auto-sync 2026-07-03T16:58:17Z workspace (part 6)

workspace/runs/summary_ctt.csv

@@ -1,5 +1,8 @@
 method,run_path,clean,status,gate,base_success,success,proposal_oracle,hidden_oracle,support_gap,selector_gap,success_support_gap,success_selector_gap,outcome_ptr,pptc_0p20,pptc_0p40,negative_near_0p20,calibration_ece,seeds,ci
 V0 residual transport diagnostic,runs/reproduce_v0,yes,measured diagnostic baseline,legacy diagnostic,0.2974,0.3890,0.4435,n/a,0.1264,0.0545,0.1264,0.0545,n/a,0.2366,0.5269,0.0533,n/a,3,see run
+CTT residual base-context-observation,runs/ctt_base_context_obs_nonlinear_dominance_basic_positive_train_to_test,yes,train feature=base_context_obs,not measured,n/a,n/a,n/a,n/a,n/a,n/a,n/a,n/a,n/a,n/a,n/a,n/a,n/a,120,see run
+CTT residual base-context-observation,runs/ctt_base_context_obs_nonlinear_dominance_context_success_train_to_test,yes,train feature=base_context_obs,not measured,n/a,n/a,n/a,n/a,n/a,n/a,n/a,n/a,n/a,n/a,n/a,n/a,n/a,120,see run
+CTT residual base-context-observation,runs/ctt_base_context_obs_nonlinear_dominance_context_tangent_positive_train_to_test,yes,train feature=base_context_obs,not measured,n/a,n/a,n/a,n/a,n/a,n/a,n/a,n/a,n/a,n/a,n/a,n/a,n/a,120,see run
 CTT residual test measured aggregate,runs/ctt_base_context_obs_test_rollout_comparison,yes,measured rollout K=8,"fail:selected<47.45,support_gap>7,selector_gap>3",0.2917,0.2708,0.5139,0.7292,0.4885,0.4814,0.2639,0.2431,0.5347,n/a,n/a,n/a,0.2247,3,see measured_metrics
 CTT residual validation measured aggregate,runs/ctt_base_context_obs_train_cal_rollout_comparison,yes,measured rollout K=8,"fail:selected<47.45,oracle<50,support_gap>7,selector_gap>3",0.2778,0.2546,0.4722,0.6597,0.4570,0.4219,0.2315,0.2176,0.4907,n/a,n/a,n/a,0.2258,3,see measured_metrics
 CTT residual validation measured aggregate,runs/ctt_base_context_obs_val_rollout_comparison,yes,measured rollout K=8,"fail:selected<47.45,oracle<50,support_gap>7,selector_gap>3",0.2754,0.2415,0.4058,0.6667,0.5080,0.3118,0.2754,0.1643,0.5024,n/a,n/a,n/a,0.2175,3,see measured_metrics
@@ -70,3 +73,6 @@ CTT residual test dominance tau0,runs/ctt_dominance_val_to_test_tau0,yes,measure
 "CTT residual test learned dominance (basic, success_weighted_margin)",runs/ctt_learned_dominance_success_weighted_val_to_test,yes,measured dominance K=8 coverage=0.1944 fallback=0.8056 target=success_weighted_margin features=basic,"fail:selected<47.45,support_gap>7,selector_gap>3",0.2847,0.2986,0.5069,0.7292,0.4833,0.4794,0.2639,0.2708,0.5278,n/a,n/a,n/a,n/a,3,see run
 "CTT residual test learned dominance (tangent, utility_margin)",runs/ctt_learned_dominance_tangent_val_to_test,yes,measured dominance K=8 coverage=0.1042 fallback=0.8958 target=utility_margin features=tangent,"fail:selected<47.45,support_gap>7,selector_gap>3",0.2847,0.2986,0.5069,0.7292,0.4833,0.4597,0.2639,0.2708,0.5278,n/a,n/a,n/a,n/a,3,see run
 "CTT residual test learned dominance (basic, utility_margin)",runs/ctt_learned_dominance_val_to_test,yes,measured dominance K=8 coverage=0.2431 fallback=0.7569,"fail:selected<47.45,support_gap>7,selector_gap>3",0.2847,0.3056,0.5069,0.7292,0.4833,0.4730,0.2639,0.2569,0.5278,n/a,n/a,n/a,n/a,3,see run
+"CTT residual test nonlinear dominance (rf_regressor, basic, positive_margin)",runs/ctt_base_context_obs_nonlinear_dominance_basic_positive_train_to_test,yes,measured dominance K=8 coverage=0.3472 fallback=0.6528 target=positive_margin features=basic,"fail:selected<47.45,support_gap>7,selector_gap>3",0.2917,0.3056,0.5139,0.7292,0.4885,0.4641,0.2639,0.2569,0.5347,n/a,n/a,n/a,n/a,3,see run
+"CTT residual test nonlinear dominance (hgb_regressor, context, success)",runs/ctt_base_context_obs_nonlinear_dominance_context_success_train_to_test,yes,measured dominance K=8 coverage=0.3750 fallback=0.6250 target=success features=context,"fail:selected<47.45,support_gap>7,selector_gap>3",0.2917,0.2569,0.5139,0.7292,0.4885,0.5467,0.2639,0.3056,0.5347,n/a,n/a,n/a,n/a,3,see run
+"CTT residual test nonlinear dominance (hgb_classifier, context_tangent, positive_margin)",runs/ctt_base_context_obs_nonlinear_dominance_context_tangent_positive_train_to_test,yes,measured dominance K=8 coverage=0.4028 fallback=0.5972 target=positive_margin features=context_tangent,"fail:selected<47.45,support_gap>7,selector_gap>3",0.2917,0.2986,0.5139,0.7292,0.4885,0.4531,0.2639,0.2431,0.5347,n/a,n/a,n/a,n/a,3,see run

