feat: close measured model review evidence

docs/AI_PIPELINES.md

@@ -172,6 +172,14 @@ not treated as reviewable model misses. This renderer does not alter persisted
 QA metrics; coverage-adjusted values remain audit diagnostics until the QA
 service evaluation population is deliberately hardened.
 
+`validate_detection_false_negative_review_decisions.py` provides the same
+fail-closed validation as the false-positive workflow. It requires an exact
+one-to-one set of reviewed reference ids and emits only explicit
+`confirmed_model_false_negative` geometries. `--require-complete` rejects any
+remaining `unreviewed` row. The July 2026 96-card review is recorded in
+`docs/reviews/2026-07-15-small-building-model-review.md`; it yielded no novel,
+leakage-free labels and therefore did not trigger model training.
+
 ### Local model asset catalog
 
 GeoIntel can list local runtime model files mounted into the backend model

docs/CODEX_EXECUTION_LOG.md

@@ -8192,3 +8192,40 @@ Next:
 - Deploy the migration and UI, verify one live review queue and map QA result,
   then complete representative manual decisions before constructing any new
   model-training corpus.
+
+## Sprint 198 - Evidence-closed model review (2026-07-15)
+
+Implemented:
+- Generated the previously missing 48-card false-negative contact-sheet bundle
+  from the exact persisted inference manifests for Geel, Herentals and
+  Turnhout; 731 references outside tile coverage remained explicitly excluded.
+- Completed all 96 FP/FN decisions through orthophoto inspection plus
+  persisted geometry-overlap diagnostics. The result is 5 confirmed model FP,
+  10 confirmed model FN, 59 QA-alignment cases, 10 reference-gap/change cases
+  and 12 uncertain/obscured cases.
+- Added `validate_detection_false_negative_review_decisions.py`, mirroring the
+  existing FP safety contract and exporting only explicit confirmed misses.
+- Audited confirmed evidence against the active training split. Geel and
+  Herentals are existing training sources; Turnhout is an excluded operational
+  holdout. The review therefore provides zero novel leakage-free labels and a
+  new fine-tuning run was deliberately rejected.
+- Added strict match count plus the existing diagnostic reference-envelope
+  result to the map analysis panel. Canonical footprint IoU metrics remain
+  unchanged.
+
+Validation:
+- Both 48-row decision CSVs completed with zero `unreviewed` records and passed
+  their validators with `--require-complete`.
+- Focused validator/readiness/map-contract tests passed.
+- Full readiness passed 596 backend tests, backend compilation, 88 documented
+  API routes, one Alembic head, frontend typecheck/build and shell syntax gates.
+
+Known limitation:
+- The active local model remains useful but imperfect. A new candidate requires
+  independently collected training-only AOIs and complete tile labels; holdout
+  review evidence must not be recycled into training.
+
+Next:
+- Collect a new training-only small-building/background evidence pack outside
+  all operational holdouts, then train an inactive candidate only if the pack
+  passes label, leakage and sample-volume audits.

docs/TODO.md

@@ -50,7 +50,8 @@ This file now starts with the current implementation status. Older preparation/b
 - [x] Persist paginated false-positive/false-negative operator decisions and expose them in the Quality workspace.
 - [x] Bound QA evidence resolution to persisted evidence ids instead of loading complete regional reference datasets.
 - [x] Compare confidence `0.10` and `0.15` over independent Mol holdouts; retain `0.15` because it wins F1 in every positive zone while both pass the empty-background control.
-- [ ] Complete manual decisions for the generated 48 false-negative and 48 false-positive review cards before constructing any new training corpus.
+- [x] Complete conservative decisions for the generated 48 false-negative and 48 false-positive review cards, validate both roles fail-closed and retain only 15 confirmed model errors.
+- [x] Audit the confirmed review evidence against the active corpus and holdouts; reject retraining because it provides zero novel leakage-free labels.
 - [x] Backend FastAPI foundation, health endpoint and service structure.
 - [x] React/TypeScript frontend foundation and MapLibre workbench.
 - [x] Map layer visibility, opacity and feature property inspection.

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.