feat: close measured model review evidence

docs/reviews/2026-07-15-small-building-model-review.md

@@ -0,0 +1,71 @@
+# Small-building model review - 2026-07-15
+
+## Scope
+
+This review uses persisted configured-YOLO detections, persisted GRB building
+features and the exact orthophoto inference tiles from Geel, Herentals and
+Turnhout. It does not infer labels from QA status alone and does not alter the
+active model.
+
+- Model asset: `geointel-building-yolov8s-smallbld-minpx3-img640-ft30-pt`
+- SHA256: `a9088b8491dfae36694b53e9e9406cb4e3511d334a5712fa34f75078a47759c1`
+- Canonical QA method: candidate polygon versus GRB footprint IoU `0.25`
+- False-positive cards reviewed: 48
+- False-negative cards reviewed: 48
+- False negatives outside persisted inference-tile coverage: 731, excluded
+
+Every card was checked against its orthophoto and the persisted candidate and
+reference overlays. Geometric overlap diagnostics were used to distinguish a
+model error from box/footprint or one-to-one matching effects. Ambiguous cards
+remain excluded from training.
+
+## Decisions
+
+| Evidence | Confirmed model error | QA alignment | Reference gap/change | Uncertain/obscured |
+| --- | ---: | ---: | ---: | ---: |
+| False positive | 5 | 34 | 5 | 4 |
+| False negative | 10 | 25 | 5 | 8 |
+| Total | 15 | 59 | 10 | 12 |
+
+The dominant finding is not a model error. In 59 of 96 reviewed cards, a real
+candidate and reference overlap but the canonical box-versus-footprint or
+one-to-one assignment does not count that pair as a match. Those records must
+not become positive or negative training labels.
+
+## Training-readiness audit
+
+The active tile corpus already uses Geel and Herentals as training sources and
+keeps Turnhout excluded as an operation-level holdout.
+
+| Confirmed evidence | Geel | Herentals | Turnhout holdout |
+| --- | ---: | ---: | ---: |
+| False positive | 3 | 1 | 1 |
+| False negative | 0 | 4 | 6 |
+
+- The four confirmed false negatives in Herentals already exist as GRB labels
+  in the current training source. Re-adding them would not add new ground
+  truth; it would only change sample weighting.
+- Confirmed false-positive detections in Geel and Herentals occur on urban
+  source tiles that also contain valid GRB buildings. Treating those complete
+  tiles as empty hard negatives would create false negative labels.
+- The seven confirmed errors in Turnhout remain holdout evidence and cannot be
+  used for training without invalidating the independent benchmark.
+
+Result: **0 novel, leakage-free training labels are available from this review
+bundle.** A new fine-tuning run is therefore rejected. The active model and
+confidence `0.15` remain unchanged.
+
+## Required next evidence before training
+
+1. Collect new training-only orthophoto AOIs outside all Mol/Turnhout/Retie/
+   Westerlo operational holdouts.
+2. Label the complete contents of each selected tile from an authoritative
+   reference snapshot; never label only the reviewed detection box.
+3. Add enough independent confirmed small-building and true empty-background
+   examples to justify a separate candidate.
+4. Keep the candidate inactive until it passes the same positive-AOI,
+   pure-empty-background and Mol holdout gates.
+
+The review CSVs, contact sheets and validator outputs remain under the
+persistent operator-data mount and are intentionally not committed as generated
+repository artifacts.