auto-sync 2026-07-02T18:12:23Z workspace (part 3)

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": "2026-07-02T17:31:38+00:00",
+  "generated_utc": "2026-07-02T18:32:12+00:00",
   "mechanism_gap": {
     "best_clean_vs_direct_same_ckpt": 0.1060869565217391,
     "best_clean_vs_h16": 0.09159420289855075,
@@ -12710,6 +12710,20 @@
       "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_advw1p0": {
+      "headline_metric": "mean_success",
+      "headline_metric_label": "deployment",
+      "label": "Generator V1 utility-weighted tangents, expert anchor, advantage weight 1.0",
+      "missing": true,
+      "source": "results/h16_transport_field_reground_fieldonly_k6clean_dropnoopwg_b12_v1_besttransport_margin0p00_k6_srcscore_task_pick001_stack005_advw1p0_summary.json"
+    },
+    "transport_field_reground_fieldonly_k6matched_b12_clean_k6_dropnoopwg_retargeted_srcscore_task_pick001_stack005_advw2p0": {
+      "headline_metric": "mean_success",
+      "headline_metric_label": "deployment",
+      "label": "Generator V1 utility-weighted tangents, expert anchor, advantage weight 2.0",
+      "missing": true,
+      "source": "results/h16_transport_field_reground_fieldonly_k6clean_dropnoopwg_b12_v1_besttransport_margin0p00_k6_srcscore_task_pick001_stack005_advw2p0_summary.json"
+    },
     "transport_field_reground_fieldonly_k6matched_b12_clean_k6_dropnoopwg_retargeted_srcscore_task_pick001_stack005_oraclecal_train800_k4_r0025_t0025": {
       "ci95_success": 0.044132081897821224,
       "headline_ci95_success": 0.044132081897821224,
@@ -13094,6 +13108,13 @@
       "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_advw2p0": {
+      "headline_metric": "mean_success",
+      "headline_metric_label": "deployment",
+      "label": "Generator V1 positive tangents, policy anchor, advantage weight 2.0",
+      "missing": true,
+      "source": "results/h16_transport_field_reground_fieldonly_k6clean_dropnoopwg_b12_v1_besttransport_margin0p00_k6_srcscore_task_pick001_stack005_policyanchor_advw2p0_summary.json"
+    },
     "transport_field_reground_fieldonly_k6matched_b12_clean_k6_dropnoopwg_retargeted_srcscore_task_pick001_stack005_scaleoracle_trainall_k4_scale0025": {
       "ci95_success": 0.043707043917985716,
       "headline_ci95_success": 0.043707043917985716,

workspace/results/paper_analysis.md

@@ -1,6 +1,6 @@
 # Paper Analysis
 
-Generated: `2026-07-02T17:31:38+00:00`
+Generated: `2026-07-02T18:32:12+00:00`
 
 ## Main Seed Statistics
 
@@ -72,6 +72,9 @@ Generated: `2026-07-02T17:31:38+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_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/scripts/build_paper_analysis.py

@@ -574,6 +574,30 @@ METHODS = [
             "besttransport_margin0p00_k6_srcscore_task_pick001_stack005_summary.json"
         ),
     ),
+    MethodSpec(
+        key="transport_field_reground_fieldonly_k6matched_b12_clean_k6_dropnoopwg_retargeted_srcscore_task_pick001_stack005_advw1p0",
+        label="Generator V1 utility-weighted tangents, expert anchor, advantage weight 1.0",
+        summary_path=(
+            "h16_transport_field_reground_fieldonly_k6clean_dropnoopwg_b12_v1_"
+            "besttransport_margin0p00_k6_srcscore_task_pick001_stack005_advw1p0_summary.json"
+        ),
+    ),
+    MethodSpec(
+        key="transport_field_reground_fieldonly_k6matched_b12_clean_k6_dropnoopwg_retargeted_srcscore_task_pick001_stack005_advw2p0",
+        label="Generator V1 utility-weighted tangents, expert anchor, advantage weight 2.0",
+        summary_path=(
+            "h16_transport_field_reground_fieldonly_k6clean_dropnoopwg_b12_v1_"
+            "besttransport_margin0p00_k6_srcscore_task_pick001_stack005_advw2p0_summary.json"
+        ),
+    ),
+    MethodSpec(
+        key="transport_field_reground_fieldonly_k6matched_b12_clean_k6_dropnoopwg_retargeted_srcscore_task_pick001_stack005_policyanchor_advw2p0",
+        label="Generator V1 positive tangents, policy anchor, advantage weight 2.0",
+        summary_path=(
+            "h16_transport_field_reground_fieldonly_k6clean_dropnoopwg_b12_v1_"
+            "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_scaleoracle_trainall_k4_scale0025",
         label="K6-matched transported residual field re-grounding, all-train tangent-length prior 0.025",

workspace/scripts/eval_maniskill_policy_rollout.py

@@ -305,6 +305,13 @@ def main(argv: list[str] | None = None) -> int:
         help="Scale for adding a train-source reward-score advantage prior relative to the "
         "source residual anchor action.",
     )
+    parser.add_argument(
+        "--retrieval-residual-source-advantage-weight-scale",
+        type=float,
+        default=0.0,
+        help="Scale for exponentiating train-source reward-score advantage into residual "
+        "consensus weights. This changes proposal geometry before field selection.",
+    )
     parser.add_argument(
         "--retrieval-residual-composite-l2-penalty-scale",
         type=float,
@@ -605,6 +612,9 @@ def main(argv: list[str] | None = None) -> int:
         retrieval_residual_source_advantage_bonus_scale=(
             args.retrieval_residual_source_advantage_bonus_scale
         ),
+        retrieval_residual_source_advantage_weight_scale=(
+            args.retrieval_residual_source_advantage_weight_scale
+        ),
         retrieval_residual_composite_l2_penalty_scale=(
             args.retrieval_residual_composite_l2_penalty_scale
         ),