workspace/runs/summary_ctt.md

@@ -3,6 +3,9 @@
 | Method | Status | Base | Selected | Proposal oracle | Hidden oracle | Utility support gap | Utility selector gap | Success support gap | Success selector gap | OutcomePTR | PPTC@0.20 | PPTC@0.40 | NegativeNear@0.20 | Calibration ECE | Run |
 | --- | --- | ---: | ---: | ---: | ---: | ---: | ---: | ---: | ---: | ---: | ---: | ---: | ---: | ---: | --- |
 | V0 residual transport diagnostic | measured diagnostic baseline; gate: legacy diagnostic | 0.2974 | 0.3890 | 0.4435 | n/a | 0.1264 | 0.0545 | 0.1264 | 0.0545 | n/a | 0.2366 | 0.5269 | 0.0533 | n/a | `runs/reproduce_v0` |
+| CTT residual base-context-observation | train feature=base_context_obs; gate: not measured | n/a | n/a | n/a | n/a | n/a | n/a | n/a | n/a | n/a | n/a | n/a | n/a | n/a | `runs/ctt_base_context_obs_nonlinear_dominance_basic_positive_train_to_test` |
+| CTT residual base-context-observation | train feature=base_context_obs; gate: not measured | n/a | n/a | n/a | n/a | n/a | n/a | n/a | n/a | n/a | n/a | n/a | n/a | n/a | `runs/ctt_base_context_obs_nonlinear_dominance_context_success_train_to_test` |
+| CTT residual base-context-observation | train feature=base_context_obs; gate: not measured | n/a | n/a | n/a | n/a | n/a | n/a | n/a | n/a | n/a | n/a | n/a | n/a | n/a | `runs/ctt_base_context_obs_nonlinear_dominance_context_tangent_positive_train_to_test` |
 | CTT residual test measured aggregate | measured rollout K=8; gate: fail:selected<47.45,support_gap>7,selector_gap>3 | 0.2917 | 0.2708 | 0.5139 | 0.7292 | 0.4885 | 0.4814 | 0.2639 | 0.2431 | 0.5347 | n/a | n/a | n/a | 0.2247 | `runs/ctt_base_context_obs_test_rollout_comparison` |
 | CTT residual validation measured aggregate | measured rollout K=8; gate: fail:selected<47.45,oracle<50,support_gap>7,selector_gap>3 | 0.2778 | 0.2546 | 0.4722 | 0.6597 | 0.4570 | 0.4219 | 0.2315 | 0.2176 | 0.4907 | n/a | n/a | n/a | 0.2258 | `runs/ctt_base_context_obs_train_cal_rollout_comparison` |
 | CTT residual validation measured aggregate | measured rollout K=8; gate: fail:selected<47.45,oracle<50,support_gap>7,selector_gap>3 | 0.2754 | 0.2415 | 0.4058 | 0.6667 | 0.5080 | 0.3118 | 0.2754 | 0.1643 | 0.5024 | n/a | n/a | n/a | 0.2175 | `runs/ctt_base_context_obs_val_rollout_comparison` |
@@ -73,5 +76,8 @@
 | CTT residual test learned dominance (basic, success_weighted_margin) | measured dominance K=8 coverage=0.1944 fallback=0.8056 target=success_weighted_margin features=basic; gate: fail:selected<47.45,support_gap>7,selector_gap>3 | 0.2847 | 0.2986 | 0.5069 | 0.7292 | 0.4833 | 0.4794 | 0.2639 | 0.2708 | 0.5278 | n/a | n/a | n/a | n/a | `runs/ctt_learned_dominance_success_weighted_val_to_test` |
 | CTT residual test learned dominance (tangent, utility_margin) | measured dominance K=8 coverage=0.1042 fallback=0.8958 target=utility_margin features=tangent; gate: fail:selected<47.45,support_gap>7,selector_gap>3 | 0.2847 | 0.2986 | 0.5069 | 0.7292 | 0.4833 | 0.4597 | 0.2639 | 0.2708 | 0.5278 | n/a | n/a | n/a | n/a | `runs/ctt_learned_dominance_tangent_val_to_test` |
 | CTT residual test learned dominance (basic, utility_margin) | measured dominance K=8 coverage=0.2431 fallback=0.7569; gate: fail:selected<47.45,support_gap>7,selector_gap>3 | 0.2847 | 0.3056 | 0.5069 | 0.7292 | 0.4833 | 0.4730 | 0.2639 | 0.2569 | 0.5278 | n/a | n/a | n/a | n/a | `runs/ctt_learned_dominance_val_to_test` |
+| CTT residual test nonlinear dominance (rf_regressor, basic, positive_margin) | measured dominance K=8 coverage=0.3472 fallback=0.6528 target=positive_margin features=basic; gate: fail:selected<47.45,support_gap>7,selector_gap>3 | 0.2917 | 0.3056 | 0.5139 | 0.7292 | 0.4885 | 0.4641 | 0.2639 | 0.2569 | 0.5347 | n/a | n/a | n/a | n/a | `runs/ctt_base_context_obs_nonlinear_dominance_basic_positive_train_to_test` |
+| CTT residual test nonlinear dominance (hgb_regressor, context, success) | measured dominance K=8 coverage=0.3750 fallback=0.6250 target=success features=context; gate: fail:selected<47.45,support_gap>7,selector_gap>3 | 0.2917 | 0.2569 | 0.5139 | 0.7292 | 0.4885 | 0.5467 | 0.2639 | 0.3056 | 0.5347 | n/a | n/a | n/a | n/a | `runs/ctt_base_context_obs_nonlinear_dominance_context_success_train_to_test` |
+| CTT residual test nonlinear dominance (hgb_classifier, context_tangent, positive_margin) | measured dominance K=8 coverage=0.4028 fallback=0.5972 target=positive_margin features=context_tangent; gate: fail:selected<47.45,support_gap>7,selector_gap>3 | 0.2917 | 0.2986 | 0.5139 | 0.7292 | 0.4885 | 0.4531 | 0.2639 | 0.2431 | 0.5347 | n/a | n/a | n/a | n/a | `runs/ctt_base_context_obs_nonlinear_dominance_context_tangent_positive_train_to_test` |
 
 `n/a` means the required measured artifact does not exist yet. Measured gates use selected >=47.45%, proposal oracle >=50%, success support gap <=7 points, success selector gap <=3 points, OutcomePTR above the V0 proposal oracle reference, and at least three train seeds.

