auto-sync 2026-07-02T14:58:13Z workspace (part 8)

workspace/results/h16_transport_field_reground_fieldonly_k6clean_dropnoopwg_b12_v1_besttransport_margin0p00_k6_srcscore_task_pick001_stack005_summary.json

@@ -0,0 +1,481 @@
+{
+  "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.json",
+  "num_completed": 3,
+  "baseline_h4_policy_success": 0.2967,
+  "baseline_h16_rank_checkpoint_success": 0.29739130434782607,
+  "mean_success": 0.3889855072463768,
+  "std_success": 0.01859595934184978,
+  "mean_progress": 0.5995674399000364,
+  "mean_action_mse_to_best": 0.5131558791707284,
+  "gain_vs_h4": 0.0922855072463768,
+  "gain_vs_h16_rank_checkpoint": 0.09159420289855075,
+  "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.json",
+      "num_groups": 575,
+      "selection_mode": "retrieval_residual",
+      "num_candidates": 48,
+      "candidate_sigma": 0.0,
+      "proposal_types": [],
+      "proposal_lattice_types": [],
+      "selection_margin": 0.0,
+      "prepend_policy_candidate": false,
+      "field_optim_steps": 0,
+      "field_optim_step_size": 0.0,
+      "field_optim_trust_radius": 0.0,
+      "field_optim_l2_penalty": 0.0,
+      "retrieval_neighbors": 6,
+      "retrieval_metric": "raw",
+      "retrieval_type_min_success": 0.0,
+      "retrieval_type_success_bonus_scale": 0.0,
+      "retrieval_residual_consensus_penalty_scale": 0.0,
+      "retrieval_residual_min_source_progress": 0.0,
+      "retrieval_residual_min_source_advantage": -1000000000.0,
+      "retrieval_residual_source_progress_bonus_scale": 0.0,
+      "retrieval_residual_source_score_bonus_scale": 0.0,
+      "retrieval_residual_source_advantage_bonus_scale": 0.0,
+      "retrieval_residual_composite_l2_penalty_scale": 0.0,
+      "retrieval_residual_scale": 1.0,
+      "retrieval_residual_scales": [
+        0.35,
+        0.4,
+        0.45
+      ],
+      "retrieval_residual_anchor": "expert",
+      "retrieval_residual_direction": "candidate_minus_anchor",
+      "retrieval_residual_reduce": "compose_mean_by_type",
+      "retrieval_residual_challenger_types": [],
+      "retrieval_residual_challenger_scales": [],
+      "retrieval_residual_challenger_margin": 0.0,
+      "retrieval_residual_challenger_type_margins": {},
+      "retrieval_residual_challenger_tasks": [],
+      "retrieval_residual_challenger_type_tasks": {},
+      "lattice_exclude_types": [
+        "residual_random_negative",
+        "residual_wrong_direction",
+        "residual_near_miss+residual_no_op",
+        "residual_no_op+residual_wrong_gripper"
+      ],
+      "lattice_exclude_type_tasks": {},
+      "candidate_type_bonuses": {},
+      "candidate_type_bonuses_by_task": {},
+      "candidate_type_bonus_components": false,
+      "candidate_oracle_rollouts": 0,
+      "candidate_oracle_unique_tolerance": 1e-06,
+      "candidate_oracle_success_rate": null,
+      "candidate_oracle_progress": null,
+      "candidate_oracle_score": null,
+      "candidate_oracle_regret": null,
+      "candidate_oracle_score_gain_over_selected": null,
+      "candidate_oracle_unique_count": null,
+      "candidate_oracle_improvement_rate": null,
+      "candidate_oracle_selected_branch_success_rate": null,
+      "candidate_oracle_selected_branch_progress": null,
+      "candidate_oracle_type_counts": {},
+      "candidate_oracle_best_branch_rank": null,
+      "candidate_oracle_best_branch_rank_counts": {},
+      "candidate_oracle_branch_success_rates": [],
+      "candidate_oracle_branch_progress": [],
+      "candidate_oracle_branch_score_gains_over_selected": [],
+      "selected_residual_scale_counts": {
+        "0.35": 336,
+        "0.45": 152,
+        "0.4": 87
+      },
+      "policy_rollout_success_rate": 0.3773913043478261,
+      "policy_rollout_progress": 0.5869067004063855,
+      "oracle_success_rate": 0.8573913043478261,
+      "action_mse_to_best": 0.4981627210083863,
+      "best_policy_val": {
+        "bc_loss": 0.3445456624031067,
+        "field_effect_loss": 0.009637858977334367,
+        "field_potential_loss": 0.28480341533819836,
+        "field_preference_loss": 0.5244891477955712,
+        "lattice_edges": 3833.3333333333335,
+        "progress_mae": 0.19576937291357252,
+        "proposal_loss": 0.0,
+        "rank_acc": 0.8192069795396593,
+        "rank_loss": 0.524484740363227,
+        "regret_mae": 0.42191632754272884,
+        "success_accuracy": 0.8693191342883639,
+        "total_loss": 0.261704473859734,
+        "transport_field_candidates": 0.0,
+        "transport_field_edges": 0.0,
+        "transport_field_loss": 0.0,
+        "transport_field_potential_loss": 0.0,
+        "transport_field_preference_loss": 0.0,
+        "transport_field_rank_acc": 0.0
+      },
+      "per_task": {
+        "LiftPegUpright-v1": {
+          "action_mse_to_best": 0.33983143654258285,
+          "expert_success_rate": 0.8865979381443299,
+          "num_groups": 97,
+          "oracle_success_rate": 0.9278350515463918,
+          "policy_expert_regret": 1.07819144181984,
+          "policy_oracle_regret": 1.1046069195282828,
+          "policy_rollout_progress": 0.5716054517890989,
+          "policy_rollout_success_rate": 0.2268041237113402,
+          "restore_max_error": 2.384185791015625e-07
+        },
+        "PickCube-v1": {
+          "action_mse_to_best": 0.46344211518477935,
+          "expert_success_rate": 0.9375,
+          "num_groups": 208,
+          "oracle_success_rate": 0.9471153846153846,
+          "policy_expert_regret": 0.9422954545230963,
+          "policy_oracle_regret": 0.9490729826233851,
+          "policy_rollout_progress": 0.6223857593991292,
+          "policy_rollout_success_rate": 0.3605769230769231,
+          "restore_max_error": 2.384185791015625e-07
+        },
+        "PullCube-v1": {
+          "action_mse_to_best": 0.6133117520963991,
+          "expert_success_rate": 0.19480519480519481,
+          "num_groups": 77,
+          "oracle_success_rate": 0.36363636363636365,
+          "policy_expert_regret": 0.2165840214432834,
+          "policy_oracle_regret": 0.383972076118573,
+          "policy_rollout_progress": 0.36198321732317473,
+          "policy_rollout_success_rate": 0.2597402597402597,
+          "restore_max_error": 2.3096799850463867e-07
+        },
+        "PushCube-v1": {
+          "action_mse_to_best": 0.5042698027447183,
+          "expert_success_rate": 0.8617021276595744,
+          "num_groups": 94,
+          "oracle_success_rate": 0.9893617021276596,
+          "policy_expert_regret": 0.37636212211974124,
+          "policy_oracle_regret": 0.4349256801478406,
+          "policy_rollout_progress": 0.7794892613558059,
+          "policy_rollout_success_rate": 0.7659574468085106,
+          "restore_max_error": 2.384185791015625e-07
+        },
+        "StackCube-v1": {
+          "action_mse_to_best": 0.6308847364376892,
+          "expert_success_rate": 0.6767676767676768,
+          "num_groups": 99,
+          "oracle_success_rate": 0.8585858585858586,
+          "policy_expert_regret": 0.8869800036302721,
+          "policy_oracle_regret": 1.0460948816152533,
+          "policy_rollout_progress": 0.5194412894321211,
+          "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.json",
+      "num_groups": 575,
+      "selection_mode": "retrieval_residual",
+      "num_candidates": 48,
+      "candidate_sigma": 0.0,
+      "proposal_types": [],
+      "proposal_lattice_types": [],
+      "selection_margin": 0.0,
+      "prepend_policy_candidate": false,
+      "field_optim_steps": 0,
+      "field_optim_step_size": 0.0,
+      "field_optim_trust_radius": 0.0,
+      "field_optim_l2_penalty": 0.0,
+      "retrieval_neighbors": 6,
+      "retrieval_metric": "raw",
+      "retrieval_type_min_success": 0.0,
+      "retrieval_type_success_bonus_scale": 0.0,
+      "retrieval_residual_consensus_penalty_scale": 0.0,
+      "retrieval_residual_min_source_progress": 0.0,
+      "retrieval_residual_min_source_advantage": -1000000000.0,
+      "retrieval_residual_source_progress_bonus_scale": 0.0,
+      "retrieval_residual_source_score_bonus_scale": 0.0,
+      "retrieval_residual_source_advantage_bonus_scale": 0.0,
+      "retrieval_residual_composite_l2_penalty_scale": 0.0,
+      "retrieval_residual_scale": 1.0,
+      "retrieval_residual_scales": [
+        0.35,
+        0.4,
+        0.45
+      ],
+      "retrieval_residual_anchor": "expert",
+      "retrieval_residual_direction": "candidate_minus_anchor",
+      "retrieval_residual_reduce": "compose_mean_by_type",
+      "retrieval_residual_challenger_types": [],
+      "retrieval_residual_challenger_scales": [],
+      "retrieval_residual_challenger_margin": 0.0,
+      "retrieval_residual_challenger_type_margins": {},
+      "retrieval_residual_challenger_tasks": [],
+      "retrieval_residual_challenger_type_tasks": {},
+      "lattice_exclude_types": [
+        "residual_random_negative",
+        "residual_wrong_direction",
+        "residual_near_miss+residual_no_op",
+        "residual_no_op+residual_wrong_gripper"
+      ],
+      "lattice_exclude_type_tasks": {},
+      "candidate_type_bonuses": {},
+      "candidate_type_bonuses_by_task": {},
+      "candidate_type_bonus_components": false,
+      "candidate_oracle_rollouts": 0,
+      "candidate_oracle_unique_tolerance": 1e-06,
+      "candidate_oracle_success_rate": null,
+      "candidate_oracle_progress": null,
+      "candidate_oracle_score": null,
+      "candidate_oracle_regret": null,
+      "candidate_oracle_score_gain_over_selected": null,
+      "candidate_oracle_unique_count": null,
+      "candidate_oracle_improvement_rate": null,
+      "candidate_oracle_selected_branch_success_rate": null,
+      "candidate_oracle_selected_branch_progress": null,
+      "candidate_oracle_type_counts": {},
+      "candidate_oracle_best_branch_rank": null,
+      "candidate_oracle_best_branch_rank_counts": {},
+      "candidate_oracle_branch_success_rates": [],
+      "candidate_oracle_branch_progress": [],
+      "candidate_oracle_branch_score_gains_over_selected": [],
+      "selected_residual_scale_counts": {
+        "0.35": 313,
+        "0.4": 46,
+        "0.45": 216
+      },
+      "policy_rollout_success_rate": 0.3791304347826087,
+      "policy_rollout_progress": 0.5928394121466124,
+      "oracle_success_rate": 0.8695652173913043,
+      "action_mse_to_best": 0.5139227381794025,
+      "best_policy_val": {
+        "bc_loss": 0.31955528176493114,
+        "field_effect_loss": 0.010269693254182736,
+        "field_potential_loss": 0.30970407525698346,
+        "field_preference_loss": 0.5266202224625481,
+        "lattice_edges": 3833.3333333333335,
+        "progress_mae": 0.20275857630703184,
+        "proposal_loss": 0.0,
+        "rank_acc": 0.8124439285861121,
+        "rank_loss": 0.5266130566596985,
+        "regret_mae": 0.40278742214043933,
+        "success_accuracy": 0.8677195310592651,
+        "total_loss": 0.2706700298521254,
+        "transport_field_candidates": 0.0,
+        "transport_field_edges": 0.0,
+        "transport_field_loss": 0.0,
+        "transport_field_potential_loss": 0.0,
+        "transport_field_preference_loss": 0.0,
+        "transport_field_rank_acc": 0.0
+      },
+      "per_task": {
+        "LiftPegUpright-v1": {
+          "action_mse_to_best": 0.4174908100528461,
+          "expert_success_rate": 0.8584070796460177,
+          "num_groups": 113,
+          "oracle_success_rate": 0.9380530973451328,
+          "policy_expert_regret": 0.9912996379147588,
+          "policy_oracle_regret": 1.0707513929995816,
+          "policy_rollout_progress": 0.6289191406912509,
+          "policy_rollout_success_rate": 0.22123893805309736,
+          "restore_max_error": 2.384185791015625e-07
+        },
+        "PickCube-v1": {
+          "action_mse_to_best": 0.4553744464950717,
+          "expert_success_rate": 0.9402173913043478,
+          "num_groups": 184,
+          "oracle_success_rate": 0.9456521739130435,
+          "policy_expert_regret": 1.0321165590658379,
+          "policy_oracle_regret": 1.032140563540768,
+          "policy_rollout_progress": 0.5804765945067629,
+          "policy_rollout_success_rate": 0.32608695652173914,
+          "restore_max_error": 2.384185791015625e-07
+        },
+        "PullCube-v1": {
+          "action_mse_to_best": 0.6725145407805317,
+          "expert_success_rate": 0.25,
+          "num_groups": 76,
+          "oracle_success_rate": 0.40789473684210525,
+          "policy_expert_regret": 0.21536361078094496,
+          "policy_oracle_regret": 0.4112854622990677,
+          "policy_rollout_progress": 0.34068188899637836,
+          "policy_rollout_success_rate": 0.2236842105263158,
+          "restore_max_error": 2.384185791015625e-07
+        },
+        "PushCube-v1": {
+          "action_mse_to_best": 0.4303124049628103,
+          "expert_success_rate": 0.8198198198198198,
+          "num_groups": 111,
+          "oracle_success_rate": 1.0,
+          "policy_expert_regret": 0.26138654270687617,
+          "policy_oracle_regret": 0.27855410983970574,
+          "policy_rollout_progress": 0.8655900343044384,
+          "policy_rollout_success_rate": 0.8558558558558559,
+          "restore_max_error": 2.3795291781425476e-07
+        },
+        "StackCube-v1": {
+          "action_mse_to_best": 0.7215871726582338,
+          "expert_success_rate": 0.6923076923076923,
+          "num_groups": 91,
+          "oracle_success_rate": 0.8571428571428571,
+          "policy_expert_regret": 1.0029281839564606,
+          "policy_oracle_regret": 1.162750510396538,
+          "policy_rollout_progress": 0.4509317398398787,
+          "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.json",
+      "num_groups": 575,
+      "selection_mode": "retrieval_residual",
+      "num_candidates": 48,
+      "candidate_sigma": 0.0,
+      "proposal_types": [],
+      "proposal_lattice_types": [],
+      "selection_margin": 0.0,
+      "prepend_policy_candidate": false,
+      "field_optim_steps": 0,
+      "field_optim_step_size": 0.0,
+      "field_optim_trust_radius": 0.0,
+      "field_optim_l2_penalty": 0.0,
+      "retrieval_neighbors": 6,
+      "retrieval_metric": "raw",
+      "retrieval_type_min_success": 0.0,
+      "retrieval_type_success_bonus_scale": 0.0,
+      "retrieval_residual_consensus_penalty_scale": 0.0,
+      "retrieval_residual_min_source_progress": 0.0,
+      "retrieval_residual_min_source_advantage": -1000000000.0,
+      "retrieval_residual_source_progress_bonus_scale": 0.0,
+      "retrieval_residual_source_score_bonus_scale": 0.0,
+      "retrieval_residual_source_advantage_bonus_scale": 0.0,
+      "retrieval_residual_composite_l2_penalty_scale": 0.0,
+      "retrieval_residual_scale": 1.0,
+      "retrieval_residual_scales": [
+        0.35,
+        0.4,
+        0.45
+      ],
+      "retrieval_residual_anchor": "expert",
+      "retrieval_residual_direction": "candidate_minus_anchor",
+      "retrieval_residual_reduce": "compose_mean_by_type",
+      "retrieval_residual_challenger_types": [],
+      "retrieval_residual_challenger_scales": [],
+      "retrieval_residual_challenger_margin": 0.0,
+      "retrieval_residual_challenger_type_margins": {},
+      "retrieval_residual_challenger_tasks": [],
+      "retrieval_residual_challenger_type_tasks": {},
+      "lattice_exclude_types": [
+        "residual_random_negative",
+        "residual_wrong_direction",
+        "residual_near_miss+residual_no_op",
+        "residual_no_op+residual_wrong_gripper"
+      ],
+      "lattice_exclude_type_tasks": {},
+      "candidate_type_bonuses": {},
+      "candidate_type_bonuses_by_task": {},
+      "candidate_type_bonus_components": false,
+      "candidate_oracle_rollouts": 0,
+      "candidate_oracle_unique_tolerance": 1e-06,
+      "candidate_oracle_success_rate": null,
+      "candidate_oracle_progress": null,
+      "candidate_oracle_score": null,
+      "candidate_oracle_regret": null,
+      "candidate_oracle_score_gain_over_selected": null,
+      "candidate_oracle_unique_count": null,
+      "candidate_oracle_improvement_rate": null,
+      "candidate_oracle_selected_branch_success_rate": null,
+      "candidate_oracle_selected_branch_progress": null,
+      "candidate_oracle_type_counts": {},
+      "candidate_oracle_best_branch_rank": null,
+      "candidate_oracle_best_branch_rank_counts": {},
+      "candidate_oracle_branch_success_rates": [],
+      "candidate_oracle_branch_progress": [],
+      "candidate_oracle_branch_score_gains_over_selected": [],
+      "selected_residual_scale_counts": {
+        "0.35": 337,
+        "0.45": 165,
+        "0.4": 73
+      },
+      "policy_rollout_success_rate": 0.41043478260869565,
+      "policy_rollout_progress": 0.6189562071471111,
+      "oracle_success_rate": 0.8765217391304347,
+      "action_mse_to_best": 0.5273821783243962,
+      "best_policy_val": {
+        "bc_loss": 0.2974168997671869,
+        "field_effect_loss": 0.009809228720971279,
+        "field_potential_loss": 0.29470594475666684,
+        "field_preference_loss": 0.5080636441707611,
+        "lattice_edges": 3833.3333333333335,
+        "progress_mae": 0.1944621221886741,
+        "proposal_loss": 0.0,
+        "rank_acc": 0.8252745270729065,
+        "rank_loss": 0.5080037944846683,
+        "regret_mae": 0.3989163488149643,
+        "success_accuracy": 0.870691223276986,
+        "total_loss": 0.2595750474267536,
+        "transport_field_candidates": 0.0,
+        "transport_field_edges": 0.0,
+        "transport_field_loss": 0.0,
+        "transport_field_potential_loss": 0.0,
+        "transport_field_preference_loss": 0.0,
+        "transport_field_rank_acc": 0.0
+      },
+      "per_task": {
+        "LiftPegUpright-v1": {
+          "action_mse_to_best": 0.44695540916291066,
+          "expert_success_rate": 0.8229166666666666,
+          "num_groups": 96,
+          "oracle_success_rate": 0.9270833333333334,
+          "policy_expert_regret": 0.8176391956706842,
+          "policy_oracle_regret": 0.9106746912002563,
+          "policy_rollout_progress": 0.6583504987259706,
+          "policy_rollout_success_rate": 0.3645833333333333,
+          "restore_max_error": 1.955777406692505e-07
+        },
+        "PickCube-v1": {
+          "action_mse_to_best": 0.4654180597927835,
+          "expert_success_rate": 0.9444444444444444,
+          "num_groups": 198,
+          "oracle_success_rate": 0.9595959595959596,
+          "policy_expert_regret": 0.960848643049372,
+          "policy_oracle_regret": 0.9679320935780803,
+          "policy_rollout_progress": 0.6404885555325885,
+          "policy_rollout_success_rate": 0.35353535353535354,
+          "restore_max_error": 2.384185791015625e-07
+        },
+        "PullCube-v1": {
+          "action_mse_to_best": 0.6722042224473423,
+          "expert_success_rate": 0.24444444444444444,
+          "num_groups": 90,
+          "oracle_success_rate": 0.4666666666666667,
+          "policy_expert_regret": 0.3065602982416749,
+          "policy_oracle_regret": 0.450493123030497,
+          "policy_rollout_progress": 0.38515357037799225,
+          "policy_rollout_success_rate": 0.2777777777777778,
+          "restore_max_error": 2.384185791015625e-07
+        },
+        "PushCube-v1": {
+          "action_mse_to_best": 0.5134861391074587,
+          "expert_success_rate": 0.8514851485148515,
+          "num_groups": 101,
+          "oracle_success_rate": 1.0,
+          "policy_expert_regret": 0.2486148280377435,
+          "policy_oracle_regret": 0.28494056705201026,
+          "policy_rollout_progress": 0.8635742844331382,
+          "policy_rollout_success_rate": 0.8514851485148515,
+          "restore_max_error": 2.384185791015625e-07
+        },
+        "StackCube-v1": {
+          "action_mse_to_best": 0.620264192753368,
+          "expert_success_rate": 0.7666666666666667,
+          "num_groups": 90,
+          "oracle_success_rate": 0.9111111111111111,
+          "policy_expert_regret": 1.0516069930460719,
+          "policy_oracle_regret": 1.1881121820873684,
+          "policy_rollout_progress": 0.48885125749640995,
+          "policy_rollout_success_rate": 0.2222222222222222,
+          "restore_max_error": 2.2351741790771484e-07
+        }
+      }
+    }
+  ]
+}

workspace/results/h16_transport_field_reground_fieldonly_k6clean_dropnoopwg_b12_v1_besttransport_margin0p00_k6_srcscore_task_pick001_stack005_summary.md

@@ -0,0 +1,19 @@
+# h=16 Best-Policy Checkpoint Rollout
+
+Run root: `/scratch/knguy52/dovla/experiments/dovla_h16_policy_ckpt_runs`
+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.json`
+Completed seeds: 3
+Baseline h=4 policy success: 29.67%
+Baseline h=16 rank-checkpoint success: 29.74%
+
+Mean success: 38.90% +/- 1.86%
+Gain vs h=16 rank checkpoint: +9.16%
+Mean progress: 59.96%
+Mean action MSE to best: 0.513
+
+| seed | mode | k | policy cand | retrieval K | retrieval metric | residual anchor | residual direction | residual reduce | min type success | type success bonus | consensus penalty | min source progress | source progress bonus | source score bonus | residual scale | residual scales | margin | sigma | opt steps | trust | success | progress | oracle | candidate oracle | oracle gain | action MSE |
+|---:|---|---:|---|---:|---|---|---|---|---:|---:|---:|---:|---:|---:|---:|---|---:|---:|---:|---:|---:|---:|---:|---:|---:|---:|
+| 0 | retrieval_residual | 48 | no | 6 | raw | expert | candidate_minus_anchor | compose_mean_by_type | 0.00 | 0.000 | 0.000 | 0.00 | 0.000 | 0.000 | 1.00 | 0.35,0.40,0.45 | 0.000 | 0.00 | 0 | 0.00 | 37.74% | 58.69% | 85.74% | n/a | n/a | 0.498 |
+| 1 | retrieval_residual | 48 | no | 6 | raw | expert | candidate_minus_anchor | compose_mean_by_type | 0.00 | 0.000 | 0.000 | 0.00 | 0.000 | 0.000 | 1.00 | 0.35,0.40,0.45 | 0.000 | 0.00 | 0 | 0.00 | 37.91% | 59.28% | 86.96% | n/a | n/a | 0.514 |
+| 2 | retrieval_residual | 48 | no | 6 | raw | expert | candidate_minus_anchor | compose_mean_by_type | 0.00 | 0.000 | 0.000 | 0.00 | 0.000 | 0.000 | 1.00 | 0.35,0.40,0.45 | 0.000 | 0.00 | 0 | 0.00 | 41.04% | 61.90% | 87.65% | n/a | n/a | 0.527 |

