evaluate models on fresh regional calibration AOIs

docs/CODEX_EXECUTION_LOG.md

@@ -12771,3 +12771,29 @@ Open:
   No independent positive or background accuracy evidence currently exists for
   this model family; new AOIs must be provisioned without ancestry or spatial
   exposure before another ranking is defensible.
+
+## 2026-08-09 - Fresh V72 regional calibration
+
+### Delivered
+
+- Added custom non-protected AOI provisioning with explicit rejection of test
+  and background-test roles.
+- Acquired six new governed orthophoto/reference pairs across all three Belgian
+  regions and froze an evaluation-only, training-prohibited corpus.
+- Added a diagnostic tiler that refuses training-eligible manifests, emits no
+  training data and excludes generic overlapping edge-cover windows.
+- Added a raster-bound spatial lineage audit. All 31 ancestral AOIs resolved;
+  every V72 AOI passes the 2 km separation floor.
+- Added strict pure-background validation: non-empty labels can no longer be
+  silently counted as pure background.
+
+### GPU result and decision
+
+- Corrected view: 24 non-overlapping tiles, 348 labels, six AOIs.
+- Active mAP50/mAP50-95: `0.056779`/`0.021750`.
+- Challenger mAP50/mAP50-95: `0.044705`/`0.016004`.
+- Confidence 0.25 F1: active `0.231527`, challenger `0.235885`.
+- The active aggregate F1 peaks at 0.30, but its Flemish subgroup regresses;
+  threshold, active weights and production image therefore remain unchanged.
+- The weak fresh-domain scores are retained as a real blocker, not hidden by
+  the older exposed Turnhout/Westerlo/Postel diagnostics.

docs/TODO.md

@@ -1185,3 +1185,14 @@ This file now starts with the current implementation status. Older preparation/b
 - [ ] Convert the AI-assisted ledger into no stronger claim than experimental
   triage; a real human must independently review and sign the frozen artifacts
   before the governed training wrapper may unlock.
+- [x] Provision and visually inspect six fresh non-protected calibration AOIs
+  across Vlaanderen, Wallonië and Brussel; prove all are at least 2 km from all
+  31 resolved ancestral AOIs and absent from every ancestral split.
+- [x] Reject overlapping edge-cover tiles in diagnostic evaluation; retain 24
+  canonical 512 px cells and record all 30 excluded overlapping views.
+- [x] Run active/challenger GPU calibration on the fresh portfolio and retain
+  aggregate, regional and per-AOI metrics without promoting either model.
+- [ ] Provision a separately frozen train-only failure corpus after human
+  review. Do not copy V72 calibration tiles or use their labels for fitting.
+- [ ] Finalise version-specific PICC/UrbIS semantic harmonisation contracts
+  before treating Walloon/Brussels metrics as release-grade ground truth.

docs/reviews/2026-08-09-ai-assisted-checkpoint-review.md

@@ -263,3 +263,41 @@ The byte-matching parent/output chain is retained separately in
 `artifacts/evidence/accuracy/model-training/20260809-v71-model-lineage-receipt.json`.
 Fresh geographically separated AOIs with complete spatial and lineage checks
 are required for the next meaningful evaluation.
+
+## Fresh V72 calibration portfolio
+
+Six AOIs were pre-registered before inference and acquired again from the
+governed regional orthophoto and building adapters: two each in Vlaanderen,
+Wallonië and Brussel. All six are absent from every split of the retained model
+lineage. A raster-bound spatial audit resolved all 31 ancestral AOIs and found
+the nearest new AOI at 4,869.63 metres, above the frozen 2 km floor.
+
+The initial generic edge-cover tiling would have counted 54 tiles because its
+final-row/final-column coverage nearly duplicated the preceding 512 px cells.
+That view was rejected. The corrected evaluation-only exporter retains exactly
+24 full grid cells, excludes 30 overlapping edge-cover views with reason codes,
+and writes no training tiles. Both the corpus and tile view contain bound
+`NO_TRAINING.json` markers.
+
+All 24 tiles and 348 retained labels were visually inspected. Havelange is a
+positive village context despite its pre-registered `field-cal-bg` name.
+Assenede and Laeken contain sparse official building labels and are difficult
+contexts, not pure-empty backgrounds. The evaluator now refuses to accept a
+pure-background prefix when any corresponding label file is non-empty.
+
+On the RTX 4080 SUPER the active model scores mAP50 `0.056779` and mAP50-95
+`0.021750`; the challenger scores `0.044705` and `0.016004`. At confidence 0.25
+and match IoU 0.25 their F1 scores are `0.231527` and `0.235885`. These low
+values expose substantial domain-generalisation failure. The active model stays
+in production because the challenger has weaker mAP, only a marginal F1 gain,
+missing historical release provenance and no valid promotion bundle.
+
+The active model's aggregate calibration F1 peaks at threshold 0.30
+(`0.250356`), but its Flemish subgroup becomes worse than at lower thresholds.
+No threshold was changed because a global average may not mask that subgroup
+regression. PICC and UrbIS remain diagnostic-only while their regional semantic
+harmonisation contracts are pending; no national release claim is made.
+
+Complete evidence is retained under
+`artifacts/evidence/accuracy/model-training/20260809-v72-fresh-calibration/`
+and summarized by `20260809-v72-fresh-calibration-summary.json`.