Add independent Flemish low-rise calibration candidates

docs/CODEX_EXECUTION_LOG.md

@@ -3779,6 +3779,35 @@ Validation:
 
 - Add an auditable image/footprint alignment stage or select lower-parallax independent Flemish AOIs, then repeat visual QA before corpus composition.
 
+## 2026-07-30 - V66 independent Flemish low-rise replacement audit
+
+### What changed
+
+- Added and provisioned three spatially independent Flemish low-rise calibration candidates: Zutendaal, Zoersel and Landen.
+- Materialized immutable temporal corpus `building-be-v66-lowrise-temporal-r1` with manifest SHA-256 `d861da48aca40121e158e6ab243a6ccbe53e14d500a186af64b620f612935b07`.
+- Applied the provider-native GRB main-building allowlist, retaining 73, 66 and 110 source instances respectively while preserving every excluded annex in label-audit evidence.
+- Tested per-instance image-edge alignment, V31 detector-consensus translation, strict SAM refinement/drop and an independent YOLO-World roof vocabulary. Every generated contact sheet remained fail-closed rejected; none entered training.
+
+### What was tested
+
+- `python -m pytest tests/test_provision_belgium_building_training_portfolio.py -q` passed (`3 passed`).
+- The server confirmed CUDA execution on the RTX 4080 SUPER.
+- YOLO-World model SHA-256: `9b2c17ab6124a913e9b3a5c170617920d91b0f01111a8479da69f00e2cf27792`; it returned 65 low-confidence detections across 12 tiles and failed visual precision.
+- The V31 consensus route accepted only Zoersel with nine agreeing pairs and failed the other two AOIs; visual review rejected the accepted shift as well.
+
+### What remains open
+
+- An independently reviewed image-visible Flemish roof-label set is still required before the national retraining loop can honestly resume.
+- Protected test/background evaluation, promotion and production redeployment remain closed.
+
+### Known limitations
+
+- Orthophoto ground displacement means current GRB ground-footprint boxes are not detector-grade roof truth at IoU 0.5. Automatic proposal models tested here do not reach the required visual precision.
+
+### Next recommended pass
+
+- Obtain or annotate image-visible roof boxes independently of the release candidate, freeze their checksums and human acceptance evidence, then rebuild and retrain the national corpus.
+
 ## Post-V1 national coverage completion: Wallonia (2026-07-22)
 
 - Located and live-validated the stable official WALOUS 2018 GeoTIFF

docs/TODO.md

@@ -11,6 +11,8 @@ Uitvoeringsbord: `docs/PYTORCH_TRAINING_ROADMAP_BELGIUM.md`.
   - [x] Verwerp de drie V64 Vlaamse kalibratiekandidaten fail-closed na gereproduceerde ruwe en SAM2-contact-sheets; verscherp de geometrische SAM2-gates.
   - [x] Scheid GRB-hoofdgebouwen bron-native van bijgebouwen en bewijs dat ook de hoofdgebouw/SAM-dropvariant de visuele gate nog niet haalt.
   - [ ] Lever onafhankelijke Vlaamse kalibratie-AOI's met aantoonbaar beeldzichtbare daklabels vóór een volgende training.
+    - [x] Provision en audit onafhankelijke V66-laagbouwkandidaten in Zutendaal, Zoersel en Landen; verwerp alle automatische labelvarianten na visuele QA.
+    - [ ] Laat beeldzichtbare Vlaamse dakboxen onafhankelijk annoteren/reviewen en freeze hun checksums.
 - [ ] Train en evalueer een Belgische gebouwchallenger onafhankelijk; promoveer alleen zonder achtergrondregressie.
 - [ ] Houd zonnepanelen en segmentatie `not_configured` tot gereviewde taaklabels en hold-outs bestaan.
 

scripts/provision_belgium_building_training_portfolio.py

@@ -141,6 +141,12 @@ AOIS = (
     Aoi("aarschot-mixed-cal-candidate-v63", "flanders", "mixed-urban", "train", 4.837, 50.985),
     Aoi("beveren-ribbon-cal-candidate-v63", "flanders", "ribbon-development", "train", 4.256, 51.212),
     Aoi("oostkamp-suburban-cal-candidate-v63", "flanders", "suburban", "train", 3.226, 51.154),
+    # v66 replacement candidates after V63/V65 visual QA rejected dense,
+    # parallax-heavy labels. These spatially independent low-rise contexts
+    # intentionally reduce roof displacement and instance ambiguity.
+    Aoi("zutendaal-lowrise-cal-candidate-v66", "flanders", "detached-lowrise", "train", 5.573, 50.933),
+    Aoi("zoersel-lowrise-cal-candidate-v66", "flanders", "detached-lowrise", "train", 4.712, 51.268),
+    Aoi("landen-rural-cal-candidate-v66", "flanders", "rural-lowrise", "train", 5.079, 50.746),
     Aoi("bruges-val", "flanders", "historic-urban", "val", 3.224, 51.209),
     Aoi("turnhout-val", "flanders", "suburban", "val", 4.944, 51.322),
     Aoi("hasselt-cal", "flanders", "suburban", "calibration", 5.340, 50.930),

tests/test_provision_belgium_building_training_portfolio.py

@@ -6,6 +6,11 @@ V63_CANDIDATES = {
     "beveren-ribbon-cal-candidate-v63",
     "oostkamp-suburban-cal-candidate-v63",
 }
+V66_CANDIDATES = {
+    "zutendaal-lowrise-cal-candidate-v66",
+    "zoersel-lowrise-cal-candidate-v66",
+    "landen-rural-cal-candidate-v66",
+}
 
 
 def test_portfolio_slugs_are_unique() -> None:
@@ -19,3 +24,11 @@ def test_v63_calibration_candidates_are_flemish_train_role_inputs() -> None:
     assert all(aoi.region == "flanders" for aoi in candidates.values())
     assert all(aoi.split == "train" for aoi in candidates.values())
     assert all(aoi.sample_role == "positive" for aoi in candidates.values())
+
+
+def test_v66_lowrise_candidates_are_flemish_train_role_inputs() -> None:
+    candidates = {aoi.slug: aoi for aoi in AOIS if aoi.slug in V66_CANDIDATES}
+    assert set(candidates) == V66_CANDIDATES
+    assert all(aoi.region == "flanders" for aoi in candidates.values())
+    assert all(aoi.split == "train" for aoi in candidates.values())
+    assert all("lowrise" in aoi.context for aoi in candidates.values())