Promote focused small-building detector

backend/README.md

@@ -345,6 +345,12 @@ Then point `OPERATOR_YOLO_DATASET_DIR` at
 training wrapper. Tile-level output remains operator tooling outside the V1
 browser product.
 
+Use `--samples` (or `OPERATOR_YOLO_SAMPLES`) when an experiment needs a
+deliberate manifest subset. The generated summary records the source manifest
+count plus selected and excluded sample slugs. Unknown samples and any selected
+manifest holdout that is omitted from `--val-samples` fail before files are
+written.
+
 The backend also exposes a read-only model asset catalog for the mounted model
 directory:
 
@@ -462,13 +468,24 @@ docker exec -it geointel python3 /app/scripts/prepare_operator_real_data_samples
 ```
 
 The helper writes GeoTIFF orthophotos, GRB GBG building GeoJSON files and
-`operator_samples_manifest.json` under `/app/storage/operator-data`. The corpus
-contains dense reference AOIs for Geel, Mol, Turnhout, Herentals, Balen, Retie
-and Westerlo plus explicitly marked background candidates for Postel-bos,
-Lommel-heide and Kasterlee-bos. Background candidates can persist empty GRB
-FeatureCollections for negative-tile training; normal reference AOIs still fail
-when GRB returns no buildings. These are runtime artifacts only and are not
-committed to Git.
+`operator_samples_manifest.json` under `/app/storage/operator-data`. In
+addition to the established positive and background AOIs, the registry contains
+Beerse, Rijkevorsel, Hoogstraten and Vorselaar as focused small-building
+training AOIs. Vosselaar and Grobbendonk are independent validation AOIs and
+must not be exported into the training split. Background candidates can persist
+empty GRB FeatureCollections for negative-tile training; normal reference AOIs
+still fail when GRB returns no buildings. These are runtime artifacts only and
+are not committed to Git.
+
+The current recommended local building model is
+`geointel-building-yolov8s-smallbld-minpx3-img640-ft30-pt` with tile size
+`512`, overlap `64` and confidence threshold `0.15`. Its SHA256 is
+`a9088b8491dfae36694b53e9e9406cb4e3511d334a5712fa34f75078a47759c1`.
+The promotion evidence covers seven positive AOIs at QA match IoU `0.25` and
+three pure-empty background AOIs. The model improves recall and persisted
+false-negative counts, but has lower precision than the previous balanced
+model; operators must review and persist QA/QC rather than treating detections
+as ground truth.
 
 For model-quality calibration, run the confidence sweep wrapper: