_residual_no_op": {
+          "count": 254.0,
+          "mean_progress": 0.5431713150542845,
+          "success_rate": 0.3543307086614173
+        },
+        "retrieval_residual_residual_wrong_gripper": {
+          "count": 422.0,
+          "mean_progress": 0.5451416262288653,
+          "success_rate": 0.34360189573459715
+        }
+      },
+      "source": "results/h16_transport_field_reground_fieldonly_k6clean_dropnoopwg_b12_v1_besttransport_margin0p00_k6_srcscore_task_pick001_stack005_policyanchor_advw1p0_summary.json",
+      "std_success": 0.011181044360571532
+    },
     "transport_field_reground_fieldonly_k6matched_b12_clean_k6_dropnoopwg_retargeted_srcscore_task_pick001_stack005_policyanchor_advw2p0": {
+      "ci95_success": 0.032999037366698375,
+      "headline_ci95_success": 0.032999037366698375,
       "headline_metric": "mean_success",
       "headline_metric_label": "deployment",
+      "headline_std_success": 0.013282828101321283,
+      "headline_success": 0.3768115942028985,
       "label": "Generator V1 positive tangents, policy anchor, advantage weight 2.0",
+      "mean_action_mse_to_best": 0.5143569296768502,
+      "mean_progress": 0.5894178291635194,
+      "mean_success": 0.3768115942028985,
+      "num_completed": 3,
+      "per_task_success": {
+        "LiftPegUpright-v1": {
+          "mean_num_groups": 102.0,
+          "mean_success": 0.26639429948605664,
+          "std_success": 0.06792534648331451
+        },
+        "PickCube-v1": {
+          "mean_num_groups": 196.66666666666666,
+          "mean_success": 0.3274931590148981,
+          "std_success": 0.03420004779247759
+        },
+        "PullCube-v1": {
+          "mean_num_groups": 81.0,
+          "mean_success": 0.2500303789777474,
+          "std_success": 0.023077786338402385
+        },
+        "PushCube-v1": {
+          "mean_num_groups": 102.0,
+          "mean_success": 0.8146349425051763,
+          "std_success": 0.06171557177404845
+        },
+        "StackCube-v1": {
+          "mean_num_groups": 93.33333333333333,
+          "mean_success": 0.22687682687682686,
+          "std_success": 0.04846784140995916
+        }
+      },
+      "seed_action_mse_to_best": {
+        "0": 0.5005510983859067,
+        "1": 0.5151173928242339,
+        "2": 0.52740229782041
+      },
+      "seed_progress": {
+        "0": 0.5787463493072985,
+        "1": 0.582736657930781,
+        "2": 0.6067704802524785
+      },
+      "seed_success": {
+        "0": 0.3739130434782609,
+        "1": 0.3652173913043478,
+        "2": 0.391304347826087
+      },
+      "selected_candidate_type_counts": {
+        "retrieval_residual_policy_residual": 584,
+        "retrieval_residual_residual_near_miss": 269,
+        "retrieval_residual_residual_near_miss+residual_wrong_gripper": 227,
+        "retrieval_residual_residual_no_op": 252,
+        "retrieval_residual_residual_wrong_gripper": 393
+      },
+      "selected_residual_scale_counts": {
+        "0.35": 1007,
+        "0.4": 191,
+        "0.45": 527
+      },
+      "selected_type_outcomes": {
+        "retrieval_residual_policy_residual": {
+          "count": 584.0,
+          "mean_progress": 0.5396331625987381,
+          "success_rate": 0.2910958904109589
+        },
+        "retrieval_residual_residual_near_miss": {
+          "count": 269.0,
+          "mean_progress": 0.7359185712931324,
+          "success_rate": 0.550185873605948
+        },
+        "retrieval_residual_residual_near_miss+residual_wrong_gripper": {
+          "count": 227.0,
+          "mean_progress": 0.7157547649340745,
+          "success_rate": 0.5330396475770925
+        },
+        "retrieval_residual_residual_no_op": {
+          "count": 252.0,
+          "mean_progress": 0.5248836046774823,
+          "success_rate": 0.31746031746031744
+        },
+        "retrieval_residual_residual_wrong_gripper": {
+          "count": 393.0,
+          "mean_progress": 0.5315289889384089,
+          "success_rate": 0.3333333333333333
+        }
+      },
+      "source": "results/h16_transport_field_reground_fieldonly_k6clean_dropnoopwg_b12_v1_besttransport_margin0p00_k6_srcscore_task_pick001_stack005_policyanchor_advw2p0_summary.json",
+      "std_success": 0.013282828101321283
+    },
+    "transport_field_reground_fieldonly_k6matched_b12_clean_k6_dropnoopwg_retargeted_srcscore_task_pick001_stack005_policyanchor_advw2p0_gate0": {
+      "ci95_success": 0.04778760648872724,
+      "headline_ci95_success": 0.04778760648872724,
+      "headline_metric": "mean_success",
+      "headline_metric_label": "deployment",
+      "headline_std_success": 0.019235547852796592,
+      "headline_success": 0.36927536231884056,
+      "label": "Generator V1 positive tangents, policy anchor, source advantage gate >= 0.0",
+      "mean_action_mse_to_best": 0.45602177684910705,
+      "mean_progress": 0.5815420930200059,
+      "mean_success": 0.36927536231884056,
+      "num_completed": 3,
+      "per_task_success": {
+        "LiftPegUpright-v1": {
+          "mean_num_groups": 102.0,
+          "mean_success": 0.26639429948605664,
+          "std_success": 0.06792534648331451
+        },
+        "PickCube-v1": {
+          "mean_num_groups": 196.66666666666666,
+          "mean_success": 0.32465893607197954,
+          "std_success": 0.018763488756388967
+        },
+        "PullCube-v1": {
+          "mean_num_groups": 81.0,
+          "mean_success": 0.23624718361560468,
+          "std_success": 0.04766015797840824
+        },
+        "PushCube-v1": {
+          "mean_num_groups": 102.0,
+          "mean_success": 0.8055743757239439,
+          "std_success": 0.04486214967261923
+        },
+        "StackCube-v1": {
+          "mean_num_groups": 93.33333333333333,
+          "mean_success": 0.2081844081844082,
+          "std_success": 0.07436028694013369
+        }
+      },
+      "seed_action_mse_to_best": {
+        "0": 0.43498127313573726,
+        "1": 0.4576410755636575,
+        "2": 0.4754429818479263
+      },
+      "seed_progress": {
+        "0": 0.5694765353704924,
+        "1": 0.5735185124835824,
+        "2": 0.6016312312059429
+      },
+      "seed_success": {
+        "0": 0.36695652173913046,
+        "1": 0.35130434782608694,
+        "2": 0.38956521739130434
+      },
+      "selected_candidate_type_counts": {
+        "retrieval_residual_policy_residual": 819,
+        "retrieval_residual_residual_near_miss": 479,
+        "retrieval_residual_residual_near_miss+residual_wrong_gripper": 80,
+        "retrieval_residual_residual_no_op": 174,
+        "retrieval_residual_residual_no_op+residual_near_miss": 14,
