Require tile quality evidence for training

backend/tests/test_belgium_training_loop.py

@@ -98,6 +98,11 @@ def test_dry_run_can_gate_existing_checkpoint_without_training(tmp_path: Path) -
         "status": "needs_human_review", "failures": [],
         "manifest_immutable": True, "spatial_leakage_status": "ok",
     }))
+    quality = tmp_path / "quality.json"
+    quality.write_text(json.dumps({
+        "status": "ok", "low_variance_positive_tile_count": 0,
+        "label_stats": {"invalid_label_count": 0, "missing_label_file_count": 0},
+    }))
     result = subprocess.run(
         [
             sys.executable, str(SCRIPT),
@@ -105,6 +110,7 @@ def test_dry_run_can_gate_existing_checkpoint_without_training(tmp_path: Path) -
             "--train-yaml", str(tmp_path / "dataset.yaml"),
             "--train-summary", str(tmp_path / "train-summary.json"),
             "--dataset-audit", str(audit),
+            "--train-quality-audit", str(quality),
             "--calibration-summary", str(tmp_path / "cal.json"),
             "--test-summary", str(tmp_path / "test.json"),
             "--background-summary", str(tmp_path / "background.json"),
@@ -128,6 +134,11 @@ def test_existing_checkpoint_iteration_directory_can_be_created_without_yolo(tmp
 def test_loop_refuses_failed_dataset_audit(tmp_path: Path) -> None:
     audit = tmp_path / "audit.json"
     audit.write_text(json.dumps({"status": "needs_attention", "low_variance_positive_tile_count": 4}))
+    quality = tmp_path / "quality.json"
+    quality.write_text(json.dumps({
+        "status": "ok", "low_variance_positive_tile_count": 0,
+        "label_stats": {"invalid_label_count": 0, "missing_label_file_count": 0},
+    }))
     result = subprocess.run(
         [
             sys.executable,
@@ -140,6 +151,8 @@ def test_loop_refuses_failed_dataset_audit(tmp_path: Path) -> None:
             str(tmp_path / "train-summary.json"),
             "--dataset-audit",
             str(audit),
+            "--train-quality-audit",
+            str(quality),
             "--calibration-summary",
             str(tmp_path / "cal.json"),
             "--test-summary",
@@ -168,7 +181,11 @@ def test_pending_human_review_does_not_block_objective_training() -> None:
         "low_variance_positive_tile_count": 0,
         "review_complete": False,
     }
-    assert MODULE.dataset_audit_failures(audit) == []
+    quality = {
+        "status": "ok", "low_variance_positive_tile_count": 0,
+        "label_stats": {"invalid_label_count": 0, "missing_label_file_count": 0},
+    }
+    assert MODULE.dataset_audit_failures(audit, quality) == []
 
 
 def test_training_audit_still_fails_closed_on_automated_integrity_gates() -> None:
@@ -179,11 +196,30 @@ def test_training_audit_still_fails_closed_on_automated_integrity_gates() -> Non
         "spatial_leakage_status": "failed",
         "low_variance_positive_tile_count": 2,
     }
-    failures = MODULE.dataset_audit_failures(audit)
+    quality = {
+        "status": "failed", "low_variance_positive_tile_count": 2,
+        "label_stats": {"invalid_label_count": 1, "missing_label_file_count": 1},
+    }
+    failures = MODULE.dataset_audit_failures(audit, quality)
     assert "wallonia/test below minimum" in failures
     assert "corpus manifest is not immutable" in failures
     assert "spatial leakage audit is not ok" in failures
     assert "dataset contains blank/low-variance positive tiles" in failures
+    assert "train tile quality audit is not ok" in failures
+    assert "train tile quality audit contains invalid labels" in failures
+    assert "train tile quality audit contains missing label files" in failures
+
+
+def test_missing_tile_quality_evidence_fails_closed() -> None:
+    audit = {
+        "status": "needs_human_review", "failures": [],
+        "manifest_immutable": True, "spatial_leakage_status": "ok",
+    }
+    failures = MODULE.dataset_audit_failures(audit, {})
+    assert "train tile quality audit is not ok" in failures
+    assert "train tile quality audit contains invalid labels" in failures
+    assert "train tile quality audit contains missing label files" in failures
+    assert "dataset contains blank/low-variance positive tiles" in failures
 
 
 def test_calibration_failure_blocks_protected_evaluation() -> None:

docs/BELGIUM_BUILDING_TRAINING_LOOP.md

@@ -131,6 +131,9 @@ the versioned JSON argv command. Invalid state and bounded launch exhaustion
 fail closed.
 The orchestrator refuses to start unless every automated frozen-dataset gate
 passes and the corpus contains zero blank/low-variance positive tiles. The
+separate train tile-quality audit is a required checksummed loop input; missing
+invalid-label, missing-label-file or low-variance evidence fails closed rather
+than being interpreted as zero. The
 audit status may remain `needs_human_review` while training and objective
 evaluation continue: final human sign-off is deliberately the last gate and
 can never be interpreted as model promotion approval in advance.

