Drive corpus sampling from calibration failures

scripts/build_failure_driven_yolo_sampling.py

@@ -34,7 +34,10 @@ def build_sampling(
 
     samples = {item["sample_slug"]: item for item in manifest["samples"]}
     gates = assessment["gates"]
-    regions = assessment["test"]["regions"]
+    evaluation = assessment.get("test") or assessment.get("calibration")
+    if not evaluation or "regions" not in evaluation:
+        raise ValueError("Assessment has no regional calibration or test evidence")
+    regions = evaluation["regions"]
     weak_recall_regions = {
         region
         for region, metrics in regions.items()
@@ -46,8 +49,10 @@ def build_sampling(
         for region, metrics in regions.items()
         if metrics["precision"] < gates["min_region_precision"]
     }
-    background_failed = (
-        assessment["background"]["pure_empty_false_positives"]
+    background = assessment.get("background")
+    background_failed = bool(
+        background
+        and background["pure_empty_false_positives"]
         > gates["max_pure_empty_false_positives"]
     )
 
@@ -77,6 +82,7 @@ def build_sampling(
         "schema_version": 1,
         "status": "ok",
         "strategy": "failed-region-positive-and-hard-negative-repeat",
+        "failure_evidence_source": "test" if assessment.get("test") else "calibration",
         "weak_recall_regions": sorted(weak_recall_regions),
         "weak_precision_regions": sorted(weak_precision_regions),
         "background_gate_failed": background_failed,