+        "retrieval_residual_residual_wrong_gripper": 143,
+        "retrieval_residual_residual_wrong_gripper+residual_near_miss": 4,
+        "retrieval_residual_residual_wrong_gripper+residual_no_op": 12
+      },
+      "selected_residual_scale_counts": {
+        "0.35": 1208,
+        "0.4": 97,
+        "0.45": 420
+      },
+      "selected_type_outcomes": {
+        "retrieval_residual_policy_residual": {
+          "count": 819.0,
+          "mean_progress": 0.5104973517751755,
+          "success_rate": 0.2857142857142857
+        },
+        "retrieval_residual_residual_near_miss": {
+          "count": 479.0,
+          "mean_progress": 0.6349816032148711,
+          "success_rate": 0.4154488517745303
+        },
+        "retrieval_residual_residual_near_miss+residual_wrong_gripper": {
+          "count": 80.0,
+          "mean_progress": 0.8649393017403781,
+          "success_rate": 0.8
+        },
+        "retrieval_residual_residual_no_op": {
+          "count": 174.0,
+          "mean_progress": 0.6662687381793713,
+          "success_rate": 0.42528735632183906
+        },
+        "retrieval_residual_residual_no_op+residual_near_miss": {
+          "count": 14.0,
+          "mean_progress": 0.7973395671163287,
+          "success_rate": 0.5
+        },
+        "retrieval_residual_residual_wrong_gripper": {
+          "count": 143.0,
+          "mean_progress": 0.5357520567809156,
+          "success_rate": 0.3776223776223776
+        },
+        "retrieval_residual_residual_wrong_gripper+residual_near_miss": {
+          "count": 4.0,
+          "mean_progress": 0.5595042705535889,
+          "success_rate": 0.25
+        },
+        "retrieval_residual_residual_wrong_gripper+residual_no_op": {
+          "count": 12.0,
+          "mean_progress": 0.48061430764695007,
+          "success_rate": 0.3333333333333333
+        }
+      },
+      "source": "results/h16_transport_field_reground_fieldonly_k6clean_dropnoopwg_b12_v1_besttransport_margin0p00_k6_srcscore_task_pick001_stack005_policyanchor_advw2p0_gate0_summary.json",
+      "std_success": 0.019235547852796592
+    },
+    "transport_field_reground_fieldonly_k6matched_b12_clean_k6_dropnoopwg_retargeted_srcscore_task_pick001_stack005_policyanchor_advw2p0_oraclek8": {
+      "headline_metric": "mean_candidate_oracle_success_rate",
+      "headline_metric_label": "candidate-oracle",
+      "label": "Generator V1 policy-anchor advantage weight 2.0, candidate-oracle K8",
       "missing": true,
-      "source": "results/h16_transport_field_reground_fieldonly_k6clean_dropnoopwg_b12_v1_besttransport_margin0p00_k6_srcscore_task_pick001_stack005_policyanchor_advw2p0_summary.json"
+      "source": "results/h16_transport_field_reground_fieldonly_k6clean_dropnoopwg_b12_v1_besttransport_margin0p00_k6_srcscore_task_pick001_stack005_policyanchor_advw2p0_oraclek8_summary.json"
+    },
+    "transport_field_reground_fieldonly_k6matched_b12_clean_k6_dropnoopwg_retargeted_srcscore_task_pick001_stack005_policyanchor_advw4p0": {
+      "ci95_success": 0.023795944253316624,
+      "headline_ci95_success": 0.023795944253316624,
+      "headline_metric": "mean_success",
+      "headline_metric_label": "deployment",
+      "headline_std_success": 0.00957838356049757,
+      "headline_success": 0.3715942028985508,
+      "label": "Generator V1 positive tangents, policy anchor, advantage weight 4.0",
+      "mean_action_mse_to_best": 0.5142619989150999,
+      "mean_progress": 0.5871968711354795,
+      "mean_success": 0.3715942028985508,
+      "num_completed": 3,
+      "per_task_success": {
+        "LiftPegUpright-v1": {
+          "mean_num_groups": 102.0,
+          "mean_success": 0.2629578733692181,
+          "std_success": 0.07185893911242767
+        },
+        "PickCube-v1": {
+          "mean_num_groups": 196.66666666666666,
+          "mean_success": 0.325890594912334,
+          "std_success": 0.0315464330915602
+        },
+        "PullCube-v1": {
+          "mean_num_groups": 81.0,
+          "mean_success": 0.238350927824612,
+          "std_success": 0.005496676719939777
+        },
+        "PushCube-v1": {
+          "mean_num_groups": 102.0,
+          "mean_success": 0.811334612472173,
+          "std_success": 0.058125770572001835
+        },
+        "StackCube-v1": {
+          "mean_num_groups": 93.33333333333333,
+          "mean_success": 0.2161024161024161,
+          "std_success": 0.0500502261749912
+        }
+      },
+      "seed_action_mse_to_best": {
+        "0": 0.4974852632974153,
+        "1": 0.5146053172963793,
+        "2": 0.5306954161515054
+      },
+      "seed_progress": {
+        "0": 0.576676478124632,
+        "1": 0.5834855123139594,
+        "2": 0.601428622967847
+      },
+      "seed_success": {
+        "0": 0.3652173913043478,
+        "1": 0.36695652173913046,
+        "2": 0.3826086956521739
+      },
+      "selected_candidate_type_counts": {
+        "retrieval_residual_policy_residual": 584,
+        "retrieval_residual_residual_near_miss": 277,
+        "retrieval_residual_residual_near_miss+residual_wrong_gripper": 230,
+        "retrieval_residual_residual_no_op": 249,
+        "retrieval_residual_residual_wrong_gripper": 385
+      },
+      "selected_residual_scale_counts": {
+        "0.35": 1010,
+        "0.4": 189,
+        "0.45": 526
+      },
+      "selected_type_outcomes": {
+        "retrieval_residual_policy_residual": {
+          "count": 584.0,
+          "mean_progress": 0.5373461956767467,
+          "success_rate": 0.2910958904109589
+        },
+        "retrieval_residual_residual_near_miss": {
+          "count": 277.0,
+          "mean_progress": 0.725734414046386,
+          "success_rate": 0.5342960288808665
+        },
+        "retrieval_residual_residual_near_miss+residual_wrong_gripper": {
+          "count": 230.0,
+          "mean_progress": 0.7039336256844841,
+          "success_rate": 0.508695652173913
+        },
+        "retrieval_residual_residual_no_op": {
+          "count": 249.0,
+          "mean_progress": 0.5242827954681274,
+          "success_rate": 0.30120481927710846
+        },
+        "retrieval_residual_residual_wrong_gripper": {
+          "count": 385.0,
+          "mean_progress": 0.5340904980873713,
+          "success_rate": 0.34025974025974026
+        }
+      },
+      "source": "results/h16_transport_field_reground_fieldonly_k6clean_dropnoopwg_b12_v1_besttransport_margin0p00_k6_srcscore_task_pick001_stack005_policyanchor_advw4p0_summary.json",
+      "std_success": 0.00957838356049757
     },
     "transport_field_reground_fieldonly_k6matched_b12_clean_k6_dropnoopwg_retargeted_srcscore_task_pick001_stack005_scaleoracle_trainall_k4_scale0025": {