workspace/scripts/audit_chart_feature_sources.py

@@ -63,6 +63,7 @@ def main(argv: list[str] | None = None) -> int:
                     counts["rows"] += 1
                     for key in (
                         "observation_embedding_path",
+                        "object_embedding_path",
                         "observation_ref",
                         "scene_id",
                         "instruction",
@@ -107,6 +108,10 @@ def main(argv: list[str] | None = None) -> int:
                 ),
                 "observation_ref_present": int(counts["observation_ref_present"]),
                 "observation_ref_rate": _rate(counts["observation_ref_present"], rows),
+                "object_embedding_path_present": int(counts["object_embedding_path_present"]),
+                "object_embedding_path_rate": _rate(
+                    counts["object_embedding_path_present"], rows
+                ),
                 "scene_id_present": int(counts["scene_id_present"]),
                 "scene_id_rate": _rate(counts["scene_id_present"], rows),
                 "instruction_present": int(counts["instruction_present"]),
@@ -151,6 +156,12 @@ def _rate(count: int, total: int) -> float:
 
 def _conclusion(rows: list[dict[str, Any]]) -> str:
     if all(float(row["observation_embedding_path_rate"]) > 0.0 for row in rows):
+        if all(float(row["object_embedding_path_rate"]) > 0.0 for row in rows):
+            return (
+                "Current chart exports expose observation and object-layout embeddings "
+                "in every split; visual/object-layout chart features can be evaluated "
+                "with leakage-audited indexes."
+            )
         return (
             "Current chart exports expose observation embeddings in every split; "
             "visual chart features can be evaluated with leakage-audited indexes."
@@ -185,15 +196,16 @@ def _metrics_by_task(rows: list[dict[str, Any]]) -> str:
 def _table(rows: list[dict[str, Any]]) -> str:
     lines = [
         "% Auto-generated by scripts/audit_chart_feature_sources.py",
-        "\\begin{tabular}{lrrrr}",
+        "\\begin{tabular}{lrrrrr}",
         "\\toprule",
-        "Split & Rows & ObsEmbed & ObsRef & Instruction \\\\",
+        "Split & Rows & ObsEmbed & ObjEmbed & ObsRef & Instruction \\\\",
         "\\midrule",
     ]
     for row in rows:
         lines.append(
             f"{row['split']} & {row['rows']} & "
             f"{row['observation_embedding_path_present']} & "
+            f"{row['object_embedding_path_present']} & "
             f"{row['observation_ref_present']} & "
             f"{row['instruction_present']} \\\\"
         )
@@ -207,13 +219,14 @@ def _report(metrics: dict[str, Any]) -> str:
         "",
         metrics["conclusion"],
         "",
-        "| Split | Rows | Obs embedding path | Obs ref | Scene id | Instruction | Feature dims |",
-        "| --- | ---: | ---: | ---: | ---: | ---: | --- |",
+        "| Split | Rows | Obs embedding path | Object embedding path | Obs ref | Scene id | Instruction | Feature dims |",
+        "| --- | ---: | ---: | ---: | ---: | ---: | ---: | --- |",
     ]
     for row in metrics["splits"]:
         lines.append(
             f"| {row['split']} | {row['rows']} | "
             f"{row['observation_embedding_path_present']} ({row['observation_embedding_path_rate']:.2%}) | "
+            f"{row['object_embedding_path_present']} ({row['object_embedding_path_rate']:.2%}) | "
             f"{row['observation_ref_present']} ({row['observation_ref_rate']:.2%}) | "
             f"{row['scene_id_present']} ({row['scene_id_rate']:.2%}) | "
             f"{row['instruction_present']} ({row['instruction_rate']:.2%}) | "

workspace/scripts/build_ctt_proxy_comparison.py

@@ -24,6 +24,14 @@ DEFAULT_RUNS = [
         "ctt_residual_base_context_obs",
         sorted(Path("runs").glob("ctt_residual_base_context_obs_seed*_val_proxy")),
     ),
+    (
+        "ctt_residual_base_context_obj",
+        sorted(Path("runs").glob("ctt_residual_base_context_obj_seed*_val_proxy")),
+    ),
+    (
+        "ctt_residual_base_context_obs_obj",
+        sorted(Path("runs").glob("ctt_residual_base_context_obs_obj_seed*_val_proxy")),
+    ),
     (
         "ctt_gated_residual_full",
         sorted(Path("runs").glob("ctt_gated_residual_full_seed*_val_proxy")),