workspace/results/paper_analysis.json

@@ -1,11 +1,11 @@
 {
-  "best_clean_key": "transport_field_reground_fieldonly_k6matched_b12_clean_k6_dropnoopwg_retargeted_srcscore001",
-  "generated_utc": "2026-07-02T14:35:03+00:00",
+  "best_clean_key": "transport_field_reground_fieldonly_k6matched_b12_clean_k6_dropnoopwg_retargeted_srcscore_task_pick001_stack005",
+  "generated_utc": "2026-07-02T14:59:38+00:00",
   "mechanism_gap": {
-    "best_clean_vs_direct_same_ckpt": 0.10550724637681164,
-    "best_clean_vs_h16": 0.09101449275362328,
+    "best_clean_vs_direct_same_ckpt": 0.1060869565217391,
+    "best_clean_vs_h16": 0.09159420289855075,
     "same_state_full_vs_no_expert": 0.12347826086956515,
-    "same_state_no_expert_vs_best_clean": 0.18144927536231886,
+    "same_state_no_expert_vs_best_clean": 0.1808695652173914,
     "same_state_no_expert_vs_h16": 0.27246376811594214
   },
   "methods": {
@@ -11634,6 +11634,294 @@
       "source": "results/h16_transport_field_reground_fieldonly_k6clean_dropnoopwg_b12_v1_besttransport_margin0p00_k6_srcscore0015_summary.json",
       "std_success": 0.018157054798105324
     },
+    "transport_field_reground_fieldonly_k6matched_b12_clean_k6_dropnoopwg_retargeted_srcscore001_metric_taskrel": {
+      "ci95_success": 0.04121578446082111,
+      "headline_ci95_success": 0.04121578446082111,
+      "headline_metric": "mean_success",
+      "headline_metric_label": "deployment",
+      "headline_std_success": 0.0165902469811643,
+      "headline_success": 0.37217391304347824,
+      "label": "K6-matched transported residual field re-grounding, source-score 0.01, task-relative retrieval chart",
+      "mean_action_mse_to_best": 0.5116538966142073,
+      "mean_progress": 0.5848761159114108,
+      "mean_success": 0.37217391304347824,
+      "num_completed": 3,
+      "per_task_success": {
+        "LiftPegUpright-v1": {
+          "mean_num_groups": 102.0,
+          "mean_success": 0.27972502255471415,
+          "std_success": 0.07423460353404222
+        },
+        "PickCube-v1": {
+          "mean_num_groups": 196.66666666666666,
+          "mean_success": 0.33342482799004536,
+          "std_success": 0.0544887761297919
+        },
+        "PullCube-v1": {
+          "mean_num_groups": 81.0,
+          "mean_success": 0.21277436014278117,
+          "std_success": 0.025520989159461067
+        },
+        "PushCube-v1": {
+          "mean_num_groups": 102.0,
+          "mean_success": 0.7912868630163764,
+          "std_success": 0.052717414948247804
+        },
+        "StackCube-v1": {
+          "mean_num_groups": 93.33333333333333,
+          "mean_success": 0.22649942649942653,
+          "std_success": 0.05958316533404071
+        }
+      },
+      "seed_action_mse_to_best": {
+        "0": 0.4960412207715537,
+        "1": 0.5053197572339813,
+        "2": 0.533600711837087
+      },
+      "seed_progress": {
+        "0": 0.5859811081583409,
+        "1": 0.5724870564092113,
+        "2": 0.5961601831666802
+      },
+      "seed_success": {
+        "0": 0.3739130434782609,
+        "1": 0.35478260869565215,
+        "2": 0.38782608695652177
+      },
+      "selected_candidate_type_counts": {
+        "retrieval_residual_policy_residual": 553,
+        "retrieval_residual_residual_near_miss": 213,
+        "retrieval_residual_residual_near_miss+residual_wrong_gripper": 251,
+        "retrieval_residual_residual_no_op": 288,
+        "retrieval_residual_residual_wrong_gripper": 420
+      },
+      "selected_residual_scale_counts": {
+        "0.35": 903,
+        "0.4": 195,
+        "0.45": 627
+      },
+      "selected_type_outcomes": {
+        "retrieval_residual_policy_residual": {
+          "count": 553.0,
+          "mean_progress": 0.545225228481062,
+          "success_rate": 0.2766726943942134
+        },
+        "retrieval_residual_residual_near_miss": {
+          "count": 213.0,
+          "mean_progress": 0.7468650269501366,
+          "success_rate": 0.596244131455399
+        },
+        "retrieval_residual_residual_near_miss+residual_wrong_gripper": {
+          "count": 251.0,
+          "mean_progress": 0.7313106328219294,
+          "success_rate": 0.5418326693227091
+        },
+        "retrieval_residual_residual_no_op": {
+          "count": 288.0,
+          "mean_progress": 0.4482968628508388,
+          "success_rate": 0.2604166666666667
+        },
+        "retrieval_residual_residual_wrong_gripper": {
+          "count": 420.0,
+          "mean_progress": 0.561073886946265,
+          "success_rate": 0.3595238095238095
+        }
+      },
+      "source": "results/h16_transport_field_reground_fieldonly_k6clean_dropnoopwg_b12_v1_besttransport_margin0p00_k6_srcscore001_metric_taskrel_summary.json",
+      "std_success": 0.0165902469811643
+    },
+    "transport_field_reground_fieldonly_k6matched_b12_clean_k6_dropnoopwg_retargeted_srcscore001_metric_taskrelz": {
+      "ci95_success": 0.05007659255026318,
+      "headline_ci95_success": 0.05007659255026318,
+      "headline_metric": "mean_success",
+      "headline_metric_label": "deployment",
+      "headline_std_success": 0.02015691437763908,
+      "headline_success": 0.36985507246376814,
+      "label": "K6-matched transported residual field re-grounding, source-score 0.01, task-relative z-score retrieval chart",
+      "mean_action_mse_to_best": 0.5108807214875908,
+      "mean_progress": 0.5848812929805928,
+      "mean_success": 0.36985507246376814,
+      "num_completed": 3,
+      "per_task_success": {
+        "LiftPegUpright-v1": {
+          "mean_num_groups": 102.0,
+          "mean_success": 0.27972502255471415,
+          "std_success": 0.07423460353404222
+        },
+        "PickCube-v1": {
+          "mean_num_groups": 196.66666666666666,
+          "mean_success": 0.32106953030866076,
+          "std_success": 0.029193894870027198
+        },
+        "PullCube-v1": {
+          "mean_num_groups": 81.0,
+          "mean_success": 0.21214905951748056,
+          "std_success": 0.030352394320702063
+        },
+        "PushCube-v1": {
+          "mean_num_groups": 102.0,
+          "mean_success": 0.8080342824391696,
+          "std_success": 0.05489764730606154
+        },
+        "StackCube-v1": {
+          "mean_num_groups": 93.33333333333333,
+          "mean_success": 0.22313242313242312,
+          "std_success": 0.05397365074207076
+        }
+      },
+      "seed_action_mse_to_best": {
+        "0": 0.4926190784978478,
+        "1": 0.5073619219069572,
+        "2": 0.5326611640579675
+      },
+      "seed_progress": {
+        "0": 0.5738849096746771,
+        "1": 0.5792882018633511,
+        "2": 0.6014707674037503
+      },
+      "seed_success": {
+        "0": 0.3565217391304348,
+        "1": 0.36,
+        "2": 0.39304347826086955
+      },
+      "selected_candidate_type_counts": {
+        "retrieval_residual_policy_residual": 557,
+        "retrieval_residual_residual_near_miss": 216,
+        "retrieval_residual_residual_near_miss+residual_wrong_gripper": 245,
+        "retrieval_residual_residual_no_op": 285,
+        "retrieval_residual_residual_wrong_gripper": 422
+      },
+      "selected_residual_scale_counts": {
+        "0.35": 918,
+        "0.4": 190,
+        "0.45": 617
+      },
+      "selected_type_outcomes": {
+        "retrieval_residual_policy_residual": {
+          "count": 557.0,
+          "mean_progress": 0.5446326207449823,
+          "success_rate": 0.27827648114901254
+        },
+        "retrieval_residual_residual_near_miss": {
+          "count": 216.0,
+          "mean_progress": 0.7734846497805847,
+          "success_rate": 0.6157407407407407
+        },
+        "retrieval_residual_residual_near_miss+residual_wrong_gripper": {
+          "count": 245.0,
+          "mean_progress": 0.7282262291531174,
+          "success_rate": 0.5346938775510204
+        },
+        "retrieval_residual_residual_no_op": {
+          "count": 285.0,
+          "mean_progress": 0.44615056981054835,
+          "success_rate": 0.24912280701754386
+        },
+        "retrieval_residual_residual_wrong_gripper": {
+          "count": 422.0,
+          "mean_progress": 0.5519403738043629,
+          "success_rate": 0.35071090047393366
+        }
+      },
+      "source": "results/h16_transport_field_reground_fieldonly_k6clean_dropnoopwg_b12_v1_besttransport_margin0p00_k6_srcscore001_metric_taskrelz_summary.json",
+      "std_success": 0.02015691437763908
+    },
+    "transport_field_reground_fieldonly_k6matched_b12_clean_k6_dropnoopwg_retargeted_srcscore001_metric_zscore": {
+      "ci95_success": 0.030346814107679115,
+      "headline_ci95_success": 0.030346814107679115,
+      "headline_metric": "mean_success",
+      "headline_metric_label": "deployment",
+      "headline_std_success": 0.012215250727945203,
+      "headline_success": 0.37855072463768114,
+      "label": "K6-matched transported residual field re-grounding, source-score 0.01, z-score retrieval chart",
+      "mean_action_mse_to_best": 0.5220838308536813,
+      "mean_progress": 0.5908983214655419,
+      "mean_success": 0.37855072463768114,
+      "num_completed": 3,
+      "per_task_success": {
+        "LiftPegUpright-v1": {
+          "mean_num_groups": 102.0,
+          "mean_success": 0.2826748750620888,
+          "std_success": 0.07248620066794445
+        },
+        "PickCube-v1": {
+          "mean_num_groups": 196.66666666666666,
+          "mean_success": 0.3291632884024188,
+          "std_success": 0.035632325041567985
+        },
+        "PullCube-v1": {
+          "mean_num_groups": 81.0,
+          "mean_success": 0.25937064884433303,
+          "std_success": 0.02465844091003031
+        },
+        "PushCube-v1": {
+          "mean_num_groups": 102.0,
+          "mean_success": 0.7981332923401597,
+          "std_success": 0.04807608059065016
+        },
+        "StackCube-v1": {
+          "mean_num_groups": 93.33333333333333,
+          "mean_success": 0.22620342620342623,
+          "std_success": 0.06728314128782689
+        }
+      },
+      "seed_action_mse_to_best": {
+        "0": 0.5014481430354973,
+        "1": 0.5208085850974464,
+        "2": 0.5439947644280999
+      },
+      "seed_progress": {
+        "0": 0.5823645833001027,
+        "1": 0.5842054872760428,
+        "2": 0.6061248938204802
+      },
+      "seed_success": {
+        "0": 0.3773913043478261,
+        "1": 0.36695652173913046,
+        "2": 0.391304347826087
+      },
+      "selected_candidate_type_counts": {
+        "retrieval_residual_policy_residual": 586,
+        "retrieval_residual_residual_near_miss": 188,
+        "retrieval_residual_residual_near_miss+residual_wrong_gripper": 232,
+        "retrieval_residual_residual_no_op": 246,
+        "retrieval_residual_residual_wrong_gripper": 473
+      },
+      "selected_residual_scale_counts": {
+        "0.35": 921,
+        "0.4": 217,
+        "0.45": 587
+      },
+      "selected_type_outcomes": {
+        "retrieval_residual_policy_residual": {
+          "count": 586.0,
+          "mean_progress": 0.5288170372090643,
+          "success_rate": 0.2627986348122867
+        },
+        "retrieval_residual_residual_near_miss": {
+          "count": 188.0,
+          "mean_progress": 0.8107344068269781,
+          "success_rate": 0.6861702127659575
+        },
+        "retrieval_residual_residual_near_miss+residual_wrong_gripper": {
+          "count": 232.0,
+          "mean_progress": 0.7362754722474573,
+          "success_rate": 0.5387931034482759
+        },
+        "retrieval_residual_residual_no_op": {
+          "count": 246.0,
+          "mean_progress": 0.5057666452934708,
+          "success_rate": 0.32113821138211385
+        },
+        "retrieval_residual_residual_wrong_gripper": {
+          "count": 473.0,
+          "mean_progress": 0.5534043296754171,
+          "success_rate": 0.35095137420718814
+        }
+      },
+      "source": "results/h16_transport_field_reground_fieldonly_k6clean_dropnoopwg_b12_v1_besttransport_margin0p00_k6_srcscore001_metric_zscore_summary.json",
+      "std_success": 0.012215250727945203
+    },
     "transport_field_reground_fieldonly_k6matched_b12_clean_k6_dropnoopwg_retargeted_srcscore001_oraclek8": {
       "candidate_oracle_best_branch_rank_counts": {
         "1": 984,
@@ -12272,6 +12560,102 @@
       "source": "results/h16_transport_field_reground_fieldonly_k6clean_dropnoopwg_b12_v1_besttransport_margin0p00_k6_srcscore005_summary.json",
       "std_success": 0.017764119937442965
     },
+    "transport_field_reground_fieldonly_k6matched_b12_clean_k6_dropnoopwg_retargeted_srcscore_task_pick001_stack005": {
+      "ci95_success": 0.04619865231337768,
+      "headline_ci95_success": 0.04619865231337768,
+      "headline_metric": "mean_success",
+      "headline_metric_label": "deployment",
+      "headline_std_success": 0.018595959341849776,
+      "headline_success": 0.3889855072463768,
+      "label": "K6-matched transported residual field re-grounding, task-conditioned source-score prior",
+      "mean_action_mse_to_best": 0.5131558791707284,
+      "mean_progress": 0.5995674399000362,
+      "mean_success": 0.3889855072463768,
+      "num_completed": 3,
+      "per_task_success": {
+        "LiftPegUpright-v1": {
+          "mean_num_groups": 102.0,
+          "mean_success": 0.27087546503259025,
+          "std_success": 0.08120108533113783
+        },
+        "PickCube-v1": {
+          "mean_num_groups": 196.66666666666666,
+          "mean_success": 0.3467330777113386,
+          "std_success": 0.018223409821817384
+        },
+        "PullCube-v1": {
+          "mean_num_groups": 81.0,
+          "mean_success": 0.2537340826814511,
+          "std_success": 0.02754240597049266
+        },
+        "PushCube-v1": {
+          "mean_num_groups": 102.0,
+          "mean_success": 0.8244328170597393,
+          "std_success": 0.05068828725828965
+        },
+        "StackCube-v1": {
+          "mean_num_groups": 93.33333333333333,
+          "mean_success": 0.24527324527324526,
+          "std_success": 0.03280317778959639
+        }
+      },
+      "seed_action_mse_to_best": {
+        "0": 0.4981627210083863,
+        "1": 0.5139227381794025,
+        "2": 0.5273821783243962
+      },
+      "seed_progress": {
+        "0": 0.5869067004063855,
+        "1": 0.5928394121466124,
+        "2": 0.6189562071471111
+      },
+      "seed_success": {
+        "0": 0.3773913043478261,
+        "1": 0.3791304347826087,
+        "2": 0.41043478260869565
+      },
+      "selected_candidate_type_counts": {
+        "retrieval_residual_policy_residual": 608,
+        "retrieval_residual_residual_near_miss": 204,
+        "retrieval_residual_residual_near_miss+residual_wrong_gripper": 212,
+        "retrieval_residual_residual_no_op": 234,
+        "retrieval_residual_residual_wrong_gripper": 467
+      },
+      "selected_residual_scale_counts": {
+        "0.35": 986,
+        "0.4": 206,
+        "0.45": 533
+      },
+      "selected_type_outcomes": {
+        "retrieval_residual_policy_residual": {
+          "count": 608.0,
+          "mean_progress": 0.5486438763006223,
+          "success_rate": 0.29605263157894735
+        },
+        "retrieval_residual_residual_near_miss": {
+          "count": 204.0,
+          "mean_progress": 0.793509843540104,
+          "success_rate": 0.6568627450980392
+        },
+        "retrieval_residual_residual_near_miss+residual_wrong_gripper": {
+          "count": 212.0,
+          "mean_progress": 0.7113450988599995,
+          "success_rate": 0.5094339622641509
+        },
+        "retrieval_residual_residual_no_op": {
+          "count": 234.0,
+          "mean_progress": 0.5392063331002226,
+          "success_rate": 0.3547008547008547
+        },
+        "retrieval_residual_residual_wrong_gripper": {
+          "count": 467.0,
+          "mean_progress": 0.5606486210938567,
+          "success_rate": 0.3554603854389722
+        }
+      },
+      "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_typesuccess001": {
       "ci95_success": 0.04240637681159422,
       "headline_ci95_success": 0.04240637681159422,
@@ -14554,38 +14938,38 @@
   },
   "paired_deltas": {
     "best_clean - canonical_h16": {
-      "ci95_delta": 0.05007659255026316,
-      "left": "transport_field_reground_fieldonly_k6matched_b12_clean_k6_dropnoopwg_retargeted_srcscore001",
-      "mean_delta": 0.09101449275362321,
+      "ci95_delta": 0.05208653786416468,
+      "left": "transport_field_reground_fieldonly_k6matched_b12_clean_k6_dropnoopwg_retargeted_srcscore_task_pick001_stack005",
+      "mean_delta": 0.09159420289855073,
       "right": "h16_policy_canonical",
       "seed_deltas": {
         "0": 0.09391304347826085,
         "1": 0.06956521739130439,
-        "2": 0.10956521739130437
+        "2": 0.11130434782608695
       },
       "seeds": [
         0,
         1,
         2
       ],
-      "std_delta": 0.02015691437763907
+      "std_delta": 0.02096596095075374
     },
     "best_clean - direct_same_ckpt": {
-      "ci95_delta": 0.05572274719395504,
-      "left": "transport_field_reground_fieldonly_k6matched_b12_clean_k6_dropnoopwg_retargeted_srcscore001",
-      "mean_delta": 0.1055072463768116,
+      "ci95_delta": 0.05812756914879615,
+      "left": "transport_field_reground_fieldonly_k6matched_b12_clean_k6_dropnoopwg_retargeted_srcscore_task_pick001_stack005",
+      "mean_delta": 0.10608695652173912,
       "right": "near_miss_policy_bc5",
       "seed_deltas": {
         "0": 0.09913043478260869,
         "1": 0.08695652173913043,
-        "2": 0.13043478260869568
+        "2": 0.13217391304347825
       },
       "seeds": [
         0,
         1,
         2
       ],
-      "std_delta": 0.02242961405467028
+      "std_delta": 0.02339760703838906
     },
     "full_lattice - no_expert_lattice": {
       "ci95_delta": 0.02628111194117426,
@@ -14645,14 +15029,14 @@
       "PickCube-v1": 0.1408757305496436,
       "PullCube-v1": 0.05587959798486114,
       "PushCube-v1": 0.07387564518383971,
-      "StackCube-v1": 0.10060310060310063
+      "StackCube-v1": 0.10430680430680431
     },
     "no_expert_vs_best_clean": {
       "LiftPegUpright-v1": 0.34628170331174174,
       "PickCube-v1": 0.27474578561535085,
       "PullCube-v1": -0.0552542973595605,
       "PushCube-v1": -0.023389831795140315,
-      "StackCube-v1": 0.24094794094794092
+      "StackCube-v1": 0.23724423724423724
     }
   }
 }

workspace/results/paper_analysis.md

@@ -1,6 +1,6 @@
 # Paper Analysis
 
-Generated: `2026-07-02T14:35:03+00:00`
+Generated: `2026-07-02T14:59:38+00:00`
 
 ## Main Seed Statistics
 
@@ -68,6 +68,10 @@ Generated: `2026-07-02T14:35:03+00:00`
 | transport_field_reground_fieldonly_k6matched_b12_clean_k6_dropnoopwg_retargeted_srcscore001_typecal005 | K6-matched transported residual field re-grounding, source-score 0.01 + transport-outcome type calibration 0.05 | 3 | deployment | 38.78% +/- 1.82 | 38.78% | +/- 4.51 | 59.90% | 0.511 | +9.04 pp |
 | transport_field_reground_fieldonly_k6matched_b12_clean_k6_dropnoopwg_retargeted_srcscore001_typecal010 | K6-matched transported residual field re-grounding, source-score 0.01 + transport-outcome type calibration 0.10 | 3 | deployment | 38.78% +/- 1.66 | 38.78% | +/- 4.12 | 59.87% | 0.509 | +9.04 pp |
 | transport_field_reground_fieldonly_k6matched_b12_clean_k6_dropnoopwg_retargeted_srcscore001_typecal020 | K6-matched transported residual field re-grounding, source-score 0.01 + transport-outcome type calibration 0.20 | 3 | deployment | 38.61% +/- 1.66 | 38.61% | +/- 4.12 | 59.87% | 0.507 | +8.87 pp |
+| transport_field_reground_fieldonly_k6matched_b12_clean_k6_dropnoopwg_retargeted_srcscore001_metric_zscore | K6-matched transported residual field re-grounding, source-score 0.01, z-score retrieval chart | 3 | deployment | 37.86% +/- 1.22 | 37.86% | +/- 3.03 | 59.09% | 0.522 | +8.12 pp |
+| 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_srcscore001_oraclek8 | K6-matched transported residual field re-grounding, source-score 0.01 candidate-oracle K8 | 3 | candidate-oracle | 44.35% +/- 1.98 | 38.84% | +/- 4.91 | 59.95% | 0.512 | +14.61 pp |
 | transport_field_reground_fieldonly_k6matched_b12_clean_k6_dropnoopwg_retargeted_srcscore0005 | K6-matched transported residual field re-grounding, exact drop-mask train source-score prior 0.005 | 3 | deployment | 38.72% +/- 1.71 | 38.72% | +/- 4.24 | 59.90% | 0.511 | +8.99 pp |
 | transport_field_reground_fieldonly_k6matched_b12_clean_k6_dropnoopwg_retargeted_srcscore0015 | K6-matched transported residual field re-grounding, exact drop-mask train source-score prior 0.015 | 3 | deployment | 38.78% +/- 1.82 | 38.78% | +/- 4.51 | 59.94% | 0.511 | +9.04 pp |
@@ -167,8 +171,8 @@ Generated: `2026-07-02T14:35:03+00:00`
 
 | comparison | seeds | mean delta | 95% CI | seed deltas |
 |---|---:|---:|---:|---|
-| best_clean - canonical_h16 | 3 | +9.10 pp | +/- 5.01 | 0:+9.39, 1:+6.96, 2:+10.96 |
-| best_clean - direct_same_ckpt | 3 | +10.55 pp | +/- 5.57 | 0:+9.91, 1:+8.70, 2:+13.04 |
+| best_clean - canonical_h16 | 3 | +9.16 pp | +/- 5.21 | 0:+9.39, 1:+6.96, 2:+11.13 |
+| best_clean - direct_same_ckpt | 3 | +10.61 pp | +/- 5.81 | 0:+9.91, 1:+8.70, 2:+13.22 |
 | no_expert_lattice - canonical_h16 | 3 | +27.25 pp | +/- 8.58 | 0:+23.30, 1:+28.70, 2:+29.74 |
 | full_lattice - no_expert_lattice | 3 | +12.35 pp | +/- 2.63 | 0:+13.57, 1:+11.83, 2:+11.65 |
 | policy_candidate_lattice - no_expert_lattice | 3 | -16.29 pp | +/- 7.55 | 0:-15.48, 1:-13.74, 2:-19.65 |
@@ -181,13 +185,13 @@ Generated: `2026-07-02T14:35:03+00:00`
 | PickCube-v1 | 20.59% | 34.67% | 58.13% | 62.15% | 84.19% | +14.09 pp | +27.47 pp |
 | PullCube-v1 | 19.79% | 25.37% | 15.02% | 19.85% | 22.41% | +5.59 pp | -5.53 pp |
 | PushCube-v1 | 75.06% | 82.44% | 82.54% | 80.10% | 81.92% | +7.39 pp | -2.34 pp |
-| StackCube-v1 | 14.10% | 24.16% | 50.41% | 48.25% | 60.83% | +10.06 pp | +24.09 pp |
+| StackCube-v1 | 14.10% | 24.53% | 50.41% | 48.25% | 60.83% | +10.43 pp | +23.72 pp |
 
 ## Mechanism Gap
 
-- Best clean residual transport improves over canonical h16 by +9.10 pp.
+- Best clean residual transport improves over canonical h16 by +9.16 pp.
 - Same-state no-expert lattice improves over canonical h16 by +27.25 pp.
-- Remaining clean-to-same-state proposal gap is +18.14 pp.
+- Remaining clean-to-same-state proposal gap is +18.09 pp.
 - Full lattice adds expert proposals and reaches 69.33%, a +12.35 pp gain over no-expert.
 
 ## Candidate-Oracle Diagnostic
@@ -206,8 +210,8 @@ Generated: `2026-07-02T14:35:03+00:00`
 - `same_state_no_expert`: lattice_near_miss=1263 (73.2%), lattice_no_op=222 (12.9%), lattice_random_negative=144 (8.3%), lattice_wrong_gripper=62 (3.6%), lattice_wrong_direction=34 (2.0%)
 - `same_state_policy_baseline`: policy_continuous=1022 (59.2%), lattice_near_miss=448 (26.0%), lattice_no_op=119 (6.9%), lattice_random_negative=75 (4.3%), lattice_wrong_gripper=45 (2.6%), lattice_wrong_direction=16 (0.9%)
 - `same_state_full`: lattice_expert=977 (56.6%), lattice_near_miss=348 (20.2%), lattice_no_op=177 (10.3%), lattice_random_negative=138 (8.0%), lattice_wrong_gripper=55 (3.2%), lattice_wrong_direction=30 (1.7%)
-- `transport_field_reground_fieldonly_k6matched_b12_clean_k6_dropnoopwg_retargeted_srcscore001`: retrieval_residual_policy_residual=609 (35.3%), retrieval_residual_residual_wrong_gripper=479 (27.8%), retrieval_residual_residual_no_op=236 (13.7%), retrieval_residual_residual_near_miss+residual_wrong_gripper=209 (12.1%), retrieval_residual_residual_near_miss=192 (11.1%)
-- `transport_field_reground_fieldonly_k6matched_b12_clean_k6_dropnoopwg_retargeted_srcscore001` residual scale counts: {'0.35': 985, '0.4': 207, '0.45': 533}
+- `transport_field_reground_fieldonly_k6matched_b12_clean_k6_dropnoopwg_retargeted_srcscore_task_pick001_stack005`: retrieval_residual_policy_residual=608 (35.2%), retrieval_residual_residual_wrong_gripper=467 (27.1%), retrieval_residual_residual_no_op=234 (13.6%), retrieval_residual_residual_near_miss+residual_wrong_gripper=212 (12.3%), retrieval_residual_residual_near_miss=204 (11.8%)
+- `transport_field_reground_fieldonly_k6matched_b12_clean_k6_dropnoopwg_retargeted_srcscore_task_pick001_stack005` residual scale counts: {'0.35': 986, '0.4': 206, '0.45': 533}
 
 ## Selected-Type Outcomes
 

workspace/results/paper_table_status.json

@@ -1695,6 +1695,82 @@
       "best_config": null,
       "gain_vs_h16_policy": 0.08869565217391306
     },
+    {
+      "key": "transport_field_reground_fieldonly_k6matched_b12_clean_k6_dropnoopwg_retargeted_srcscore001_metric_zscore",
+      "label": "K6-matched transported residual field re-grounding, source-score 0.01, z-score retrieval chart",
+      "path": "h16_transport_field_reground_fieldonly_k6clean_dropnoopwg_b12_v1_besttransport_margin0p00_k6_srcscore001_metric_zscore_summary.json",
+      "clean_deployment": "yes",
+      "same_state_proposals": "no",
+      "expert_proposal": "no",
+      "story_role": "train-bank chart normalization ablation for transported residual support selection",
+      "fallback_success": null,
+      "pending_job": "",
+      "path_exists": true,
+      "status": "complete",
+      "success": 0.3785507246376812,
+      "std_success": 0.012215250727945201,
+      "completed_seeds": null,
+      "num_completed": 3,
+      "best_config": null,
+      "gain_vs_h16_policy": 0.08115942028985512
+    },
+    {
+      "key": "transport_field_reground_fieldonly_k6matched_b12_clean_k6_dropnoopwg_retargeted_srcscore001_metric_taskrel",
+      "label": "K6-matched transported residual field re-grounding, source-score 0.01, task-relative retrieval chart",
+      "path": "h16_transport_field_reground_fieldonly_k6clean_dropnoopwg_b12_v1_besttransport_margin0p00_k6_srcscore001_metric_taskrel_summary.json",
+      "clean_deployment": "yes",
+      "same_state_proposals": "no",
+      "expert_proposal": "no",
+      "story_role": "object-relative local chart ablation for transported residual support selection",
+      "fallback_success": null,
+      "pending_job": "",
+      "path_exists": true,
+      "status": "complete",
+      "success": 0.37217391304347824,
+      "std_success": 0.0165902469811643,
+      "completed_seeds": null,
+      "num_completed": 3,
+      "best_config": null,
+      "gain_vs_h16_policy": 0.07478260869565218
+    },
+    {
+      "key": "transport_field_reground_fieldonly_k6matched_b12_clean_k6_dropnoopwg_retargeted_srcscore001_metric_taskrelz",
+      "label": "K6-matched transported residual field re-grounding, source-score 0.01, task-relative z-score retrieval chart",
+      "path": "h16_transport_field_reground_fieldonly_k6clean_dropnoopwg_b12_v1_besttransport_margin0p00_k6_srcscore001_metric_taskrelz_summary.json",
+      "clean_deployment": "yes",
+      "same_state_proposals": "no",
+      "expert_proposal": "no",
+      "story_role": "normalized object-relative local chart ablation for transported residual support selection",
+      "fallback_success": null,
+      "pending_job": "",
+      "path_exists": true,
+      "status": "complete",
+      "success": 0.3698550724637681,
+      "std_success": 0.02015691437763908,
+      "completed_seeds": null,
+      "num_completed": 3,
+      "best_config": null,
+      "gain_vs_h16_policy": 0.07246376811594202
+    },
+    {
+      "key": "transport_field_reground_fieldonly_k6matched_b12_clean_k6_dropnoopwg_retargeted_srcscore_task_pick001_stack005",
+      "label": "K6-matched transported residual field re-grounding, task-conditioned source-score prior",
+      "path": "h16_transport_field_reground_fieldonly_k6clean_dropnoopwg_b12_v1_besttransport_margin0p00_k6_srcscore_task_pick001_stack005_summary.json",
+      "clean_deployment": "yes",
+      "same_state_proposals": "no",
+      "expert_proposal": "no",
+      "story_role": "task-conditioned train-source utility prior chosen from per-task source-score sensitivity",
+      "fallback_success": null,
+      "pending_job": "",
+      "path_exists": true,
+      "status": "complete",
+      "success": 0.3889855072463768,
+      "std_success": 0.01859595934184978,
+      "completed_seeds": null,
+      "num_completed": 3,
+      "best_config": null,
+      "gain_vs_h16_policy": 0.09159420289855075
+    },
     {
       "key": "transport_field_reground_fieldonly_k6matched_b12_clean_k6_dropnoopwg_retargeted_srcscore001_oraclek8",
       "label": "K6-matched transported residual field re-grounding, source-score 0.01 candidate-oracle K8",
@@ -3882,23 +3958,23 @@
     }
   ],
   "best_clean": {
-    "key": "transport_field_reground_fieldonly_k6matched_b12_clean_k6_dropnoopwg_retargeted_srcscore001",
-    "label": "K6-matched transported residual field re-grounding, exact drop-mask train source-score prior 0.01",
-    "path": "h16_transport_field_reground_fieldonly_k6clean_dropnoopwg_b12_v1_besttransport_margin0p00_k6_srcscore001_summary.json",
+    "key": "transport_field_reground_fieldonly_k6matched_b12_clean_k6_dropnoopwg_retargeted_srcscore_task_pick001_stack005",
+    "label": "K6-matched transported residual field re-grounding, task-conditioned source-score prior",
+    "path": "h16_transport_field_reground_fieldonly_k6clean_dropnoopwg_b12_v1_besttransport_margin0p00_k6_srcscore_task_pick001_stack005_summary.json",
     "clean_deployment": "yes",
     "same_state_proposals": "no",
     "expert_proposal": "no",
-    "story_role": "small train-source utility prior sensitivity for selector calibration",
+    "story_role": "task-conditioned train-source utility prior chosen from per-task source-score sensitivity",
     "fallback_success": null,
     "pending_job": "",
     "path_exists": true,
     "status": "complete",
-    "success": 0.3884057971014493,
-    "std_success": 0.01759303293390554,
+    "success": 0.3889855072463768,
+    "std_success": 0.01859595934184978,
     "completed_seeds": null,
     "num_completed": 3,
     "best_config": null,
-    "gain_vs_h16_policy": 0.09101449275362322
+    "gain_vs_h16_policy": 0.09159420289855075
   },
   "best_mechanism_no_expert": {
     "key": "no_expert_lattice",
@@ -3923,7 +3999,7 @@
     "Use no-expert same-state lattice as the conservative mechanism result, not as deployment-clean inference.",
     "Use full lattice only as an upper result because it includes expert proposals.",
     "Do not claim external SOTA from this table alone; add current external baselines separately.",
-    "Current best clean deployment row is K6-matched transported residual field re-grounding, exact drop-mask train source-score prior 0.01 at 38.84%.",
+    "Current best clean deployment row is K6-matched transported residual field re-grounding, task-conditioned source-score prior at 38.90%.",
     "Trust-region field optimization should be framed as a negative/diagnostic ablation.",
     "Train-state counterfactual residual retrieval is a positive clean bridge but remains below the current clean best.",
     "KNN counterfactual residual retrieval is a positive clean bridge but remains below the current clean best.",

workspace/results/paper_table_status.md

@@ -92,6 +92,10 @@ Baseline h=16 policy: 29.74%
 | transport_field_reground_fieldonly_k6matched_b12_clean_k6_dropnoopwg_retargeted_srcscore001_typecal005 | K6-matched transported residual field re-grounding, source-score 0.01 + transport-outcome type calibration 0.05 | complete | 38.78% | +9.04 pp | yes | no | no | train counterfactual outcome-calibrated candidate-family prior over transported residual types |
 | transport_field_reground_fieldonly_k6matched_b12_clean_k6_dropnoopwg_retargeted_srcscore001_typecal010 | K6-matched transported residual field re-grounding, source-score 0.01 + transport-outcome type calibration 0.10 | complete | 38.78% | +9.04 pp | yes | no | no | stronger train counterfactual outcome-calibrated candidate-family prior |
 | transport_field_reground_fieldonly_k6matched_b12_clean_k6_dropnoopwg_retargeted_srcscore001_typecal020 | K6-matched transported residual field re-grounding, source-score 0.01 + transport-outcome type calibration 0.20 | complete | 38.61% | +8.87 pp | yes | no | no | aggressive train counterfactual outcome-calibrated candidate-family prior |
+| transport_field_reground_fieldonly_k6matched_b12_clean_k6_dropnoopwg_retargeted_srcscore001_metric_zscore | K6-matched transported residual field re-grounding, source-score 0.01, z-score retrieval chart | complete | 37.86% | +8.12 pp | yes | no | no | train-bank chart normalization ablation for transported residual support selection |
+| 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 | complete | 37.22% | +7.48 pp | yes | no | no | object-relative local chart ablation for transported residual support selection |
+| 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 | complete | 36.99% | +7.25 pp | yes | no | no | normalized object-relative local chart ablation for transported residual support selection |
+| 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 | complete | 38.90% | +9.16 pp | yes | no | no | task-conditioned train-source utility prior chosen from per-task source-score sensitivity |
 | transport_field_reground_fieldonly_k6matched_b12_clean_k6_dropnoopwg_retargeted_srcscore001_oraclek8 | K6-matched transported residual field re-grounding, source-score 0.01 candidate-oracle K8 | complete | 44.35% | +14.61 pp | diagnostic | no | no | diagnostic top-8 proposal-oracle ceiling for the current best source-score-calibrated selector |
 | transport_field_reground_fieldonly_k6matched_b12_clean_k6_dropnoopwg_retargeted_srcscore0005 | K6-matched transported residual field re-grounding, exact drop-mask train source-score prior 0.005 | complete | 38.72% | +8.99 pp | yes | no | no | very small train-source utility prior sensitivity for selector calibration |
 | transport_field_reground_fieldonly_k6matched_b12_clean_k6_dropnoopwg_retargeted_srcscore0015 | K6-matched transported residual field re-grounding, exact drop-mask train source-score prior 0.015 | complete | 38.78% | +9.04 pp | yes | no | no | midpoint train-source utility prior sensitivity for selector calibration |
@@ -213,7 +217,7 @@ Baseline h=16 policy: 29.74%
 - Use no-expert same-state lattice as the conservative mechanism result, not as deployment-clean inference.
 - Use full lattice only as an upper result because it includes expert proposals.
 - Do not claim external SOTA from this table alone; add current external baselines separately.
-- Current best clean deployment row is K6-matched transported residual field re-grounding, exact drop-mask train source-score prior 0.01 at 38.84%.
+- Current best clean deployment row is K6-matched transported residual field re-grounding, task-conditioned source-score prior at 38.90%.
 - Trust-region field optimization should be framed as a negative/diagnostic ablation.
 - Train-state counterfactual residual retrieval is a positive clean bridge but remains below the current clean best.
 - KNN counterfactual residual retrieval is a positive clean bridge but remains below the current clean best.