       "ci95_success": 0.043707043917985716,

workspace/results/paper_analysis.md

@@ -1,6 +1,6 @@
 # Paper Analysis
 
-Generated: `2026-07-02T18:32:12+00:00`
+Generated: `2026-07-02T19:30:26+00:00`
 
 ## Main Seed Statistics
 
@@ -72,9 +72,18 @@ Generated: `2026-07-02T18:32:12+00:00`
 | transport_field_reground_fieldonly_k6matched_b12_clean_k6_dropnoopwg_retargeted_srcscore001_metric_taskrel | K6-matched transported residual field re-grounding, source-score 0.01, task-relative retrieval chart | 3 | deployment | 37.22% +/- 1.66 | 37.22% | +/- 4.12 | 58.49% | 0.512 | +7.48 pp |
 | transport_field_reground_fieldonly_k6matched_b12_clean_k6_dropnoopwg_retargeted_srcscore001_metric_taskrelz | K6-matched transported residual field re-grounding, source-score 0.01, task-relative z-score retrieval chart | 3 | deployment | 36.99% +/- 2.02 | 36.99% | +/- 5.01 | 58.49% | 0.511 | +7.25 pp |
 | transport_field_reground_fieldonly_k6matched_b12_clean_k6_dropnoopwg_retargeted_srcscore_task_pick001_stack005 | K6-matched transported residual field re-grounding, task-conditioned source-score prior | 3 | deployment | 38.90% +/- 1.86 | 38.90% | +/- 4.62 | 59.96% | 0.513 | +9.16 pp |
-| transport_field_reground_fieldonly_k6matched_b12_clean_k6_dropnoopwg_retargeted_srcscore_task_pick001_stack005_advw1p0 | Generator V1 utility-weighted tangents, expert anchor, advantage weight 1.0 | 0 | deployment | missing | missing | missing | missing | missing | missing |
-| transport_field_reground_fieldonly_k6matched_b12_clean_k6_dropnoopwg_retargeted_srcscore_task_pick001_stack005_advw2p0 | Generator V1 utility-weighted tangents, expert anchor, advantage weight 2.0 | 0 | deployment | missing | missing | missing | missing | missing | missing |
-| transport_field_reground_fieldonly_k6matched_b12_clean_k6_dropnoopwg_retargeted_srcscore_task_pick001_stack005_policyanchor_advw2p0 | Generator V1 positive tangents, policy anchor, advantage weight 2.0 | 0 | deployment | missing | missing | missing | missing | missing | missing |
+| transport_field_reground_fieldonly_k6matched_b12_clean_k6_dropnoopwg_retargeted_srcscore_task_pick001_stack005_advw1p0 | Generator V1 utility-weighted tangents, expert anchor, advantage weight 1.0 | 3 | deployment | 38.38% +/- 1.12 | 38.38% | +/- 2.78 | 59.48% | 0.514 | +8.64 pp |
+| transport_field_reground_fieldonly_k6matched_b12_clean_k6_dropnoopwg_retargeted_srcscore_task_pick001_stack005_advw2p0 | Generator V1 utility-weighted tangents, expert anchor, advantage weight 2.0 | 3 | deployment | 37.68% +/- 1.33 | 37.68% | +/- 3.30 | 58.94% | 0.514 | +7.94 pp |
+| transport_field_reground_fieldonly_k6matched_b12_clean_k6_dropnoopwg_retargeted_srcscore_task_pick001_stack005_policyanchor_advw2p0 | Generator V1 positive tangents, policy anchor, advantage weight 2.0 | 3 | deployment | 37.68% +/- 1.33 | 37.68% | +/- 3.30 | 58.94% | 0.514 | +7.94 pp |
+| transport_field_reground_fieldonly_k6matched_b12_clean_k6_dropnoopwg_retargeted_srcscore_task_pick001_stack005_advw2p0_gate0 | Generator V1 utility-weighted tangents, expert anchor, source advantage gate >= 0.0 | 3 | deployment | 36.93% +/- 1.92 | 36.93% | +/- 4.78 | 58.15% | 0.456 | +7.19 pp |
+| transport_field_reground_fieldonly_k6matched_b12_clean_k6_dropnoopwg_retargeted_srcscore_task_pick001_stack005_policyanchor_advw2p0_gate0 | Generator V1 positive tangents, policy anchor, source advantage gate >= 0.0 | 3 | deployment | 36.93% +/- 1.92 | 36.93% | +/- 4.78 | 58.15% | 0.456 | +7.19 pp |
+| transport_field_reground_fieldonly_k6matched_b12_clean_k6_dropnoopwg_retargeted_srcscore_task_pick001_stack005_advw1p0_oraclek8 | Generator V1 expert-anchor advantage weight 1.0, candidate-oracle K8 | 0 | candidate-oracle | missing | missing | missing | missing | missing | missing |
+| transport_field_reground_fieldonly_k6matched_b12_clean_k6_dropnoopwg_retargeted_srcscore_task_pick001_stack005_advw2p0_oraclek8 | Generator V1 expert-anchor advantage weight 2.0, candidate-oracle K8 | 0 | candidate-oracle | missing | missing | missing | missing | missing | missing |
+| transport_field_reground_fieldonly_k6matched_b12_clean_k6_dropnoopwg_retargeted_srcscore_task_pick001_stack005_policyanchor_advw2p0_oraclek8 | Generator V1 policy-anchor advantage weight 2.0, candidate-oracle K8 | 0 | candidate-oracle | missing | missing | missing | missing | missing | missing |
+| transport_field_reground_fieldonly_k6matched_b12_clean_k6_dropnoopwg_retargeted_srcscore_task_pick001_stack005_advw0p5 | Generator V1 utility-weighted tangents, expert anchor, advantage weight 0.5 | 3 | deployment | 38.72% +/- 1.74 | 38.72% | +/- 4.33 | 59.71% | 0.514 | +8.99 pp |
+| transport_field_reground_fieldonly_k6matched_b12_clean_k6_dropnoopwg_retargeted_srcscore_task_pick001_stack005_advw4p0 | Generator V1 utility-weighted tangents, expert anchor, advantage weight 4.0 | 3 | deployment | 37.16% +/- 0.96 | 37.16% | +/- 2.38 | 58.72% | 0.514 | +7.42 pp |
+| transport_field_reground_fieldonly_k6matched_b12_clean_k6_dropnoopwg_retargeted_srcscore_task_pick001_stack005_policyanchor_advw1p0 | Generator V1 positive tangents, policy anchor, advantage weight 1.0 | 3 | deployment | 38.38% +/- 1.12 | 38.38% | +/- 2.78 | 59.48% | 0.514 | +8.64 pp |
+| transport_field_reground_fieldonly_k6matched_b12_clean_k6_dropnoopwg_retargeted_srcscore_task_pick001_stack005_policyanchor_advw4p0 | Generator V1 positive tangents, policy anchor, advantage weight 4.0 | 3 | deployment | 37.16% +/- 0.96 | 37.16% | +/- 2.38 | 58.72% | 0.514 | +7.42 pp |
 | transport_field_reground_fieldonly_k6matched_b12_clean_k6_dropnoopwg_retargeted_srcscore_task_pick001_stack005_scaleoracle_trainall_k4_scale0025 | K6-matched transported residual field re-grounding, all-train tangent-length prior 0.025 | 3 | deployment | 38.84% +/- 1.76 | 38.84% | +/- 4.37 | 59.94% | 0.513 | +9.10 pp |
 | transport_field_reground_fieldonly_k6matched_b12_clean_k6_dropnoopwg_retargeted_srcscore_task_pick001_stack005_scaleoracle_trainall_k4_scale005 | K6-matched transported residual field re-grounding, all-train tangent-length prior 0.05 | 3 | deployment | 38.78% +/- 1.66 | 38.78% | +/- 4.12 | 59.89% | 0.513 | +9.04 pp |
 | transport_field_reground_fieldonly_k6matched_b12_clean_k6_dropnoopwg_retargeted_srcscore_task_pick001_stack005_scaleoracle_trainall_k4_scale010 | K6-matched transported residual field re-grounding, all-train tangent-length prior 0.10 | 3 | deployment | 38.67% +/- 1.62 | 38.67% | +/- 4.01 | 59.82% | 0.513 | +8.93 pp |

workspace/results/v1_generator_decision.json

@@ -0,0 +1,68 @@
+{
+  "generated_utc": "2026-07-02T19:13:00.427335+00:00",
+  "round": 0,
+  "baseline_success": 0.3889855072463768,