docs/CODEX_EXECUTION_LOG.md

@@ -11,6 +11,10 @@
   and nine Walloon dense/rural/architecture or difficult-negative samples.
   Added mode-0600 session-token-file support to the official provisioner so
   authenticated Tower acquisition does not expose session material in argv.
+- Closed a fail-open evidence gap between corpus and tile audits. The training
+  loop now requires and hashes the train tile-quality report, and refuses
+  missing counters instead of treating absent invalid/blank-label evidence as
+  zero.
 
 ## 2026-07-27 - Guest demo and product professionalization
 

scripts/run_belgium_building_training_loop.py

@@ -29,7 +29,10 @@ def write_json(path: Path, value: dict[str, Any]) -> None:
     temporary.replace(path)
 
 
-def dataset_audit_failures(audit: dict[str, Any]) -> list[str]:
+def dataset_audit_failures(
+    audit: dict[str, Any],
+    train_quality_audit: dict[str, Any],
+) -> list[str]:
     """Return automated corpus blockers while leaving final human review deferred."""
     failures = [str(item) for item in audit.get("failures") or []]
     status = audit.get("status")
@@ -39,7 +42,13 @@ def dataset_audit_failures(audit: dict[str, Any]) -> list[str]:
         failures.append("corpus manifest is not immutable")
     if audit.get("spatial_leakage_status") != "ok":
         failures.append("spatial leakage audit is not ok")
-    if int(audit.get("low_variance_positive_tile_count") or 0) != 0:
+    if train_quality_audit.get("status") != "ok":
+        failures.append("train tile quality audit is not ok")
+    if int(train_quality_audit.get("label_stats", {}).get("invalid_label_count", -1)) != 0:
+        failures.append("train tile quality audit contains invalid labels")
+    if int(train_quality_audit.get("label_stats", {}).get("missing_label_file_count", -1)) != 0:
+        failures.append("train tile quality audit contains missing label files")
+    if int(train_quality_audit.get("low_variance_positive_tile_count", -1)) != 0:
         failures.append("dataset contains blank/low-variance positive tiles")
     return failures
 
@@ -188,6 +197,7 @@ def main() -> int:
     parser.add_argument("--train-yaml", type=Path, required=True)
     parser.add_argument("--train-summary", type=Path, required=True)
     parser.add_argument("--dataset-audit", type=Path, required=True)
+    parser.add_argument("--train-quality-audit", type=Path, required=True)
     parser.add_argument("--calibration-summary", type=Path, required=True)
     parser.add_argument("--test-summary", type=Path, required=True)
     parser.add_argument("--background-summary", type=Path, required=True)
@@ -230,7 +240,8 @@ def main() -> int:
     if args.iterations < 1:
         raise SystemExit("--iterations must be positive")
     dataset_audit = json.loads(args.dataset_audit.read_text(encoding="utf-8"))
-    audit_failures = dataset_audit_failures(dataset_audit)
+    train_quality_audit = json.loads(args.train_quality_audit.read_text(encoding="utf-8"))
+    audit_failures = dataset_audit_failures(dataset_audit, train_quality_audit)
     if audit_failures:
         raise SystemExit(f"Dataset audit is not eligible for training: {audit_failures}")
 
@@ -242,6 +253,9 @@ def main() -> int:
         "initial_model": str(args.initial_model),
         "train_yaml": str(args.train_yaml),
         "dataset_audit": str(args.dataset_audit),
+        "dataset_audit_sha256": sha256(args.dataset_audit),
+        "train_quality_audit": str(args.train_quality_audit),
+        "train_quality_audit_sha256": sha256(args.train_quality_audit),
         "corpus_manifest": str(args.corpus_manifest),
         "iterations": [],
     }

scripts/tower-v37-completion-command.json

@@ -13,6 +13,8 @@
   "/app/storage/operator-data/building-be-v30-rotated-holdouts-r1/train/yolo_tile_dataset_summary.json",
   "--dataset-audit",
   "/app/storage/operator-data/building-be-v30-rotated-holdouts-r1/audit/belgium-building-corpus-audit.json",
+  "--train-quality-audit",
+  "/app/storage/operator-data/building-be-v30-rotated-holdouts-r1/train/audit/operator_yolo_dataset_quality_audit.json",
   "--calibration-summary",
   "/app/storage/operator-data/building-be-v30-rotated-holdouts-r1/calibration/yolo_tile_dataset_summary.json",
   "--test-summary",