workspace/scripts/build_paper_analysis.py

@@ -553,6 +553,14 @@ METHODS = [
             "besttransport_margin0p00_k6_srcscore001_metric_taskrelz_summary.json"
         ),
     ),
+    MethodSpec(
+        key="transport_field_reground_fieldonly_k6matched_b12_clean_k6_dropnoopwg_retargeted_srcscore_task_pick001_stack005",
+        label="K6-matched transported residual field re-grounding, task-conditioned source-score prior",
+        summary_path=(
+            "h16_transport_field_reground_fieldonly_k6clean_dropnoopwg_b12_v1_"
+            "besttransport_margin0p00_k6_srcscore_task_pick001_stack005_summary.json"
+        ),
+    ),
     MethodSpec(
         key="transport_field_reground_fieldonly_k6matched_b12_clean_k6_dropnoopwg_retargeted_srcscore001_oraclek8",
         label="K6-matched transported residual field re-grounding, source-score 0.01 candidate-oracle K8",

workspace/scripts/build_paper_table_status.py

@@ -918,6 +918,15 @@ SPECS = [
         expert_proposal="no",
         story_role="normalized object-relative local chart ablation for transported residual support selection",
     ),
+    ResultSpec(
+        key="transport_field_reground_fieldonly_k6matched_b12_clean_k6_dropnoopwg_retargeted_srcscore_task_pick001_stack005",
+        label="K6-matched transported residual field re-grounding, task-conditioned source-score prior",
+        path="h16_transport_field_reground_fieldonly_k6clean_dropnoopwg_b12_v1_besttransport_margin0p00_k6_srcscore_task_pick001_stack005_summary.json",
+        clean_deployment="yes",
+        same_state_proposals="no",
+        expert_proposal="no",
+        story_role="task-conditioned train-source utility prior chosen from per-task source-score sensitivity",
+    ),
     ResultSpec(
         key="transport_field_reground_fieldonly_k6matched_b12_clean_k6_dropnoopwg_retargeted_srcscore001_oraclek8",
         label="K6-matched transported residual field re-grounding, source-score 0.01 candidate-oracle K8",

workspace/scripts/eval_maniskill_policy_rollout.py

@@ -68,6 +68,28 @@ def _load_candidate_type_bonus_map(path: Path | None) -> dict[str, dict[str, flo
     return bonuses
 
 
+def _load_source_score_bonus_map(path: Path | None) -> dict[str, float]:
+    if path is None:
+        return {}
+    payload = json.loads(path.read_text())
+    if not isinstance(payload, dict):
+        raise SystemExit(
+            "--retrieval-residual-source-score-bonus-map must contain a JSON object"
+        )
+    raw_bonuses = payload.get("retrieval_residual_source_score_bonus_by_task", payload)
+    if not isinstance(raw_bonuses, dict):
+        raise SystemExit(
+            "--retrieval-residual-source-score-bonus-map must contain a "
+            "task->scale object or retrieval_residual_source_score_bonus_by_task"
+        )
+    try:
+        return {str(task_id): float(scale) for task_id, scale in raw_bonuses.items()}
+    except (TypeError, ValueError) as exc:
+        raise SystemExit(
+            "--retrieval-residual-source-score-bonus-map values must be numeric"
+        ) from exc
+
+
 def main(argv: list[str] | None = None) -> int:
     parser = argparse.ArgumentParser(
         description="Execute a DoVLA policy checkpoint from restored ManiSkill CIL states."
@@ -77,6 +99,15 @@ def main(argv: list[str] | None = None) -> int:
     parser.add_argument("--out", type=Path, required=True)
     parser.add_argument("--device", default="auto")
     parser.add_argument("--all-groups", action="store_true")
+    parser.add_argument(
+        "--split",
+        choices=("validation", "train", "all"),
+        default="validation",
+        help=(
+            "Dataset split to evaluate. Retrieval modes hold out the evaluated split "
+            "from the retrieval bank; --all-groups is kept as an alias for --split all."
+        ),
+    )
     parser.add_argument("--max-groups", type=int, default=None)
     parser.add_argument("--group-batch-size", type=int, default=16)
     parser.add_argument(
@@ -233,6 +264,13 @@ def main(argv: list[str] | None = None) -> int:
         help="Scale for adding a train-source reward-score prior to each retrieved residual "
         "candidate before field selection. Score is progress plus terminal success.",
     )
+    parser.add_argument(
+        "--retrieval-residual-source-score-bonus-map",
+        type=Path,
+        default=None,
+        help="Optional JSON task->scale map overriding the global train-source reward-score "
+        "prior for retrieval_residual candidates.",
+    )
     parser.add_argument(
         "--retrieval-residual-source-advantage-bonus-scale",
         type=float,
@@ -484,12 +522,16 @@ def main(argv: list[str] | None = None) -> int:
     candidate_type_bonuses_by_task = _load_candidate_type_bonus_map(
         args.candidate_type_bonus_map
     )
+    source_score_bonus_by_task = _load_source_score_bonus_map(
+        args.retrieval_residual_source_score_bonus_map
+    )
     result = evaluate_maniskill_policy_rollout(
         args.checkpoint,
         args.dataset,
         output_path=args.out,
         device=args.device,
         all_groups=args.all_groups,
+        eval_split=args.split,
         max_groups=args.max_groups,
         group_batch_size=args.group_batch_size,
         sim_backend=args.sim_backend,
@@ -520,6 +562,7 @@ def main(argv: list[str] | None = None) -> int:
         retrieval_residual_source_score_bonus_scale=(
             args.retrieval_residual_source_score_bonus_scale
         ),
+        retrieval_residual_source_score_bonus_by_task=source_score_bonus_by_task,
         retrieval_residual_source_advantage_bonus_scale=(
             args.retrieval_residual_source_advantage_bonus_scale
         ),

workspace/scripts/slurm/eval_maniskill_policy_rollout.sbatch

@@ -39,6 +39,7 @@ GROUP_BATCH_SIZE="${GROUP_BATCH_SIZE:-8}"
 SIM_BACKEND="${SIM_BACKEND:-physx_cuda:0}"
 RENDER_BACKEND="${RENDER_BACKEND:-none}"
 ALL_GROUPS="${ALL_GROUPS:-0}"
+EVAL_SPLIT="${EVAL_SPLIT:-validation}"
 DEVICE="${DEVICE:-cuda}"
 SELECTION_MODE="${SELECTION_MODE:-policy}"
 NUM_CANDIDATES="${NUM_CANDIDATES:-1}"
@@ -63,6 +64,7 @@ RETRIEVAL_RESIDUAL_MIN_SOURCE_PROGRESS="${RETRIEVAL_RESIDUAL_MIN_SOURCE_PROGRESS
 RETRIEVAL_RESIDUAL_MIN_SOURCE_ADVANTAGE="${RETRIEVAL_RESIDUAL_MIN_SOURCE_ADVANTAGE:--1000000000.0}"
 RETRIEVAL_RESIDUAL_SOURCE_PROGRESS_BONUS_SCALE="${RETRIEVAL_RESIDUAL_SOURCE_PROGRESS_BONUS_SCALE:-0.0}"
 RETRIEVAL_RESIDUAL_SOURCE_SCORE_BONUS_SCALE="${RETRIEVAL_RESIDUAL_SOURCE_SCORE_BONUS_SCALE:-0.0}"
+RETRIEVAL_RESIDUAL_SOURCE_SCORE_BONUS_MAP="${RETRIEVAL_RESIDUAL_SOURCE_SCORE_BONUS_MAP:-}"
 RETRIEVAL_RESIDUAL_SOURCE_ADVANTAGE_BONUS_SCALE="${RETRIEVAL_RESIDUAL_SOURCE_ADVANTAGE_BONUS_SCALE:-0.0}"
 RETRIEVAL_RESIDUAL_COMPOSITE_L2_PENALTY_SCALE="${RETRIEVAL_RESIDUAL_COMPOSITE_L2_PENALTY_SCALE:-0.0}"
 RETRIEVAL_RESIDUAL_ACTION_L2_PENALTY="${RETRIEVAL_RESIDUAL_ACTION_L2_PENALTY:-0.0}"
@@ -155,6 +157,9 @@ fi
 if [[ -n "$FIELD_RANK_BIAS_MAP" ]]; then
   EXTRA_ARGS+=(--field-rank-bias-map "$FIELD_RANK_BIAS_MAP")
 fi
+if [[ -n "$RETRIEVAL_RESIDUAL_SOURCE_SCORE_BONUS_MAP" ]]; then
+  EXTRA_ARGS+=(--retrieval-residual-source-score-bonus-map "$RETRIEVAL_RESIDUAL_SOURCE_SCORE_BONUS_MAP")
+fi
 
 apptainer exec --nv \
   --env "LD_LIBRARY_PATH=$CPU_RENDER_LIBS/lib:$NATIVE_LIBS:/.singularity.d/libs,VK_ICD_FILENAMES=$VULKAN_ICD,VK_DRIVER_FILES=$VULKAN_ICD,XDG_RUNTIME_DIR=$RUNTIME_DIR,MESA_SHADER_CACHE_DIR=$CACHE_DIR,LIBGL_ALWAYS_SOFTWARE=1,LP_NUM_THREADS=1,SSL_CERT_FILE=$CA_BUNDLE,REQUESTS_CA_BUNDLE=$CA_BUNDLE,OMP_NUM_THREADS=1,OPENBLAS_NUM_THREADS=1,MKL_NUM_THREADS=1,DOVLA_TORCH_THREADS=1,MPLBACKEND=Agg,PYTHONDONTWRITEBYTECODE=1" \
@@ -165,6 +170,7 @@ apptainer exec --nv \
     --dataset "$DATASET" \
     --out "$OUT" \
     --device "$DEVICE" \
+    --split "$EVAL_SPLIT" \
     --group-batch-size "$GROUP_BATCH_SIZE" \
     --sim-backend "$SIM_BACKEND" \
     --render-backend "$RENDER_BACKEND" \