Audit Belgian corpus splits and record CUDA evidence

backend/tests/test_building_label_normalization.py

@@ -15,6 +15,13 @@ assert SPEC and SPEC.loader
 module = importlib.util.module_from_spec(SPEC)
 SPEC.loader.exec_module(module)
 
+ASSEMBLER_SPEC = importlib.util.spec_from_file_location(
+    "assemble_building_corpus", ROOT / "scripts" / "assemble_belgium_building_corpus.py"
+)
+assert ASSEMBLER_SPEC and ASSEMBLER_SPEC.loader
+assembler = importlib.util.module_from_spec(ASSEMBLER_SPEC)
+ASSEMBLER_SPEC.loader.exec_module(assembler)
+
 
 def test_normalizer_retains_native_identity_and_records_rejections(tmp_path: Path) -> None:
     raster_path = tmp_path / "image.tif"
@@ -69,3 +76,15 @@ def test_normalizer_retains_native_identity_and_records_rejections(tmp_path: Pat
         "excluded_canopy": 1,
     }
     assert audit["temporal_mismatch_days"] == 31
+
+
+def test_spatial_leakage_audit_fails_cross_split_neighbors() -> None:
+    samples = [
+        {"sample_slug": "train-a", "split": "train", "bbox_epsg4326": [4.0, 50.0, 4.01, 50.01]},
+        {"sample_slug": "val-a", "split": "val", "bbox_epsg4326": [4.005, 50.005, 4.02, 50.02]},
+        {"sample_slug": "test-far", "split": "test", "bbox_epsg4326": [5.0, 51.0, 5.01, 51.01]},
+    ]
+    audit = assembler.audit_spatial_leakage(samples)
+    assert audit["status"] == "failed"
+    assert audit["findings"][0]["left"] == "train-a"
+    assert audit["findings"][0]["right"] == "val-a"

docs/CODEX_EXECUTION_LOG.md

@@ -11467,3 +11467,34 @@ Next gate:
 - Verification: frontend build and 78 focused contracts passed. The full
   backend run reported 1124 passed plus five unrelated local Windows/WSL
   `bash` path failures.
+## 2026-07-26 - Belgian building corpus candidate and CUDA matrix
+
+- Added provider-aware canonical label normalization for GRB, PICC and UrbIS.
+  Every source feature retains its native ID/class and receives an explicit
+  accept/reject reason covering geometry repair, duplicates, semantic
+  exclusions and resolvable pixel size.
+- Added an immutable corpus assembler that resolves only persisted governed
+  Dataset IDs, validates regional provider pairing, copies checksum-bound
+  artifacts and refuses a non-empty output directory.
+- Frozen experimental corpus `building-be-v1-candidate-20260726` contains 19
+  geographically separated AOIs spanning all three land regions and explicit
+  train, validation, calibration, test and background-test roles. Manifest
+  SHA-256 is `0450ce782c35c5955e519fae489ffdbef6075d871adb0d196cc6a244571788fa`.
+- The training export contains 36 tiles and 7,219 tile-level labels; its
+  automated label-size, variance and split audit passed. A 12-tile regional
+  pilot contact sheet was rendered for human review.
+- CUDA training on the RTX 4080 SUPER completed a generic YOLOv8s candidate
+  (`mAP50=0.0883`, `mAP50-95=0.0248`) and an incumbent fine-tune
+  (`mAP50=0.187`, `mAP50-95=0.0617`) on the held-out regional validation
+  samples. Both are immutable candidates and neither is promoted.
+- A 60-epoch incumbent fine-tune completed with artifact SHA-256
+  `594f9fef356940e7f7839da36561387a9a2f436c1e498e5733f99e1170c22fa6`.
+  On the unopened regional test AOIs (Leuven, Mons and Brussels rail) it
+  achieved precision `0.315`, recall `0.229`, mAP50 `0.120` and mAP50-95
+  `0.0340`. The incumbent scored `0.227`, `0.233`, `0.0930` and `0.0262`
+  respectively. The challenger improves precision/AP but slightly lowers
+  recall and remains far below a credible national acceptance floor.
+- The scores prove that the current small candidate corpus is insufficient for
+  a national production claim. Human review, broader negative coverage,
+  leakage audit and independent calibration/test evaluation remain blocking
+  gates; the active production asset was left unchanged.

docs/PYTORCH_TRAINING_ROADMAP_BELGIUM.md

@@ -121,21 +121,21 @@ manual URL editing and without crossing provider coverage zones.
 
 ### WP2 — Reference-label normalization
 
-- [ ] normalize GRB, PICC and UrbIS building semantics to canonical `building`;
-- [ ] retain original feature IDs and source classes;
-- [ ] quantify invalid geometry repairs, duplicates and temporal mismatches;
-- [ ] define exclusion rules for ruins, canopies, underground structures and
+- [x] normalize GRB, PICC and UrbIS building semantics to canonical `building`;
+- [x] retain original feature IDs and source classes;
+- [x] quantify invalid geometry repairs, duplicates and temporal mismatches;
+- [x] define exclusion rules for ruins, canopies, underground structures and
   objects below the resolvable pixel size;
-- [ ] generate overlay/contact-sheet review packs.
+- [x] generate overlay/contact-sheet review packs.
 
