Promote focused small-building detector

CHANGELOG.md

@@ -7,6 +7,16 @@
 
 # Changelog
 
+## Sprint 174 Focused small-building model promotion (2026-07-13)
+
+- Expanded the real operator corpus with four focused training AOIs and two independent validation AOIs, while keeping Turnhout, Retie and Westerlo outside the tile-training corpus as operation-level holdouts.
+- Exported and visually audited 198 tiles with 58,820 real GRB-derived labels; the accepted corpus contained no invalid labels, missing files or low-variance review selections.
+- Trained `geointel-building-yolov8s-smallbld-minpx3-img640-ft30-pt` from the previous active local model without downloading weights; the trained asset SHA256 is `a9088b8491dfae36694b53e9e9406cb4e3511d334a5712fa34f75078a47759c1`.
+- At tile `512`, overlap `64`, threshold `0.15` and QA match IoU `0.25`, seven persisted AOIs reached mean precision `0.5898`, recall `0.5770`, F1 `0.5825` and minimum F1 `0.5528`; all three pure-empty controls remained at zero detections.
+- Fixed-reference evidence reduced false negatives from 7,753 to 6,182, including 745 fewer 25-100 m2 misses and 181 fewer sub-25 m2 misses. Precision is lower, so the UI states the increased false-positive review load explicitly.
+- Added persistent false-negative area summaries/GeoJSON, explicit tile-corpus sample selection provenance, missing runtime evaluation scripts, Docker COPY-source regression checks and warning-free Pydantic model-field schemas.
+- Guarded activation updated only the Tower AI/model environment values; no fake outputs, provider fetching, API contract or database migration changed.
+
 ## Sprint 173 Expanded building model promotion (2026-07-13)
 
 - Trained and fully gated the inactive `geointel-building-yolov8s-aoi1024expandedminpx4vis035e50-pt` candidate from the 20-source expanded real-data corpus.

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:
 

backend/app/schemas/detection.py

@@ -3,10 +3,12 @@ from __future__ import annotations
 from datetime import datetime
 from uuid import UUID
 
-from pydantic import BaseModel, Field
+from pydantic import BaseModel, ConfigDict, Field
 
 
 class DetectionModelCapability(BaseModel):
+    model_config = ConfigDict(protected_namespaces=())
+
     model_id: str
     display_name: str
     framework: str
@@ -23,6 +25,8 @@ class DetectionModelsResponse(BaseModel):
 
 
 class ModelAssetRead(BaseModel):
+    model_config = ConfigDict(protected_namespaces=())
+
     model_asset_id: str
     filename: str
     display_name: str
@@ -39,12 +43,16 @@ class ModelAssetRead(BaseModel):
 
 
 class ModelAssetListResponse(BaseModel):
+    model_config = ConfigDict(protected_namespaces=())
+
     items: list[ModelAssetRead]
     total: int
     model_directory: str
 
 
 class DetectionRunRequest(BaseModel):
+    model_config = ConfigDict(protected_namespaces=())
+
     project_id: UUID
     dataset_id: UUID
     model_id: str
@@ -63,6 +71,8 @@ class DetectionQaRequest(BaseModel):
 
 
 class DetectionRunResponse(BaseModel):
+    model_config = ConfigDict(protected_namespaces=())
+
     analysis_run_id: UUID
     job_id: UUID
     project_id: UUID
@@ -75,6 +85,8 @@ class DetectionRunResponse(BaseModel):
 
 
 class DetectionRunRead(BaseModel):
+    model_config = ConfigDict(from_attributes=True, protected_namespaces=())
+
     id: UUID
     project_id: UUID
     dataset_id: UUID | None = None
@@ -90,8 +102,6 @@ class DetectionRunRead(BaseModel):
     started_at: datetime | None = None
     finished_at: datetime | None = None
 
