docs: record fail-closed accuracy challenger

CHANGELOG.md

@@ -7,6 +7,25 @@
 
 # Changelog
 
+## Sprint 199 Reviewed accuracy expansion (2026-07-15)
+
+- Added six leakage-free training AOIs in Arendonk, Dessel, Meerhout, Laakdal,
+  Nijlen and Hulshout, backed by 9,964 paged GRB building references.
+- Exported and audited a 252-tile, 79,192-label corpus; the configured audit and
+  balanced 64-tile visual review found no invalid, missing or low-variance
+  selections.
+- Fine-tuned one inactive local YOLOv8s challenger for 20 CPU epochs without
+  downloads. Its best checkpoint SHA256 is
+  `038f1f97a6afd534f29e1f392a730a58207b928ca01e31ab8d8fed6106705820`.
+- Re-ran active and challenger models through the same current persisted QA/QC
+  pipeline on four Mol and three regional holdouts. The challenger improved
+  mean F1 from `0.6069` to `0.6248` and improved every zone.
+- Retained the active model because the challenger produced two detections in
+  empty Postel-bos; the active profile remained zero across all three empty
+  controls. No runtime model or `.env` setting was changed.
+- Replaced stale Detection Lab profile averages with coverage-aligned active
+  evidence: precision `0.6141`, recall `0.6062`, F1 `0.6069`.
+
 ## Sprint 198 Evidence-closed model review (2026-07-15)
 
 - Completed visual and geometric review of 48 persisted false-positive and 48