 Exit: every accepted/rejected label has a machine-readable reason and provenance.
 
 ### WP3 — Frozen Belgium building corpus
 
-- [ ] select and persist the geographic strata and split assignments;
-- [ ] acquire imagery and references through governed services;
-- [ ] export tiles at candidate resolutions with deterministic negative sampling;
-- [ ] run leakage, checksum, class, label-size and visual-variance audits;
+- [x] select and persist the initial geographic strata and split assignments;
+- [x] acquire the candidate imagery and references through governed services;
+- [x] export tiles at candidate resolutions with deterministic negative sampling;
+- [x] run checksum, class, label-size and visual-variance audits;
 - [ ] complete representative human label review;
 - [ ] freeze dataset version `building-be-v1` and make it immutable.
 

docs/TODO.md

@@ -940,3 +940,13 @@ This file now starts with the current implementation status. Older preparation/b
 - [ ] Resolve governed North Sea bathymetry access without bypassing TLS.
 - [x] Enforce the active building-only model classes, CUDA readiness and persisted Mol/Kempen Area scope.
 - [ ] Add national/regional holdout evidence before expanding model classes or validated Areas.
+## Belgian PyTorch training programme (2026-07-26)
+
+- [x] Add governed SPW and UrbIS orthophoto acquisition.
+- [x] Normalize GRB/PICC/UrbIS building labels with decision provenance.
+- [x] Assemble and checksum an initial 19-AOI Belgian candidate corpus.
+- [x] Run generic and incumbent-based CUDA candidate training without promotion.
+- [ ] Complete representative human label/contact-sheet review.
+- [ ] Expand each region/context/split until the national minimum-composition gate passes.
+- [ ] Run explicit spatial leakage and independent calibration/test evaluation.
+- [ ] Promote only if every regional and pure-background gate passes.

scripts/assemble_belgium_building_corpus.py

@@ -7,14 +7,25 @@ import argparse
 import hashlib
 import json
 import shutil
+import sys
 from pathlib import Path
 from typing import Any
 from uuid import UUID
 
-from app.db.session import SessionLocal
-from app.models import Dataset
+from pyproj import Transformer
+from shapely.geometry import box
+from shapely.ops import transform as shapely_transform
 
-from normalize_belgium_building_labels import normalize
+REPO_ROOT = Path(__file__).resolve().parents[1]
+APP_ROOT = REPO_ROOT if (REPO_ROOT / "app").is_dir() else REPO_ROOT / "backend"
+for import_root in (Path(__file__).resolve().parent, APP_ROOT):
+    if str(import_root) not in sys.path:
+        sys.path.insert(0, str(import_root))
+
+from app.db.session import SessionLocal  # noqa: E402
+from app.models import Dataset  # noqa: E402
+
+from normalize_belgium_building_labels import normalize  # noqa: E402
 
 REGION_SOURCES = {
     "flanders": ("digitaal_vlaanderen_orthophoto", "grb"),
@@ -58,6 +69,33 @@ def _validate_pair(sample: dict[str, Any], raster: Dataset, reference: Dataset)
     return region, expected_reference
 
 
+def audit_spatial_leakage(samples: list[dict[str, Any]], buffer_m: float = 64.0) -> dict[str, Any]:
+    transformer = Transformer.from_crs("EPSG:4326", "EPSG:31370", always_xy=True)
+    findings: list[dict[str, Any]] = []
+    metric_boxes: list[tuple[dict[str, Any], Any]] = []
+    for sample in samples:
+        bounds = sample.get("bbox_epsg4326")
+        if not isinstance(bounds, list) or len(bounds) != 4:
+            raise SystemExit(f"Missing governed bbox for leakage audit: {sample['sample_slug']}")
+        metric_boxes.append((sample, shapely_transform(transformer.transform, box(*map(float, bounds)))))
+    for index, (left, left_geometry) in enumerate(metric_boxes):
+        for right, right_geometry in metric_boxes[index + 1 :]:
+            if left["split"] == right["split"]:
+                continue
+            distance_m = left_geometry.distance(right_geometry)
+            if distance_m < buffer_m:
+                findings.append(
+                    {
+                        "left": left["sample_slug"],
+                        "left_split": left["split"],
+                        "right": right["sample_slug"],
+                        "right_split": right["split"],
+                        "distance_m": distance_m,
+                    }
+                )
+    return {"status": "ok" if not findings else "failed", "buffer_m": buffer_m, "findings": findings}
+
+
 def main() -> int:
     parser = argparse.ArgumentParser()
     parser.add_argument("--spec", type=Path, required=True)
@@ -137,6 +175,12 @@ def main() -> int:
     }
     manifest_path = output_dir / "operator_samples_manifest.json"
     manifest_path.write_text(json.dumps(manifest, ensure_ascii=False, indent=2), encoding="utf-8")
+    leakage_audit = audit_spatial_leakage(manifest_samples)
+    (output_dir / "spatial-leakage-audit.json").write_text(
+        json.dumps(leakage_audit, ensure_ascii=False, indent=2), encoding="utf-8"
+    )
+    if leakage_audit["status"] != "ok":
+        raise SystemExit("Spatial split leakage audit failed")
     freeze = {
         "dataset_version": args.version,
         "manifest_sha256": sha256(manifest_path),