docs(accuracy): refresh governed scan and training evidence

backend/tests/test_belgium_training_loop.py

@@ -175,6 +175,31 @@ def test_failed_iteration_builds_train_only_sampling_for_next_checkpoint(tmp_pat
     assert command[command.index("--output-dir") + 1].endswith("failure-driven-training")
 
 
+def test_protected_assessment_uses_frozen_threshold_and_configured_gates(
+    tmp_path: Path,
+) -> None:
+    command = MODULE.protected_assessment_command(
+        scripts_dir=tmp_path / "scripts",
+        calibration=tmp_path / "calibration.json",
+        test=tmp_path / "test.json",
+        background=tmp_path / "background.json",
+        output=tmp_path / "assessment.json",
+        selected_threshold=0.275,
+        min_aggregate_f1=0.71,
+        min_region_f1=0.62,
+        min_region_precision=0.73,
+        min_region_recall=0.58,
+        max_pure_empty_fp=1,
+    )
+
+    assert command[command.index("--selected-threshold") + 1] == "0.275"
+    assert command[command.index("--min-aggregate-f1") + 1] == "0.71"
+    assert command[command.index("--min-region-f1") + 1] == "0.62"
+    assert command[command.index("--min-region-precision") + 1] == "0.73"
+    assert command[command.index("--min-region-recall") + 1] == "0.58"
+    assert command[command.index("--max-pure-empty-fp") + 1] == "1"
+
+
 def test_partial_iteration_resume_uses_exact_checkpoint_and_cuda(tmp_path: Path) -> None:
     checkpoint = tmp_path / "runs" / "iteration-002" / "weights" / "last.pt"
     assert MODULE.resumable_training_command("yolo", checkpoint) == [