+  "complete": true,
+  "summaries": [
+    {
+      "key": "advw1p0",
+      "label": "Generator V1 utility-weighted tangents, expert anchor, advantage weight 1.0",
+      "num_completed": 3,
+      "mean_success": 0.383768115942029,
+      "delta_vs_baseline": -0.00521739130434784,
+      "summary_path": "results/h16_transport_field_reground_fieldonly_k6clean_dropnoopwg_b12_v1_besttransport_margin0p00_k6_srcscore_task_pick001_stack005_advw1p0_summary.json"
+    },
+    {
+      "key": "advw2p0",
+      "label": "Generator V1 utility-weighted tangents, expert anchor, advantage weight 2.0",
+      "num_completed": 3,
+      "mean_success": 0.37681159420289856,
+      "delta_vs_baseline": -0.012173913043478257,
+      "summary_path": "results/h16_transport_field_reground_fieldonly_k6clean_dropnoopwg_b12_v1_besttransport_margin0p00_k6_srcscore_task_pick001_stack005_advw2p0_summary.json"
+    },
+    {
+      "key": "policyanchor_advw2p0",
+      "label": "Generator V1 positive tangents, policy anchor, advantage weight 2.0",
+      "num_completed": 3,
+      "mean_success": 0.37681159420289856,
+      "delta_vs_baseline": -0.012173913043478257,
+      "summary_path": "results/h16_transport_field_reground_fieldonly_k6clean_dropnoopwg_b12_v1_besttransport_margin0p00_k6_srcscore_task_pick001_stack005_policyanchor_advw2p0_summary.json"
+    }
+  ],
+  "best_key": "advw1p0",
+  "best_success": 0.383768115942029,
+  "best_delta_vs_baseline": -0.00521739130434784,
+  "recommendation": "submit_wider_advantage_weight_support_sweep",
+  "submitted": [
+    {
+      "key": "advw0p5",
+      "eval_job": "15069074",
+      "summary_job": "15069075"
+    },
+    {
+      "key": "advw4p0",
+      "eval_job": "15069076",
+      "summary_job": "15069077"
+    },
+    {
+      "key": "policyanchor_advw1p0",
+      "eval_job": "15069078",
+      "summary_job": "15069079"
+    },
+    {
+      "key": "policyanchor_advw4p0",
+      "eval_job": "15069080",
+      "summary_job": "15069081"
+    },
+    {
+      "key": "advw2p0_gate0",
+      "eval_job": "15069082",
+      "summary_job": "15069083"
+    },
+    {
+      "key": "policyanchor_advw2p0_gate0",
+      "eval_job": "15069084",
+      "summary_job": "15069085"
+    }
+  ]
+}

workspace/results/v1_generator_decision.md

@@ -0,0 +1,20 @@
+# Generator V1 Decision
+
+Generated: `2026-07-02T19:13:00.427335+00:00`
+Baseline clean success: 38.90%
+Recommendation: `submit_wider_advantage_weight_support_sweep`
+
+| key | completed | success | delta |
+|---|---:|---:|---:|
+| advw1p0 | 3 | 38.38% | -0.52% |
+| advw2p0 | 3 | 37.68% | -1.22% |
+| policyanchor_advw2p0 | 3 | 37.68% | -1.22% |
+
+Submitted follow-up jobs:
+
+- `advw0p5`: eval `15069074`, summary `15069075`
+- `advw4p0`: eval `15069076`, summary `15069077`
+- `policyanchor_advw1p0`: eval `15069078`, summary `15069079`
+- `policyanchor_advw4p0`: eval `15069080`, summary `15069081`
+- `advw2p0_gate0`: eval `15069082`, summary `15069083`
+- `policyanchor_advw2p0_gate0`: eval `15069084`, summary `15069085`

workspace/scripts/advance_v1_generator.py

@@ -0,0 +1,665 @@
+#!/usr/bin/env python3
+from __future__ import annotations
+
+import argparse
+import json
+import math
+import os
+import statistics
+import subprocess
+from collections import Counter
+from dataclasses import dataclass
+from datetime import datetime, timezone
+from pathlib import Path
+from typing import Any
+
+
+DEFAULT_OBJECTIVE = "transport_field_reground_fieldonly_k6clean_dropnoopwg_b12_v1"
+BEST_CLEAN_SUMMARY = (
+    "h16_transport_field_reground_fieldonly_k6clean_dropnoopwg_b12_v1_"
+    "besttransport_margin0p00_k6_srcscore_task_pick001_stack005_summary.json"
+)
+BEST_CLEAN_SUCCESS = 0.3889855072463768
+SOURCE_SCORE_MAP = "manifests/source_score_bonus_pick001_stack005.json"
+DATASET = "/scratch/{user}/dovla/experiments/six_task_h16_collection"
+RUN_ROOT = "/scratch/{user}/dovla/experiments/dovla_h16_policy_ckpt_runs"
+EXCLUDE_TYPES = (
+    "residual_random_negative:"
+    "residual_wrong_direction:"
+    "residual_near_miss+residual_no_op:"
+    "residual_no_op+residual_wrong_gripper"
+)
+
+
+@dataclass(frozen=True)
+class V1Config:
+    key: str
+    label: str
+    out_name: str
+    summary_tag: str
+    anchor: str
+    advantage_weight: float
+    min_source_advantage: float = -1.0e9
+    candidate_oracle_rollouts: int = 0
+
+
+INITIAL_CONFIGS = (
+    V1Config(
+        key="advw1p0",
+        label="Generator V1 utility-weighted tangents, expert anchor, advantage weight 1.0",
+        out_name=(
+            "policy_rollout_fieldonly_k6clean_dropnoopwg_b12_v1_"
+            "besttransport_margin0p00_k6_srcscore_task_pick001_stack005_advw1p0.json"
+        ),
+        summary_tag=(
+            "transport_field_reground_fieldonly_k6clean_dropnoopwg_b12_v1_"
+            "besttransport_margin0p00_k6_srcscore_task_pick001_stack005_advw1p0"
+        ),
+        anchor="expert",
+        advantage_weight=1.0,
+    ),
+    V1Config(
+        key="advw2p0",
+        label="Generator V1 utility-weighted tangents, expert anchor, advantage weight 2.0",
+        out_name=(
+            "policy_rollout_fieldonly_k6clean_dropnoopwg_b12_v1_"
+            "besttransport_margin0p00_k6_srcscore_task_pick001_stack005_advw2p0.json"
+        ),
+        summary_tag=(
+            "transport_field_reground_fieldonly_k6clean_dropnoopwg_b12_v1_"
+            "besttransport_margin0p00_k6_srcscore_task_pick001_stack005_advw2p0"
+        ),
+        anchor="expert",
+        advantage_weight=2.0,
+    ),
+    V1Config(
+        key="policyanchor_advw2p0",
+        label="Generator V1 positive tangents, policy anchor, advantage weight 2.0",
+        out_name=(
+            "policy_rollout_fieldonly_k6clean_dropnoopwg_b12_v1_"
+            "besttransport_margin0p00_k6_srcscore_task_pick001_stack005_policyanchor_advw2p0.json"
+        ),
+        summary_tag=(
+            "transport_field_reground_fieldonly_k6clean_dropnoopwg_b12_v1_"
+            "besttransport_margin0p00_k6_srcscore_task_pick001_stack005_policyanchor_advw2p0"
+        ),
+        anchor="policy",
+        advantage_weight=2.0,
+    ),
+)
+
+NEXT_SWEEP_CONFIGS = (
+    V1Config(
+        key="advw0p5",
+        label="Generator V1 utility-weighted tangents, expert anchor, advantage weight 0.5",
+        out_name=(
+            "policy_rollout_fieldonly_k6clean_dropnoopwg_b12_v1_"
+            "besttransport_margin0p00_k6_srcscore_task_pick001_stack005_advw0p5.json"
+        ),
+        summary_tag=(
+            "transport_field_reground_fieldonly_k6clean_dropnoopwg_b12_v1_"
+            "besttransport_margin0p00_k6_srcscore_task_pick001_stack005_advw0p5"
