auto-sync 2026-07-03T15:40:37Z workspace (part 4)

workspace/runs/leakage_audit_rgb_refs/report.json

@@ -3,7 +3,7 @@
   "indexes": {
     "test": {
       "audience": "evaluator_only",
-      "content_hash": "a249a4d627b8113988660be3aecd867ba1b7b8b8d00f66eb0e3070439a395554",
+      "content_hash": "d258ed27d587e6f41ef6ffd520832087fe7626b5e5cba2bd6352f50a1f45125a",
       "include_outcomes": true,
       "num_groups_exported": 410,
       "num_rows": 6560,
@@ -13,7 +13,7 @@
     },
     "train": {
       "audience": "train_retrieval",
-      "content_hash": "2cad4dbf330d10ecaeb4ed4e86ceb162850223d2d9618170b6800b1c73e3d531",
+      "content_hash": "1d15143588697e89f7fc2f6375b2745e4d479c726c9a947ff7eb7b2705280e1a",
       "include_outcomes": true,
       "num_groups_exported": 2044,
       "num_rows": 32704,
@@ -23,7 +23,7 @@
     },
     "val": {
       "audience": "evaluator_only",
-      "content_hash": "bad9b7241350f81c31e394e231d066a14138a61bb8e23a68e13e15eb5ac65d64",
+      "content_hash": "577761809f93f760fc7de6558a25d8085b2a0e51ca5d969680e99bcfa8586850",
       "include_outcomes": true,
       "num_groups_exported": 419,
       "num_rows": 6704,
@@ -37,6 +37,8 @@
   "warnings": [
     "runs/chart_feature_audit/metrics.json: missing data_hash",
     "runs/chart_feature_audit/metrics.json: missing split_hash",
+    "runs/chart_feature_audit_rgb_refs/metrics.json: missing data_hash",
+    "runs/chart_feature_audit_rgb_refs/metrics.json: missing split_hash",
     "runs/ctt_dominance_utility_energy_smoke_val_to_test/metrics.json: missing data_hash",
     "runs/ctt_dominance_utility_energy_smoke_val_to_test/metrics.json: missing split_hash",
     "runs/ctt_dominance_utility_energy_val_to_test_seed0/metrics.json: missing data_hash",

workspace/runs/leakage_audit_rgb_refs/report.md

@@ -16,6 +16,8 @@ Status: `pass`
 
 - runs/chart_feature_audit/metrics.json: missing data_hash
 - runs/chart_feature_audit/metrics.json: missing split_hash
+- runs/chart_feature_audit_rgb_refs/metrics.json: missing data_hash
+- runs/chart_feature_audit_rgb_refs/metrics.json: missing split_hash
 - runs/ctt_dominance_utility_energy_smoke_val_to_test/metrics.json: missing data_hash
 - runs/ctt_dominance_utility_energy_smoke_val_to_test/metrics.json: missing split_hash
 - runs/ctt_dominance_utility_energy_val_to_test_seed0/metrics.json: missing data_hash

workspace/scripts/audit_chart_feature_sources.py

@@ -150,12 +150,28 @@ 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["observation_embedding_path_rate"]) > 0.0 for row in rows):
+        return (
+            "Current chart exports expose observation embeddings in every split; "
+            "visual chart features can be evaluated with leakage-audited indexes."
+        )
+    if all(float(row["observation_ref_rate"]) > 0.0 for row in rows):
+        return (
+            "Current chart exports expose raw observation refs in every split but "
+            "not observation embeddings; run scripts/export_chart_observation_embeddings.py "
+            "before claiming embedded visual chart tokens."
+        )
+    if all(float(row["observation_embedding_path_rate"]) == 0.0 for row in rows) and all(
+        float(row["observation_ref_rate"]) == 0.0 for row in rows
+    ):
         return (
             "Current chart exports do not contain observation embeddings or raw observation refs; "
             "visual/object-centric chart tokens require a new export or embedding pass."
         )
-    return "At least one split exposes observation references; visual chart features can be wired next."
+    return (
+        "Observation feature availability is partial across splits; inspect per-split "
+        "embedding/ref rates before running visual chart-token experiments."
+    )
 
 
 def _metrics_by_task(rows: list[dict[str, Any]]) -> str:

workspace/scripts/eval_ctt_generated_rollout.py

@@ -293,7 +293,7 @@ def load_chart_items(
                     _json_loads(str(metadata_values[row]))
                     if metadata_values is not None
                     else {}
-                )
+                ) | {"_chart_root": str(index_path.parent)}
                 shape = tuple(int(value) for value in action_shapes[row])
                 flat_count = int(math.prod(shape))
                 base_action = np.asarray(