Unblock objective Belgian training loop

backend/tests/test_belgium_training_loop.py

@@ -91,7 +91,34 @@ def test_loop_refuses_failed_dataset_audit(tmp_path: Path) -> None:
         check=False,
     )
     assert result.returncode != 0
-    assert "Dataset audit is not ok" in result.stderr
+    assert "Dataset audit is not eligible for training" in result.stderr
+
+
+def test_pending_human_review_does_not_block_objective_training() -> None:
+    audit = {
+        "status": "needs_human_review",
+        "failures": [],
+        "manifest_immutable": True,
+        "spatial_leakage_status": "ok",
+        "low_variance_positive_tile_count": 0,
+        "review_complete": False,
+    }
+    assert MODULE.dataset_audit_failures(audit) == []
+
+
+def test_training_audit_still_fails_closed_on_automated_integrity_gates() -> None:
+    audit = {
+        "status": "needs_human_review",
+        "failures": ["wallonia/test below minimum"],
+        "manifest_immutable": False,
+        "spatial_leakage_status": "failed",
+        "low_variance_positive_tile_count": 2,
+    }
+    failures = MODULE.dataset_audit_failures(audit)
+    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
 
 
 def test_calibration_failure_blocks_protected_evaluation() -> None:

docs/BELGIUM_BUILDING_TRAINING_LOOP.md

@@ -92,8 +92,11 @@ recall are repeated, while true negative train tiles are repeated when a
 regional precision gate or the pure-background gate fails. Calibration, test,
 background-test and validation AOIs are excluded by their frozen corpus split;
 the generated evidence records that no protected sample entered training.
-The orchestrator refuses to start unless the frozen dataset audit is `ok` and
+The orchestrator refuses to start unless every automated frozen-dataset gate
-contains zero blank/low-variance positive tiles.
+passes and the corpus contains zero blank/low-variance positive tiles. 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.
 For dated imagery, GRB `BEGINDATUM` and PICC `DATE_CREAT` are compared with the
 end of the imagery period. A feature created afterward is retained in the
 audit but excluded from training as `created_after_imagery_period`. UrbIS does

docs/CODEX_EXECUTION_LOG.md

@@ -11747,3 +11747,41 @@ Deployment evidence:
 - Repeated the live landing, guest bootstrap, map navigation and quality
   navigation against `http://192.168.10.150:1202`; the inspected live flow
   produced no browser-console errors.
+
+## 2026-07-29 - V36 assessment and objective-loop repair
+
+- Audited the current Tower evidence rather than relying on the earlier v6
+  checkpoint. The governed v30 rotated-holdout corpus contains 124 samples,
+  21,830 accepted labels, zero temporal-unknown samples, an immutable manifest
+  and no spatial leakage. Its split composition exceeds the frozen regional
+  minima and protected samples remain outside training.
+- Assessed the completed YOLO11x v36 checkpoint calibration-first on the RTX
+  4080. At the selected threshold `0.10`, aggregate F1 is `0.552`, Brussels F1
+  is `0.662`, Wallonia F1 is `0.532`, and Flanders F1 is `0.265`. The candidate
+  failed Flanders F1/precision/recall and Wallonia precision, so test and
+  background evidence were not opened and the production model was unchanged.
+- Built checksummed v37 failure-driven sampling from that rejection: 3,900
+  train entries across 94 train AOIs, with extra Flanders recall and
+  Flanders/Wallonia precision evidence. The sampling audit records zero
+  protected samples in training.
+- Started the inactive v37 YOLO11x fine-tune on Tower CUDA device 0 with
+  deterministic seed `20260806`, `max_det=1000`, AdamW and aerial rotation
+  augmentation. NVIDIA process evidence confirmed GPU execution.
+- Corrected a loop deadlock: an automatically clean corpus awaiting the final
+  human review has status `needs_human_review`, while the orchestrator formerly
+  required `ok`. Training now accepts that status only when the manifest is
+  immutable, automated failures are empty, spatial leakage is `ok`, and blank
+  positive-tile count is zero. Human sign-off remains a separate mandatory
+  final promotion gate.
+
+Verified in this pass:
+
+- `py -3 -m pytest -q backend/tests/test_belgium_training_loop.py backend/tests/test_belgium_training_iteration_assessment.py backend/tests/test_belgium_training_portfolio.py`
+  (`12 passed`).
+
+Open:
+
+- Let v37 finish, run calibration-only assessment, and expose protected test
+  and pure-background results only if every regional calibration gate passes.
+- Continue failure-driven, train-only corpus iterations until all objective
+  gates pass; only then request the queued representative human review.

docs/TODO.md

@@ -971,6 +971,9 @@ This file now starts with the current implementation status. Older preparation/b
 - [x] Replace rolling-mosaic training inputs with governed dated 2025 Flanders/Brussels and complete 2023 SPW imagery; retain exact flight-day limitations.
 - [x] Reject positive labels over blank/no-data imagery and replace partial SPW 2024 coverage with the complete dated SPW 2023 campaign.
 - [x] Exclude GRB/PICC features created after the corresponding dated imagery period while retaining auditable rejection evidence.
+- [x] Allow the objective CUDA loop to consume an automatically clean `needs_human_review` corpus while keeping final human sign-off as a separate, mandatory promotion gate.
+- [x] Evaluate the completed v36 YOLO11x checkpoint calibration-first on the rotated v30 holdouts; reject it before opening test/background because the regional calibration gate failed.
+- [ ] Finish and assess the leak-free v37 YOLO11x failure-driven CUDA iteration; open test/background evidence only if every calibration gate passes.
 
 # Sprint 229 - Visual release correction and Tower redeploy
 

scripts/run_belgium_building_training_loop.py

@@ -29,6 +29,21 @@ def write_json(path: Path, value: dict[str, Any]) -> None:
     temporary.replace(path)
 
 
+def dataset_audit_failures(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")
+    if status not in {"ok", "needs_human_review"}:
+        failures.append(f"unsupported audit status: {status}")
+    if audit.get("manifest_immutable") is not True:
+        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:
+        failures.append("dataset contains blank/low-variance positive tiles")
+    return failures
+
+
 def select_calibration_threshold(report: dict[str, Any]) -> dict[str, Any]:
     """Choose a threshold without consulting test or background evidence."""
     eligible = [item for item in report["sweeps"] if item["pure_empty_false_positives"] == 0]
@@ -161,10 +176,9 @@ 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"))
-    if dataset_audit.get("status") != "ok":
+    audit_failures = dataset_audit_failures(dataset_audit)
-        raise SystemExit(f"Dataset audit is not ok: {args.dataset_audit}")
+    if audit_failures:
-    if int(dataset_audit.get("low_variance_positive_tile_count") or 0) != 0:
+        raise SystemExit(f"Dataset audit is not eligible for training: {audit_failures}")
-        raise SystemExit("Dataset audit contains blank/low-variance positive tiles")
 
     state_path = args.output_dir / "training-loop-state.json"
     state: dict[str, Any] = {