+        ),
+        anchor="expert",
+        advantage_weight=0.5,
+    ),
+    V1Config(
+        key="advw4p0",
+        label="Generator V1 utility-weighted tangents, expert anchor, advantage weight 4.0",
+        out_name=(
+            "policy_rollout_fieldonly_k6clean_dropnoopwg_b12_v1_"
+            "besttransport_margin0p00_k6_srcscore_task_pick001_stack005_advw4p0.json"
+        ),
+        summary_tag=(
+            "transport_field_reground_fieldonly_k6clean_dropnoopwg_b12_v1_"
+            "besttransport_margin0p00_k6_srcscore_task_pick001_stack005_advw4p0"
+        ),
+        anchor="expert",
+        advantage_weight=4.0,
+    ),
+    V1Config(
+        key="policyanchor_advw1p0",
+        label="Generator V1 positive tangents, policy anchor, advantage weight 1.0",
+        out_name=(
+            "policy_rollout_fieldonly_k6clean_dropnoopwg_b12_v1_"
+            "besttransport_margin0p00_k6_srcscore_task_pick001_stack005_policyanchor_advw1p0.json"
+        ),
+        summary_tag=(
+            "transport_field_reground_fieldonly_k6clean_dropnoopwg_b12_v1_"
+            "besttransport_margin0p00_k6_srcscore_task_pick001_stack005_policyanchor_advw1p0"
+        ),
+        anchor="policy",
+        advantage_weight=1.0,
+    ),
+    V1Config(
+        key="policyanchor_advw4p0",
+        label="Generator V1 positive tangents, policy anchor, advantage weight 4.0",
+        out_name=(
+            "policy_rollout_fieldonly_k6clean_dropnoopwg_b12_v1_"
+            "besttransport_margin0p00_k6_srcscore_task_pick001_stack005_policyanchor_advw4p0.json"
+        ),
+        summary_tag=(
+            "transport_field_reground_fieldonly_k6clean_dropnoopwg_b12_v1_"
+            "besttransport_margin0p00_k6_srcscore_task_pick001_stack005_policyanchor_advw4p0"
+        ),
+        anchor="policy",
+        advantage_weight=4.0,
+    ),
+    V1Config(
+        key="advw2p0_gate0",
+        label="Generator V1 utility-weighted tangents, expert anchor, positive-source gate",
+        out_name=(
+            "policy_rollout_fieldonly_k6clean_dropnoopwg_b12_v1_"
+            "besttransport_margin0p00_k6_srcscore_task_pick001_stack005_advw2p0_gate0.json"
+        ),
+        summary_tag=(
+            "transport_field_reground_fieldonly_k6clean_dropnoopwg_b12_v1_"
+            "besttransport_margin0p00_k6_srcscore_task_pick001_stack005_advw2p0_gate0"
+        ),
+        anchor="expert",
+        advantage_weight=2.0,
+        min_source_advantage=0.0,
+    ),
+    V1Config(
+        key="policyanchor_advw2p0_gate0",
+        label="Generator V1 positive tangents, policy anchor, positive-source gate",
+        out_name=(
+            "policy_rollout_fieldonly_k6clean_dropnoopwg_b12_v1_"
+            "besttransport_margin0p00_k6_srcscore_task_pick001_stack005_policyanchor_advw2p0_gate0.json"
+        ),
+        summary_tag=(
+            "transport_field_reground_fieldonly_k6clean_dropnoopwg_b12_v1_"
+            "besttransport_margin0p00_k6_srcscore_task_pick001_stack005_policyanchor_advw2p0_gate0"
+        ),
+        anchor="policy",
+        advantage_weight=2.0,
+        min_source_advantage=0.0,
+    ),
+)
+
+
+def main() -> int:
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--project-dir", type=Path, default=Path.cwd())
+    parser.add_argument("--run-root", type=Path, default=None)
+    parser.add_argument("--dataset", type=Path, default=None)
+    parser.add_argument("--objective", default=DEFAULT_OBJECTIVE)
+    parser.add_argument("--results-dir", type=Path, default=Path("results"))
+    parser.add_argument("--submit-next", action="store_true")
+    parser.add_argument("--dry-run", action="store_true")
+    parser.add_argument("--round", type=int, default=int(os.environ.get("ADVANCE_ROUND", "0")))
+    args = parser.parse_args()
+
+    user = os.environ.get("USER", "knguy52")
+    run_root = args.run_root or Path(RUN_ROOT.format(user=user))
+    dataset = args.dataset or Path(DATASET.format(user=user))
+    project_dir = args.project_dir.resolve()
+    results_dir = args.results_dir
+    results_dir.mkdir(parents=True, exist_ok=True)
+
+    summaries = [
+        _summarize_config(config, run_root=run_root, objective=args.objective, results_dir=results_dir)
+        for config in INITIAL_CONFIGS
+    ]
+    baseline_success = _baseline_success(results_dir)
+    complete = [item for item in summaries if item["num_completed"] == 3]
+    best = max(complete, key=lambda item: float(item["mean_success"]), default=None)
+    decision = _decision_payload(
+        summaries,
+        best=best,
+        baseline_success=baseline_success,
+        round_index=args.round,
+    )
+    decision_path = results_dir / "v1_generator_decision.json"
+    decision_md_path = results_dir / "v1_generator_decision.md"
+    decision_path.write_text(json.dumps(decision, indent=2))
+    decision_md_path.write_text(_render_decision_markdown(decision))
+
+    if complete:
+        _run_build_paper_analysis(project_dir, dry_run=args.dry_run)
+
+    if args.submit_next and not decision["complete"]:
+        submitted_missing = _maybe_resubmit_missing(
+            summaries,
+            INITIAL_CONFIGS,
+            project_dir=project_dir,
+            run_root=run_root,
+            dataset=dataset,
+            objective=args.objective,
+            dry_run=args.dry_run,
+        )
+        decision["submitted_missing"] = submitted_missing
+        decision_path.write_text(json.dumps(decision, indent=2))
+        decision_md_path.write_text(_render_decision_markdown(decision))
+    elif args.submit_next and complete:
+        submitted = _maybe_submit_next(
+            decision,
+            project_dir=project_dir,
+            run_root=run_root,
+            dataset=dataset,
+            objective=args.objective,
+            dry_run=args.dry_run,
+        )
+        decision["submitted"] = submitted
+        decision_path.write_text(json.dumps(decision, indent=2))
+        decision_md_path.write_text(_render_decision_markdown(decision))
+
+    print(json.dumps(decision, indent=2))
+    return 0
+
+
+def _load_json(path: Path) -> dict[str, Any]:
+    return json.loads(path.read_text())
+
+
+def _mean(values: list[float]) -> float:
+    return statistics.mean(values) if values else 0.0
+
+
+def _std(values: list[float]) -> float:
+    return statistics.stdev(values) if len(values) > 1 else 0.0
+
+
+def _baseline_success(results_dir: Path) -> float:
+    path = results_dir / BEST_CLEAN_SUMMARY
+    if path.exists():
+        value = _load_json(path).get("mean_success")
+        if isinstance(value, (int, float)) and math.isfinite(float(value)):
+            return float(value)
+    return BEST_CLEAN_SUCCESS
+
+
+def _summarize_config(
+    config: V1Config,
+    *,
+    run_root: Path,
+    objective: str,
+    results_dir: Path,
+) -> dict[str, Any]:
+    rows = []
+    base_dir = run_root / objective
+    for result_path in sorted(base_dir.glob(f"seed_*/{config.out_name}")):
+        raw = _load_json(result_path)
+        seed = int(result_path.parent.name.split("_")[-1])
+        selected_scale_counts = Counter(
+            str(row.get("selected_residual_scale"))
+            for row in raw.get("rows", [])
+            if row.get("selected_residual_scale") is not None
+        )
+        rows.append(
+            {
+                "seed": seed,
+                "path": str(result_path),
+                "num_groups": raw.get("num_groups", 0),
+                "selection_mode": raw.get("selection_mode"),
+                "num_candidates": raw.get("num_candidates"),
+                "candidate_sigma": raw.get("candidate_sigma", 0.0),
+                "selection_margin": raw.get("selection_margin", 0.0),
+                "retrieval_neighbors": raw.get("retrieval_neighbors", 0),
+                "retrieval_metric": raw.get("retrieval_metric", "none"),
+                "retrieval_residual_anchor": raw.get("retrieval_residual_anchor", "none"),
+                "retrieval_residual_direction": raw.get("retrieval_residual_direction", "none"),
+                "retrieval_residual_reduce": raw.get("retrieval_residual_reduce", "none"),
+                "retrieval_residual_scales": raw.get("retrieval_residual_scales", []),
+                "retrieval_residual_source_score_bonus_by_task": raw.get(
+                    "retrieval_residual_source_score_bonus_by_task", {}
+                ),
+                "retrieval_residual_source_advantage_weight_scale": raw.get(
+                    "retrieval_residual_source_advantage_weight_scale", 0.0
+                ),
+                "retrieval_residual_min_source_advantage": raw.get(
+                    "retrieval_residual_min_source_advantage", -1.0e9
+                ),
+                "lattice_exclude_types": raw.get("lattice_exclude_types", []),
+                "selected_residual_scale_counts": dict(selected_scale_counts),
+                "policy_rollout_success_rate": raw.get("policy_rollout_success_rate", 0.0),
+                "policy_rollout_progress": raw.get("policy_rollout_progress", 0.0),
+                "oracle_success_rate": raw.get("oracle_success_rate", 0.0),
+                "action_mse_to_best": raw.get("action_mse_to_best", 0.0),
+                "per_task": raw.get("per_task", {}),
+            }
+        )
+    successes = [float(row["policy_rollout_success_rate"]) for row in rows]
+    progresses = [float(row["policy_rollout_progress"]) for row in rows]
+    mses = [float(row["action_mse_to_best"]) for row in rows]
+    summary = {
+        "run_root": str(run_root),
+        "objective": objective,
+        "out_name": config.out_name,
+        "key": config.key,
+        "label": config.label,
+        "num_completed": len(rows),
+        "mean_success": _mean(successes),
+        "std_success": _std(successes),
+        "mean_progress": _mean(progresses),
+        "mean_action_mse_to_best": _mean(mses),
+        "rows": rows,
+    }
+    json_path = results_dir / f"h16_{config.summary_tag}_summary.json"
+    md_path = results_dir / f"h16_{config.summary_tag}_summary.md"
+    if len(rows) == 3:
+        json_path.write_text(json.dumps(summary, indent=2))
+        md_path.write_text(_render_summary_markdown(summary))
+    return summary | {"summary_path": str(json_path)}
+
+
+def _render_summary_markdown(summary: dict[str, Any]) -> str:
+    lines = [
+        "# Generator V1 Rollout Summary",
+        "",
+        f"Objective: `{summary['objective']}`",
+        f"Result file: `{summary['out_name']}`",
+        f"Completed seeds: {summary['num_completed']}",
+        f"Mean success: {summary['mean_success']:.2%} +/- {summary['std_success']:.2%}",
+        f"Mean progress: {summary['mean_progress']:.2%}",
+        "",
+        "| seed | success | progress | mse | anchor | adv weight | min src adv | scales |",
+        "|---:|---:|---:|---:|---|---:|---|---|",
+    ]
+    for row in summary["rows"]:
+        scales = ",".join(str(item) for item in row.get("retrieval_residual_scales", []))
+        lines.append(
+            "| {seed} | {success:.2%} | {progress:.2%} | {mse:.3f} | {anchor} | {weight:.2f} | {min_adv} | {scales} |".format(
+                seed=row["seed"],
+                success=row["policy_rollout_success_rate"],
+                progress=row["policy_rollout_progress"],
+                mse=row["action_mse_to_best"],
+                anchor=row.get("retrieval_residual_anchor", "none"),
+                weight=row.get("retrieval_residual_source_advantage_weight_scale", 0.0),
+                min_adv=_format_min_source_advantage(
+                    row.get("retrieval_residual_min_source_advantage", -1.0e9)
+                ),
+                scales=scales or "none",
+            )
+        )
+    return "\n".join(lines) + "\n"
+
+
+def _format_min_source_advantage(value: Any) -> str:
+    try:
+        numeric = float(value)
+    except (TypeError, ValueError):
+        return "none"
+    if numeric <= -1.0e8:
+        return "none"
+    return f"{numeric:.2f}"
+
+
+def _decision_payload(
+    summaries: list[dict[str, Any]],
+    *,
+    best: dict[str, Any] | None,
+    baseline_success: float,
+    round_index: int,
+) -> dict[str, Any]:
+    return {
+        "generated_utc": datetime.now(timezone.utc).isoformat(),
+        "round": round_index,
+        "baseline_success": baseline_success,
+        "complete": all(item["num_completed"] == 3 for item in summaries),
+        "summaries": [
+            {
+                "key": item["key"],
+                "label": item["label"],
+                "num_completed": item["num_completed"],
+                "mean_success": item["mean_success"],
+                "delta_vs_baseline": item["mean_success"] - baseline_success,
+                "summary_path": item["summary_path"],
+            }
+            for item in summaries
+        ],
+        "best_key": best["key"] if best else None,
+        "best_success": best["mean_success"] if best else None,
+        "best_delta_vs_baseline": (
+            best["mean_success"] - baseline_success if best else None
+        ),
+        "recommendation": _recommendation(best, baseline_success),
+    }
+
+
+def _recommendation(best: dict[str, Any] | None, baseline_success: float) -> str:
+    if best is None:
+        return "wait_for_v1_rollouts_or_debug_failures"
+    if float(best["mean_success"]) > baseline_success + 0.002:
+        return "submit_candidate_oracle_for_best_v1"
+    return "submit_wider_advantage_weight_support_sweep"
+
+
+def _render_decision_markdown(decision: dict[str, Any]) -> str:
+    lines = [
+        "# Generator V1 Decision",
+        "",
+        f"Generated: `{decision['generated_utc']}`",
+        f"Baseline clean success: {decision['baseline_success']:.2%}",
+        f"Recommendation: `{decision['recommendation']}`",
+        "",
+        "| key | completed | success | delta |",
+        "|---|---:|---:|---:|",
+    ]
+    for item in decision["summaries"]:
+        lines.append(
+            f"| {item['key']} | {item['num_completed']} | {item['mean_success']:.2%} | {item['delta_vs_baseline']:+.2%} |"
+        )
+    if decision.get("submitted"):
+        lines.extend(["", "Submitted follow-up jobs:", ""])
+        for item in decision["submitted"]:
+            lines.append(f"- `{item['key']}`: eval `{item.get('eval_job')}`, summary `{item.get('summary_job')}`")
+    if decision.get("submitted_missing"):
+        lines.extend(["", "Resubmitted missing seeds:", ""])
+        for item in decision["submitted_missing"]:
+            lines.append(
+                f"- `{item['key']}` seeds `{item.get('missing_seeds')}`: "
+                f"eval `{item.get('eval_job')}`, summary `{item.get('summary_job')}`"
+            )
+    return "\n".join(lines) + "\n"
+
+
+def _run_build_paper_analysis(project_dir: Path, *, dry_run: bool) -> None:
+    if dry_run:
+        return
+    subprocess.run(
+        ["python3", "scripts/build_paper_analysis.py"],
+        cwd=project_dir,
+        check=True,
+    )
+
+
+def _maybe_submit_next(
+    decision: dict[str, Any],
+    *,
+    project_dir: Path,
+    run_root: Path,
+    dataset: Path,
+    objective: str,
+    dry_run: bool,
+) -> list[dict[str, str]]:
+    marker = project_dir / "results" / "v1_generator_next_submitted.json"
+    if marker.exists():
+        return _load_json(marker).get("submitted", [])
+    best_key = decision.get("best_key")
+    recommendation = str(decision.get("recommendation"))
+    if recommendation == "submit_candidate_oracle_for_best_v1":
+        configs = [
+            _oracle_config(config)
+            for config in INITIAL_CONFIGS
+            if config.key == best_key
+        ]
+    elif recommendation == "submit_wider_advantage_weight_support_sweep":
+        configs = list(NEXT_SWEEP_CONFIGS)
+    else:
+        configs = []
+    submitted = [
+        _submit_config(
+            config,
+            project_dir=project_dir,
+            run_root=run_root,
+            dataset=dataset,
+            objective=objective,
+            dry_run=dry_run,
+        )
+        for config in configs
+    ]
+    marker.write_text(
+        json.dumps(
+            {
+                "generated_utc": datetime.now(timezone.utc).isoformat(),
+                "recommendation": recommendation,
+                "submitted": submitted,
+            },
+            indent=2,
+        )
+    )
+    return submitted
+
+
+def _maybe_resubmit_missing(
+    summaries: list[dict[str, Any]],
+    configs: tuple[V1Config, ...],
+    *,
+    project_dir: Path,
+    run_root: Path,
+    dataset: Path,
+    objective: str,
+    dry_run: bool,
+) -> list[dict[str, str]]:
+    marker = project_dir / "results" / "v1_generator_missing_resubmitted.json"
+    if marker.exists():
+        return _load_json(marker).get("submitted", [])
+    by_key = {item.key: item for item in configs}
+    submitted: list[dict[str, str]] = []
+    for summary in summaries:
+        config = by_key.get(str(summary.get("key")))
+        if config is None:
+            continue
+        present = {int(row["seed"]) for row in summary.get("rows", [])}
+        missing = [str(seed) for seed in (0, 1, 2) if seed not in present]
+        if not missing:
+            continue
+        submitted.append(
+            _submit_config(
+                config,
+                project_dir=project_dir,
+                run_root=run_root,
+                dataset=dataset,
+                objective=objective,
+                array_spec=",".join(missing),
+                dry_run=dry_run,
+            )
+        )
+        submitted[-1]["missing_seeds"] = ",".join(missing)
+    marker.write_text(
+        json.dumps(
+            {
+                "generated_utc": datetime.now(timezone.utc).isoformat(),
+                "submitted": submitted,
+            },
+            indent=2,
+        )
+    )
+    return submitted
+
+
+def _oracle_config(config: V1Config) -> V1Config:
+    return V1Config(
+        key=f"{config.key}_oraclek8",
+        label=f"{config.label}, candidate oracle K8",
+        out_name=config.out_name.replace(".json", "_oraclek8.json"),
+        summary_tag=f"{config.summary_tag}_oraclek8",
+        anchor=config.anchor,
+        advantage_weight=config.advantage_weight,
+        min_source_advantage=config.min_source_advantage,
+        candidate_oracle_rollouts=8,
+    )
+
+
+def _submit_config(
+    config: V1Config,
+    *,
+    project_dir: Path,
+    run_root: Path,
+    dataset: Path,
+    objective: str,
+    array_spec: str = "0-2",
+    dry_run: bool,
+) -> dict[str, str]:
+    export = {
+        "PROJECT_DIR": str(project_dir),
+        "RUN_ROOT": str(run_root),
+        "DATASET": str(dataset),
+        "OBJECTIVE": objective,
+        "CHECKPOINT_NAME": "best_transport.pt",
+        "MAX_GROUPS": "all",
+        "GROUP_BATCH_SIZE": "8",
+        "EVAL_SPLIT": "validation",
+        "SELECTION_MODE": "retrieval_residual",
+        "NUM_CANDIDATES": "1",
+        "CANDIDATE_SIGMA": "0.2",
+        "SELECTION_MARGIN": "0.0",
+        "RETRIEVAL_NEIGHBORS": "6",
+        "RETRIEVAL_METRIC": "raw",
+        "RETRIEVAL_RESIDUAL_REDUCE": "compose_mean_by_type",
+        "RETRIEVAL_RESIDUAL_DIRECTION": "candidate_minus_anchor",
+        "RETRIEVAL_RESIDUAL_SCALES_COLON": "0.35:0.4:0.45",
+        "LATTICE_EXCLUDE_TYPES_COLON": EXCLUDE_TYPES,
+        "RETRIEVAL_RESIDUAL_SOURCE_SCORE_BONUS_MAP": str(project_dir / SOURCE_SCORE_MAP),
+        "OUT_NAME": config.out_name,
+        "RETRIEVAL_RESIDUAL_ANCHOR": config.anchor,
+        "RETRIEVAL_RESIDUAL_SOURCE_ADVANTAGE_WEIGHT_SCALE": str(config.advantage_weight),
+        "RETRIEVAL_RESIDUAL_MIN_SOURCE_ADVANTAGE": str(config.min_source_advantage),
+        "CANDIDATE_ORACLE_ROLLOUTS": str(config.candidate_oracle_rollouts),
+    }
+    eval_cmd = [
+        "sbatch",
+        "--array=0-2",
+        "--job-name",
+        _job_name(config.key),
+        "--export",
