Complete Belgian building corpus v2 workflow

docs/CODEX_EXECUTION_LOG.md

@@ -11498,3 +11498,36 @@ Next gate:
   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.
+
+## 2026-07-27 - Belgian building corpus v2 and independent CUDA evaluation
+
+- Added a checkpoint-safe provisioner for 42 geographically independent AOIs:
+  14 per region, with six training, two validation, two calibration, two test
+  and two background-test AOIs. The portfolio includes dense urban, suburban,
+  rural, industrial, forest, heath, quarry, rail, park and port contexts.
+- Acquired imagery at an explicit 25 cm resolution and made the API reject a
+  requested resolution finer than the governed source's native resolution.
+  Rolling imagery whose per-pixel observation date is unavailable is now
+  recorded as `unknown_per_pixel`; it is never declared temporally aligned to
+  PICC/UrbIS/GRB merely from the download date.
+- Frozen corpus `building-be-v2-20260727-r3` contains 42 samples and 6,761
+  accepted labels from 6,828 inputs. Sixty-seven sub-pixel labels were rejected
+  explicitly. Manifest SHA-256 is
+  `8a2ccd39642be30a58bd12b52b117ea4e2055437b2ed9d49e4022b47605e5f19`.
+  Spatial leakage passed and all 42 temporal relations remain honestly unknown.
+- Exported a 96-tile training/validation set (5,711 labels; 82 positive and 14
+  negative tiles) plus independent 24-tile calibration, test and background
+  sets. Complete, calibration, test and background contact sheets were rendered.
+- Trained `building-be-v2-active-ft-e50.pt` for 50 epochs with CUDA on the Tower
+  NVIDIA GeForce RTX 4080 SUPER. Artifact SHA-256 is
+  `615769c585ff96af4f4be3b7bdeb26e6fb7d59ba6d8f25f90fefe85fe366fb2e`.
+- On the independent test set the candidate achieved precision `0.353`, recall
+  `0.228`, mAP50 `0.164` and mAP50-95 `0.0553`, versus incumbent `0.118`,
+  `0.127`, `0.0385` and `0.0118`. On the background/hard-negative set it
+  achieved `0.540`, `0.392`, `0.375` and `0.172`, versus incumbent `0.193`,
+  `0.129`, `0.0815` and `0.0348`. At confidence 0.25 it emitted zero detections
+  on all 15 pure-background tiles.
+- The challenger is materially better but remains below a credible national
+  production gate, so it was not promoted. The active model remains unchanged.
+  The deterministic audit status is `needs_human_review`: an AI-assisted visual
+  inspection cannot be represented as the required human approval.

docs/TODO.md

@@ -946,7 +946,7 @@ This file now starts with the current implementation status. Older preparation/b
 - [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.
+- [ ] Complete representative human label/contact-sheet review (the complete 42-AOI review queue and four contact sheets are ready; AI-assisted inspection is recorded separately and does not count as human sign-off).
+- [x] Expand each region/context/split until the national minimum-composition gate passes (42 independent 256 m AOIs at 25 cm, including pure-background and hard-negative contexts).
+- [x] Run explicit spatial leakage and independent calibration/test evaluation.
 - [ ] Promote only if every regional and pure-background gate passes.