docs(accuracy): refresh governed scan and training evidence

scripts/run_belgium_building_training_loop.py

@@ -277,6 +277,54 @@ def failure_sampling_command(
     return command
 
 
+def protected_assessment_command(
+    *,
+    scripts_dir: Path,
+    calibration: Path,
+    test: Path,
+    background: Path,
+    output: Path,
+    selected_threshold: float,
+    min_aggregate_f1: float,
+    min_region_f1: float,
+    min_region_precision: float,
+    min_region_recall: float,
+    max_pure_empty_fp: int,
+) -> list[str]:
+    """Build the one-shot protected assessment command from frozen calibration gates.
+
+    The protected assessor must not silently reselect an operating point or
+    fall back to its own default gates.  Keeping command construction in one
+    helper also makes the release boundary directly testable without opening
+    any protected inputs.
+    """
+
+    return [
+        sys.executable,
+        str(scripts_dir / "assess_belgium_building_training_iteration.py"),
+        "--calibration",
+        str(calibration),
+        "--test",
+        str(test),
+        "--background",
+        str(background),
+        "--output",
+        str(output),
+        "--selected-threshold",
+        str(selected_threshold),
+        "--min-aggregate-f1",
+        str(min_aggregate_f1),
+        "--min-region-f1",
+        str(min_region_f1),
+        "--min-region-precision",
+        str(min_region_precision),
+        "--min-region-recall",
+        str(min_region_recall),
+        "--max-pure-empty-fp",
+        str(max_pure_empty_fp),
+    ]
+
+
 def resumable_training_command(yolo: str, checkpoint: Path) -> list[str]:
     return [yolo, "train", f"resume={checkpoint}", "device=0"]
 
@@ -572,14 +620,19 @@ def main() -> int:
                     iteration_dir / f"{role}.log",
                 )
             run(
-                [
-                    sys.executable,
-                    str(scripts_dir / "assess_belgium_building_training_iteration.py"),
-                    "--calibration", str(reports["calibration"]),
-                    "--test", str(reports["test"]),
-                    "--background", str(reports["background"]),
-                    "--output", str(assessment),
-                ],
+                protected_assessment_command(
+                    scripts_dir=scripts_dir,
+                    calibration=reports["calibration"],
+                    test=reports["test"],
+                    background=reports["background"],
+                    output=assessment,
+                    selected_threshold=float(chosen["threshold"]),
+                    min_aggregate_f1=args.min_aggregate_f1,
+                    min_region_f1=args.min_region_f1,
+                    min_region_precision=args.min_region_precision,
+                    min_region_recall=args.min_region_recall,
+                    max_pure_empty_fp=args.max_pure_empty_fp,
+                ),
                 iteration_dir / "assessment.log",
                 allowed={0, 2},
             )