+        "ALL," + ",".join(f"{key}={value}" for key, value in export.items()),
+        "scripts/slurm/eval_maniskill_policy_rollout.sbatch",
+    ]
+    eval_cmd[1] = f"--array={array_spec}"
+    eval_job = _submit(eval_cmd, project_dir=project_dir, dry_run=dry_run)
+    summary_export = {
+        "PROJECT_DIR": str(project_dir),
+        "RUN_ROOT": str(run_root),
+        "OBJECTIVE": objective,
+        "OUT_NAME": config.out_name,
+        "SUMMARY_TAG": config.summary_tag,
+    }
+    summary_cmd = [
+        "sbatch",
+        "--dependency",
+        f"afterok:{eval_job}" if eval_job != "dry-run" else "afterok:0",
+        "--job-name",
+        _job_name(f"sum_{config.key}"),
+        "--export",
+        "ALL," + ",".join(f"{key}={value}" for key, value in summary_export.items()),
+        "scripts/slurm/summarize_h16_policy_ckpt.sbatch",
+    ]
+    summary_job = _submit(summary_cmd, project_dir=project_dir, dry_run=dry_run)
+    return {"key": config.key, "eval_job": eval_job, "summary_job": summary_job}
+
+
+def _job_name(key: str) -> str:
+    return ("v1_" + key).replace("policyanchor", "pol")[:20]
+
+
+def _submit(cmd: list[str], *, project_dir: Path, dry_run: bool) -> str:
+    if dry_run:
+        print("DRY RUN:", " ".join(cmd))
+        return "dry-run"
+    result = subprocess.run(
+        cmd,
+        cwd=project_dir,
+        check=True,
+        text=True,
+        capture_output=True,
+    )
+    for token in result.stdout.split():
+        if token.isdigit():
+            return token
+    raise RuntimeError(f"Could not parse sbatch job id from: {result.stdout}")
+
+
+if __name__ == "__main__":
+    raise SystemExit(main())

workspace/scripts/build_paper_analysis.py

@@ -598,6 +598,81 @@ METHODS = [
             "besttransport_margin0p00_k6_srcscore_task_pick001_stack005_policyanchor_advw2p0_summary.json"
         ),
     ),
+    MethodSpec(
+        key="transport_field_reground_fieldonly_k6matched_b12_clean_k6_dropnoopwg_retargeted_srcscore_task_pick001_stack005_advw2p0_gate0",
+        label="Generator V1 utility-weighted tangents, expert anchor, source advantage gate >= 0.0",
+        summary_path=(
+            "h16_transport_field_reground_fieldonly_k6clean_dropnoopwg_b12_v1_"
+            "besttransport_margin0p00_k6_srcscore_task_pick001_stack005_advw2p0_gate0_summary.json"
+        ),
+    ),
+    MethodSpec(
+        key="transport_field_reground_fieldonly_k6matched_b12_clean_k6_dropnoopwg_retargeted_srcscore_task_pick001_stack005_policyanchor_advw2p0_gate0",
+        label="Generator V1 positive tangents, policy anchor, source advantage gate >= 0.0",
+        summary_path=(
+            "h16_transport_field_reground_fieldonly_k6clean_dropnoopwg_b12_v1_"
+            "besttransport_margin0p00_k6_srcscore_task_pick001_stack005_policyanchor_advw2p0_gate0_summary.json"
+        ),
+    ),
+    MethodSpec(
+        key="transport_field_reground_fieldonly_k6matched_b12_clean_k6_dropnoopwg_retargeted_srcscore_task_pick001_stack005_advw1p0_oraclek8",
+        label="Generator V1 expert-anchor advantage weight 1.0, candidate-oracle K8",
+        summary_path=(
+            "h16_transport_field_reground_fieldonly_k6clean_dropnoopwg_b12_v1_"
+            "besttransport_margin0p00_k6_srcscore_task_pick001_stack005_advw1p0_oraclek8_summary.json"
+        ),
+        headline_metric="mean_candidate_oracle_success_rate",
+    ),
+    MethodSpec(
+        key="transport_field_reground_fieldonly_k6matched_b12_clean_k6_dropnoopwg_retargeted_srcscore_task_pick001_stack005_advw2p0_oraclek8",
+        label="Generator V1 expert-anchor advantage weight 2.0, candidate-oracle K8",
+        summary_path=(
+            "h16_transport_field_reground_fieldonly_k6clean_dropnoopwg_b12_v1_"
+            "besttransport_margin0p00_k6_srcscore_task_pick001_stack005_advw2p0_oraclek8_summary.json"
+        ),
+        headline_metric="mean_candidate_oracle_success_rate",
+    ),
+    MethodSpec(
+        key="transport_field_reground_fieldonly_k6matched_b12_clean_k6_dropnoopwg_retargeted_srcscore_task_pick001_stack005_policyanchor_advw2p0_oraclek8",
+        label="Generator V1 policy-anchor advantage weight 2.0, candidate-oracle K8",
+        summary_path=(
+            "h16_transport_field_reground_fieldonly_k6clean_dropnoopwg_b12_v1_"
+            "besttransport_margin0p00_k6_srcscore_task_pick001_stack005_policyanchor_advw2p0_oraclek8_summary.json"
+        ),
+        headline_metric="mean_candidate_oracle_success_rate",
+    ),
+    MethodSpec(
+        key="transport_field_reground_fieldonly_k6matched_b12_clean_k6_dropnoopwg_retargeted_srcscore_task_pick001_stack005_advw0p5",
+        label="Generator V1 utility-weighted tangents, expert anchor, advantage weight 0.5",
+        summary_path=(
+            "h16_transport_field_reground_fieldonly_k6clean_dropnoopwg_b12_v1_"
+            "besttransport_margin0p00_k6_srcscore_task_pick001_stack005_advw0p5_summary.json"
+        ),
+    ),
+    MethodSpec(
+        key="transport_field_reground_fieldonly_k6matched_b12_clean_k6_dropnoopwg_retargeted_srcscore_task_pick001_stack005_advw4p0",
+        label="Generator V1 utility-weighted tangents, expert anchor, advantage weight 4.0",
+        summary_path=(
+            "h16_transport_field_reground_fieldonly_k6clean_dropnoopwg_b12_v1_"
+            "besttransport_margin0p00_k6_srcscore_task_pick001_stack005_advw4p0_summary.json"
+        ),
+    ),
+    MethodSpec(
+        key="transport_field_reground_fieldonly_k6matched_b12_clean_k6_dropnoopwg_retargeted_srcscore_task_pick001_stack005_policyanchor_advw1p0",
+        label="Generator V1 positive tangents, policy anchor, advantage weight 1.0",
+        summary_path=(
+            "h16_transport_field_reground_fieldonly_k6clean_dropnoopwg_b12_v1_"
+            "besttransport_margin0p00_k6_srcscore_task_pick001_stack005_policyanchor_advw1p0_summary.json"
+        ),
+    ),
+    MethodSpec(
+        key="transport_field_reground_fieldonly_k6matched_b12_clean_k6_dropnoopwg_retargeted_srcscore_task_pick001_stack005_policyanchor_advw4p0",
+        label="Generator V1 positive tangents, policy anchor, advantage weight 4.0",
+        summary_path=(
+            "h16_transport_field_reground_fieldonly_k6clean_dropnoopwg_b12_v1_"
+            "besttransport_margin0p00_k6_srcscore_task_pick001_stack005_policyanchor_advw4p0_summary.json"
+        ),
+    ),
     MethodSpec(
         key="transport_field_reground_fieldonly_k6matched_b12_clean_k6_dropnoopwg_retargeted_srcscore_task_pick001_stack005_scaleoracle_trainall_k4_scale0025",
         label="K6-matched transported residual field re-grounding, all-train tangent-length prior 0.025",