-    model_config = {"from_attributes": True}
-
 
 class DetectionRunListResponse(BaseModel):
     items: list[DetectionRunRead]
@@ -99,6 +109,8 @@ class DetectionRunListResponse(BaseModel):
 
 
 class DetectionRead(BaseModel):
+    model_config = ConfigDict(from_attributes=True, protected_namespaces=())
+
     id: UUID
     project_id: UUID
     dataset_id: UUID | None = None
@@ -113,8 +125,6 @@ class DetectionRead(BaseModel):
     properties_json: dict | None = None
     created_at: datetime | None = None
 
-    model_config = {"from_attributes": True}
-
 
 class DetectionListResponse(BaseModel):
     items: list[DetectionRead]

backend/app/schemas/segmentation.py

@@ -3,7 +3,7 @@ from __future__ import annotations
 from datetime import datetime
 from uuid import UUID
 
-from pydantic import BaseModel, Field
+from pydantic import BaseModel, ConfigDict, Field
 
 from app.schemas.detection import DetectionModelCapability
 
@@ -16,6 +16,8 @@ class SegmentationModelsResponse(BaseModel):
 
 
 class SegmentationRunRequest(BaseModel):
+    model_config = ConfigDict(protected_namespaces=())
+
     project_id: UUID
     dataset_id: UUID
     model_id: str
@@ -33,6 +35,8 @@ class SegmentationQaRequest(BaseModel):
 
 
 class SegmentationRunResponse(BaseModel):
+    model_config = ConfigDict(protected_namespaces=())
+
     analysis_run_id: UUID
     job_id: UUID
     project_id: UUID
@@ -45,6 +49,8 @@ class SegmentationRunResponse(BaseModel):
 
 
 class SegmentationRunRead(BaseModel):
+    model_config = ConfigDict(from_attributes=True, protected_namespaces=())
+
     id: UUID
     project_id: UUID
     dataset_id: UUID | None = None
@@ -60,8 +66,6 @@ class SegmentationRunRead(BaseModel):
     started_at: datetime | None = None
     finished_at: datetime | None = None
 
-    model_config = {"from_attributes": True}
-
 
 class SegmentationRunListResponse(BaseModel):
     items: list[SegmentationRunRead]
@@ -69,6 +73,8 @@ class SegmentationRunListResponse(BaseModel):
 
 
 class SegmentationRead(BaseModel):
+    model_config = ConfigDict(from_attributes=True, protected_namespaces=())
+
     id: UUID
     project_id: UUID
     dataset_id: UUID | None = None
@@ -87,8 +93,6 @@ class SegmentationRead(BaseModel):
     provenance_json: dict | None = None
     created_at: datetime | None = None
 
-    model_config = {"from_attributes": True}
-
 
 class SegmentationListResponse(BaseModel):
     items: list[SegmentationRead]

backend/tests/test_docker_runtime_config.py

@@ -58,14 +58,33 @@ def test_all_in_one_dockerfile_can_opt_into_ai_dependencies_without_base_install
 def test_all_in_one_dockerfile_copies_operator_scripts_for_runtime_use() -> None:
     dockerfile = (ROOT / "deploy" / "unraid" / "Dockerfile.all-in-one").read_text(encoding="utf-8")
 
-    assert "COPY scripts/prepare_operator_real_data_samples.py /app/scripts/prepare_operator_real_data_samples.py" in dockerfile
-    assert "COPY scripts/export_operator_yolo_tile_dataset.py /app/scripts/export_operator_yolo_tile_dataset.py" in dockerfile
-    assert "COPY scripts/audit_operator_yolo_dataset_quality.py /app/scripts/audit_operator_yolo_dataset_quality.py" in dockerfile
-    assert (
-        "COPY scripts/render_operator_yolo_label_qa_contact_sheets.py "
-        "/app/scripts/render_operator_yolo_label_qa_contact_sheets.py"
-    ) in dockerfile
-    assert "COPY scripts/train_operator_yolo_detector.sh /app/scripts/train_operator_yolo_detector.sh" in dockerfile
+    for line in dockerfile.splitlines():
+        if line.startswith("COPY scripts/"):
+            source_path = line.split()[1]
+            assert (ROOT / source_path).is_file()
+
+    required_runtime_scripts = {
+        "prepare_operator_real_data_samples.py",
+        "export_operator_yolo_tile_dataset.py",
+        "audit_operator_yolo_dataset_quality.py",
+        "render_operator_yolo_label_qa_contact_sheets.py",
+        "train_operator_yolo_detector.sh",
+        "verify_real_data_detection_qa_workflow.sh",
+        "run_detection_quality_matrix.sh",
+        "run_multi_sample_detection_quality_matrix.sh",
+        "export_detection_calibration_evidence.sh",
+        "assemble_detection_calibration_evidence_portfolio.sh",
+        "build_fixed_threshold_evidence_portfolio_inputs.py",
+        "audit_detection_false_negative_evidence.py",
+        "run_operator_hard_negative_detection_matrix.sh",
+        "run_background_corpus_split_matrix.sh",
+        "build_background_corpus_split_report.py",
+        "build_detection_model_promotion_report.py",
+        "run_split_background_promotion_workflow.sh",
+        "activate_promoted_yolo_candidate.py",
+    }
+    for script_name in required_runtime_scripts:
+        assert f"COPY scripts/{script_name} /app/scripts/{script_name}" in dockerfile
 
 
 def test_all_in_one_dockerfile_copies_operator_scripts_after_dependency_install() -> None:

backend/tests/test_schema_model_field_warnings.py

@@ -0,0 +1,17 @@
+from pydantic import BaseModel
+
+from app.schemas import detection, segmentation
+
+
+def test_model_prefixed_api_fields_are_explicitly_supported() -> None:
+    schemas = [
+        value
+        for module in (detection, segmentation)
+        for value in vars(module).values()
+        if isinstance(value, type)
+        and issubclass(value, BaseModel)
+        and any(field_name.startswith("model_") for field_name in value.model_fields)
+    ]
+
+    assert schemas
+    assert all(schema.model_config.get("protected_namespaces") == () for schema in schemas)

backend/tests/test_sprint130_operator_yolo_tile_dataset.py

@@ -69,6 +69,7 @@ def test_operator_yolo_tile_dataset_export_help_does_not_require_gis_dependencie
     assert "Export operator real-data samples to a tile-level YOLO detection dataset" in result.stdout
     assert "--tile-size" in result.stdout
     assert "--stride" in result.stdout
+    assert "--samples" in result.stdout
     assert "--negative-keep-ratio" in result.stdout
     assert "--min-label-visible-ratio" in result.stdout
     assert "--background-negative-repeat" in result.stdout
@@ -84,10 +85,19 @@ def test_default_validation_split_is_explicit_and_rejects_holdout_leakage() -> N
         {"sample_slug": "retie", "recommended_split": "val"},
         {"sample_slug": "westerlo", "recommended_split": "val"},
         {"sample_slug": "arendonk_heide", "recommended_split": "val"},
+        {"sample_slug": "vosselaar_center", "recommended_split": "val"},
+        {"sample_slug": "grobbendonk_center", "recommended_split": "val"},
     ]
 
     assert module.DEFAULT_VALIDATION_SAMPLE_SLUGS == frozenset(
-        {"turnhout", "retie", "westerlo", "arendonk_heide"}
+        {
+            "turnhout",
+            "retie",
+            "westerlo",
+            "arendonk_heide",
+            "vosselaar_center",
+            "grobbendonk_center",
+        }
     )
     assert module.validate_validation_split(
         samples,
@@ -100,6 +110,35 @@ def test_default_validation_split_is_explicit_and_rejects_holdout_leakage() -> N
         module.validate_validation_split(samples, {"turnhout", "missing"})
 
 
+def test_manifest_sample_selection_keeps_external_holdouts_out_of_targeted_dataset() -> None:
+    module = load_tile_exporter()
+    samples = [
+        {"sample_slug": "geel", "recommended_split": "train"},
+        {"sample_slug": "beerse_center", "recommended_split": "train"},
+        {"sample_slug": "vosselaar_center", "recommended_split": "val"},
+        {"sample_slug": "turnhout", "recommended_split": "val"},
+        {"sample_slug": "retie", "recommended_split": "val"},
+        {"sample_slug": "westerlo", "recommended_split": "val"},
+    ]
+
+    selected, excluded = module.select_manifest_samples(
+        samples,
+        {"geel", "beerse_center", "vosselaar_center"},
+    )
+
+    assert [sample["sample_slug"] for sample in selected] == [
+        "geel",
+        "beerse_center",
+        "vosselaar_center",
+    ]
+    assert excluded == ["retie", "turnhout", "westerlo"]
+    assert module.validate_validation_split(selected, {"vosselaar_center"}) == {
+        "vosselaar_center"
+    }
+    with pytest.raises(SystemExit, match="unknown samples"):
+        module.select_manifest_samples(samples, {"geel", "missing"})
+
+
 def test_validation_coverage_reports_holdouts_without_retained_tiles() -> None:
     module = load_tile_exporter()
     coverage = module.validation_sample_coverage(

backend/tests/test_sprint131_operator_sample_expansion.py

@@ -51,7 +51,7 @@ def test_operator_training_expansion_preserves_geographically_separate_holdouts(
     expected_holdouts = {"turnhout", "retie", "westerlo", "arendonk_heide"}
 
     assert module.TRAINING_EXPANSION_SAMPLE_SLUGS == frozenset(expected_expansion)
-    assert module.DEFAULT_VALIDATION_SAMPLE_SLUGS == frozenset(expected_holdouts)
+    assert expected_holdouts.issubset(module.DEFAULT_VALIDATION_SAMPLE_SLUGS)
     assert all(module.SAMPLES[slug].sample_role == "reference" for slug in expected_expansion)
     assert all(not module.SAMPLES[slug].allow_empty_reference for slug in expected_expansion)
     assert all(module.recommended_split_for_sample(module.SAMPLES[slug]) == "train" for slug in expected_expansion)
@@ -78,6 +78,51 @@ def test_operator_training_expansion_preserves_geographically_separate_holdouts(
         ) >= 2_000
 
 
+def test_small_building_expansion_has_separate_training_and_validation_centers() -> None:
+    module = load_sample_preparer()
+
+    expected_training = {
+        "beerse_center",
+        "rijkevorsel_center",
+        "hoogstraten_center",
+        "vorselaar_center",
+    }
+    expected_validation = {"vosselaar_center", "grobbendonk_center"}
+
+    assert module.SMALL_BUILDING_TRAINING_SAMPLE_SLUGS == frozenset(expected_training)
+    assert module.SMALL_BUILDING_VALIDATION_SAMPLE_SLUGS == frozenset(expected_validation)
+    assert expected_validation.issubset(module.DEFAULT_VALIDATION_SAMPLE_SLUGS)
+    assert all(module.SAMPLES[slug].sample_role == "reference" for slug in expected_training | expected_validation)
+    assert all(
+        module.recommended_split_for_sample(module.SAMPLES[slug]) == "train"
+        for slug in expected_training
+    )
+    assert all(
+        module.recommended_split_for_sample(module.SAMPLES[slug]) == "val"
+        for slug in expected_validation
+    )
+
+    def distance_m(left, right) -> float:
+        radius_m = 6_371_008.8
+        left_lat = math.radians(left.center_lat)
+        right_lat = math.radians(right.center_lat)
+        delta_lat = right_lat - left_lat
+        delta_lon = math.radians(right.center_lon - left.center_lon)
+        haversine = (
+            math.sin(delta_lat / 2) ** 2
+            + math.cos(left_lat) * math.cos(right_lat) * math.sin(delta_lon / 2) ** 2
+        )
+        return 2 * radius_m * math.asin(math.sqrt(haversine))
+
+    protected_holdouts = expected_validation | {"turnhout", "retie", "westerlo"}
+    for training_slug in expected_training:
+        training_sample = module.SAMPLES[training_slug]
+        assert min(
+            distance_m(training_sample, module.SAMPLES[holdout_slug])
+            for holdout_slug in protected_holdouts
+        ) >= 2_000
+
+
 def test_operator_background_candidates_are_unique_enough_for_hard_negative_training() -> None:
     module = load_sample_preparer()
 

backend/tests/test_sprint155_detection_operator_profiles.py

@@ -9,8 +9,10 @@ def test_detection_operator_profiles_define_explicit_yolo_candidates_and_promote
     source = profiles.read_text(encoding="utf-8")
 
     assert "DETECTION_OPERATOR_PROFILES" in source
+    assert "geointel-building-yolov8s-smallbld-minpx3-img640-ft30-pt" in source
     assert "geointel-building-yolov8s-aoi1024expandedminpx4vis035e50-pt" in source
     assert "geointel-building-yolov8s-aoi1024bg512r3e50-pt" in source
+    assert "small-building-balanced-review" in source
     assert "expanded-balanced-review" in source
     assert "conservative-review" in source
     assert "confidenceThreshold: 0.15" in source
@@ -18,10 +20,12 @@ def test_detection_operator_profiles_define_explicit_yolo_candidates_and_promote
     assert "defaultApproved: true" in source
     assert "promotionRecommendation: 'promote_candidate'" in source
     assert "positiveSampleCount: 7" in source
+    assert "f1: 0.5824578631584316" in source
     assert "f1: 0.5432865390636915" in source
     assert "maxBackgroundDetections: 0" in source
     assert "pure-empty gate passed" in source
-    assert "persistent small-building misses" in source
+    assert "1,571 fewer false negatives" in source
+    assert "higher false-positive review load" in source
 
 
 def test_detection_lab_surfaces_profiles_as_deliberate_operator_actions() -> None:

backend/tests/test_sprint170_detection_false_negative_audit.py

@@ -199,6 +199,28 @@ def test_false_negative_audit_finds_persistent_reference_misses(tmp_path: Path)
     assert candidate["false_negative_count"] == 1
     assert candidate["false_negative_rate"] == 1 / 3
     assert active["false_negative_area_m2"]["median"] > 0
+    assert sample["persistent_false_negative_area_m2"]["count"] == 1
+    assert sample["persistent_false_negative_area_m2"]["median"] > 0
+    assert sum(
+        bucket["count"] for bucket in sample["persistent_area_buckets"].values()
+    ) == 1
+    assert sum(
+        bucket["share"] for bucket in sample["persistent_area_buckets"].values()
+    ) == 1.0
+    persistent_evidence = json.loads(
+        (output_dir / "persistent_false_negatives.geojson").read_text(encoding="utf-8")
+    )
+    assert persistent_evidence["type"] == "FeatureCollection"
+    assert len(persistent_evidence["features"]) == 1
+    persistent_feature = persistent_evidence["features"][0]
+    assert persistent_feature["properties"]["qa_evidence_role"] == "persistent_false_negative"
+    assert persistent_feature["properties"]["sample_slug"] == "geel"
+    assert persistent_feature["properties"]["persistent_reference_id"] == "source:persistent-small"
+    assert persistent_feature["properties"]["area_m2"] > 0
+    assert persistent_feature["properties"]["area_bucket"] in sample["persistent_area_buckets"]
+    assert report["persistent_evidence_geojson_path"] == str(
+        output_dir / "persistent_false_negatives.geojson"
+    )
     assert report["recommendations"]
     assert (output_dir / "detection_false_negative_audit.md").is_file()
 

deploy/unraid/Dockerfile.all-in-one

@@ -75,12 +75,34 @@ COPY scripts/export_operator_yolo_tile_dataset.py /app/scripts/export_operator_y
 COPY scripts/audit_operator_yolo_dataset_quality.py /app/scripts/audit_operator_yolo_dataset_quality.py
 COPY scripts/render_operator_yolo_label_qa_contact_sheets.py /app/scripts/render_operator_yolo_label_qa_contact_sheets.py
 COPY scripts/train_operator_yolo_detector.sh /app/scripts/train_operator_yolo_detector.sh
+COPY scripts/verify_real_data_detection_qa_workflow.sh /app/scripts/verify_real_data_detection_qa_workflow.sh
+COPY scripts/run_detection_quality_matrix.sh /app/scripts/run_detection_quality_matrix.sh
+COPY scripts/run_multi_sample_detection_quality_matrix.sh /app/scripts/run_multi_sample_detection_quality_matrix.sh
+COPY scripts/export_detection_calibration_evidence.sh /app/scripts/export_detection_calibration_evidence.sh
+COPY scripts/assemble_detection_calibration_evidence_portfolio.sh /app/scripts/assemble_detection_calibration_evidence_portfolio.sh
+COPY scripts/build_fixed_threshold_evidence_portfolio_inputs.py /app/scripts/build_fixed_threshold_evidence_portfolio_inputs.py
+COPY scripts/audit_detection_false_negative_evidence.py /app/scripts/audit_detection_false_negative_evidence.py
+COPY scripts/run_operator_hard_negative_detection_matrix.sh /app/scripts/run_operator_hard_negative_detection_matrix.sh
+COPY scripts/run_background_corpus_split_matrix.sh /app/scripts/run_background_corpus_split_matrix.sh
+COPY scripts/build_background_corpus_split_report.py /app/scripts/build_background_corpus_split_report.py
+COPY scripts/build_detection_model_promotion_report.py /app/scripts/build_detection_model_promotion_report.py
+COPY scripts/run_split_background_promotion_workflow.sh /app/scripts/run_split_background_promotion_workflow.sh
+COPY scripts/activate_promoted_yolo_candidate.py /app/scripts/activate_promoted_yolo_candidate.py
 COPY deploy/unraid/nginx-all-in-one.conf /etc/nginx/conf.d/default.conf
 COPY deploy/unraid/all-in-one-start.sh /usr/local/bin/geointel-all-in-one-start
 COPY --from=frontend-build /frontend/dist/ /usr/share/nginx/html/
 
 RUN chmod +x /usr/local/bin/geointel-all-in-one-start \
-    && chmod +x /app/scripts/train_operator_yolo_detector.sh
+    && chmod +x \
+      /app/scripts/train_operator_yolo_detector.sh \
+      /app/scripts/verify_real_data_detection_qa_workflow.sh \
+      /app/scripts/run_detection_quality_matrix.sh \
+      /app/scripts/run_multi_sample_detection_quality_matrix.sh \
+      /app/scripts/export_detection_calibration_evidence.sh \
+      /app/scripts/assemble_detection_calibration_evidence_portfolio.sh \
+      /app/scripts/run_operator_hard_negative_detection_matrix.sh \
+      /app/scripts/run_background_corpus_split_matrix.sh \
+      /app/scripts/run_split_background_promotion_workflow.sh
 
 VOLUME ["/var/lib/postgresql/data", "/app/storage"]
 

docs/AI_PIPELINES.md

@@ -335,17 +335,22 @@ hard-negative gates, then run `sparse_building_context` as a separate review
 matrix. The first expanded local model improved dense AOI F1, but Kasterlee-bos
 false positives block default promotion.
 
-The expanded-AOI local model asset,
-`geointel-building-yolov8s-aoi1024expandedminpx4vis035e50-pt`, is the current
+The focused small-building local model asset,
+`geointel-building-yolov8s-smallbld-minpx3-img640-ft30-pt`, is the current
 recommended Detection Lab operator profile. Use tile size `512`, overlap `64`
-and confidence threshold `0.15`. Persisted QA/QC across seven positive AOIs
-measured mean precision `0.6471`, recall `0.4700` and F1 `0.5433`; the strict
-three-sample pure-empty background gate produced zero detections. The previous
-`geointel-building-yolov8s-aoi1024bg512r3e50-pt` model remains available as a
-legacy conservative `0.35` review profile. Sparse-context detections remain
-review-only evidence, not a default-promotion blocker. Persistent misses are
-concentrated in small buildings, so every production-like run still requires
-persisted QA/QC against suitable reference data.
+and confidence threshold `0.15`. Persisted QA/QC at match IoU `0.25` across
+seven positive AOIs measured mean precision `0.5898`, recall `0.5770` and F1
+`0.5825`; minimum per-AOI F1 was `0.5528`. The strict three-sample pure-empty
+background gate produced zero detections. Compared with the previous balanced
+profile, the same persisted reference populations contain 1,571 fewer false
+negatives, including 745 fewer misses in the 25-100 m2 bucket and 181 fewer
+below 25 m2. This recall gain increases the false-positive review load, so the
+previous `geointel-building-yolov8s-aoi1024expandedminpx4vis035e50-pt` profile
+remains available as a higher-precision legacy `0.15` choice. The older
+`geointel-building-yolov8s-aoi1024bg512r3e50-pt` remains the conservative
+`0.35` profile. Sparse-context detections remain review-only evidence, not a
+default-promotion blocker. Every production-like run still requires persisted
+QA/QC against suitable reference data.
 
 To update a Tower/Unraid `.env` from a promoted report, use the guarded
 activation helper. It validates the exact report candidate key, verifies that
@@ -355,8 +360,8 @@ when `--apply` is supplied:
 
 ```bash
 python scripts/activate_promoted_yolo_candidate.py \
-  --promotion-report artifacts/detection-model-promotion/split-aware/aoi1024expandedminpx4vis035e50-split/detection_model_promotion_report.json \
-  --candidate-key 'geointel-building-yolov8s-aoi1024expandedminpx4vis035e50-pt|512|64|0.15' \
+  --promotion-report storage/operator-data/model-review/small-building-candidate/promotion/detection_model_promotion_report.json \
+  --candidate-key 'geointel-building-yolov8s-smallbld-minpx3-img640-ft30-pt|512|64|0.15' \
   --models-dir /mnt/user/appdata/geointel/models \
   --env-file /mnt/user/appdata/geointel/.env \
   --json
@@ -440,6 +445,13 @@ and records the positive/negative tile counts. This gives the training smoke
 more image samples while preserving the same explicit operator-data and QA/QC
 validation boundary.
 
+Focused small-building experiments use Beerse, Rijkevorsel, Hoogstraten and
+Vorselaar as training AOIs, with Vosselaar and Grobbendonk retained as
+independent validation AOIs. The exporter accepts an explicit `--samples`
+subset and records `source_manifest_sample_count`, `selected_sample_slugs` and
+`excluded_sample_slugs` in its summary. Manifest-backed validation samples
+cannot silently enter training.
+
 For visual error inspection, export the persisted QA evidence from a calibration
 summary:
 

docs/CODEX_EXECUTION_LOG.md

@@ -7071,3 +7071,68 @@ Open:
 - Embedded PostGIS live migration smoke passed with PostGIS `3.6`, required tables/indexes, database collation and the single Alembic head `202606120900`.
 - Browser validation confirmed that the recommended profile selects the exact active asset and threshold `0.15`, while live preflight displays CPU dependency/model readiness and the expected missing-manifest guard before dataset handoff.
 - Browser console warnings/errors: `0`.
+
+# Sprint 174 - Focused small-building recovery and promotion
+
+## Data and training evidence
+
+- Converted the Sprint 173 persistent false-negative audit into one focused real-data experiment instead of extending the same corpus blindly.
+- Added Beerse, Rijkevorsel, Hoogstraten and Vorselaar as training AOIs and Vosselaar/Grobbendonk as independent tile-level validation AOIs.
+- Kept Turnhout, Retie and Westerlo outside the tile corpus as operation-level holdouts.
+- Exported `/app/storage/operator-data/yolo-building-aoi1024-smallbld-minpx3vis035` from an explicit 23-sample manifest subset:
+  - 198 retained tiles;
+  - 180 positive and 18 negative tiles;
+  - 58,820 real GRB-derived labels;
+  - 48 visually reviewed tiles;
+  - zero invalid labels, missing images, missing label files or low-variance review selections.
+- The accepted `min-label-px=3` corpus retained 1,228 more genuine small-building labels than the comparable `min-label-px=4` export.
+- Trained one inactive 30-epoch CPU candidate from the previous active local model:
+  - model: `geointel-building-yolov8s-smallbld-minpx3-img640-ft30.pt`;
+  - model SHA256: `a9088b8491dfae36694b53e9e9406cb4e3511d334a5712fa34f75078a47759c1`;
+  - dataset-summary SHA256: `49b2a07d2105d08356431757b83eafc1498eaf1fb76965b1efe05b776824942a`;
+  - dataset-YAML SHA256: `3a2ea97c35a18072a1ab6738cd673c0ecec5344b19461c91d72a15e138d46e8d`;
+  - no model download and no fake training or QA data.
+
+## Persisted promotion evidence
+
+- Evaluated the exact fixed profile `tile=512`, `overlap=64`, `confidence=0.15` with QA match IoU explicitly fixed at `0.25`.
+- Seven positive AOIs produced:
+  - mean precision `0.5898197518`;
+  - mean recall `0.5769921004`;
+  - mean F1 `0.5824578632`;
+  - minimum per-AOI F1 `0.5527837436`.
+- Every AOI improved F1 relative to the previous balanced model. Turnhout improved from `0.4897494305` to `0.5527837436`.
+- The strict pure-empty gate covered Postel, Lommel and Arendonk and produced zero detections for every sample.
+- The formal promotion report recommended the exact key `geointel-building-yolov8s-smallbld-minpx3-img640-ft30-pt|512|64|0.15`.
+- Fixed-reference object evidence used identical GRB feature identities and reduced false negatives from `7,753` to `6,182`:
+  - 1,571 fewer total false negatives;
+  - 745 fewer misses in the 25-100 m2 bucket;
+  - 181 fewer misses below 25 m2;
+  - all seven AOIs improved.
+- Remaining persistent misses total `5,838`, concentrated in Turnhout, Herentals and Geel and still dominated by small buildings.
+- Mean precision decreased from `0.6470590036` to `0.5898197518`. The new profile is therefore a recall-balanced operator default with a higher false-positive review load, not ground truth.
+
+## Repository hardening
+
+- Added explicit `--samples` / `OPERATOR_YOLO_SAMPLES` corpus selection with selected/excluded sample provenance and unknown-sample rejection.
+- Added persistent false-negative area statistics, size buckets and combined GeoJSON review evidence.
+- Copied the complete operator evaluation/promotion toolchain into the all-in-one image and added a regression that rejects every Docker `COPY scripts/...` source that does not exist.
+- Removed Pydantic protected-namespace warnings for legitimate `model_*` API fields while preserving all schema field names and response contracts.
+- Updated Detection Lab profiles: the new small-building profile is recommended, the previous expanded profile remains the higher-precision legacy choice, and the background-aware `0.35` profile remains conservative.
+- Guarded activation first returned `ready_to_apply`; the reviewed `--apply` pass updated only `GEOINTEL_INSTALL_AI`, `YOLO_ENABLED`, `YOLO_MODELS_DIR` and `YOLO_MODEL_PATH` in the Tower environment.
+
+## Local validation
+
+- `python -m compileall backend/app`: passed.
+- `python -m pytest`: 472 passed.
+- `python -m ruff check` for all changed Python modules/tests: passed.
+- `npm run typecheck`: passed.
+- `npm run build`: passed; app bundle `215.64 kB`, MapLibre bundle `801.82 kB` before gzip.
+- `bash scripts/run_readiness_check.sh`: passed with 472 tests.
+- `python -m alembic heads`: one head, `202606120900`.
+- `python -m alembic upgrade head --sql`: complete migration chain rendered successfully.
+- Shell syntax checks passed for live migration and the full operator evaluation/promotion chain.
+
+## Next recommended pass
+
+- After redeploy, verify the active model SHA, local model-load preflight, live PostGIS migration smoke and browser profile selection. Then review false-positive evidence and the remaining 5,838 persistent misses before any further training.

docs/TODO.md

@@ -137,7 +137,8 @@ This file now starts with the current implementation status. Older preparation/b
 - [x] Add deterministic dataset/base/trained-model SHA256 provenance to future operator training summaries.
 - [x] Review per-AOI false-negative evidence, expand positive sample/label coverage and verify the resulting candidate improves false-negative rate in every validated AOI.
 - [x] Complete rebuild/restart and browser/runtime smoke for the guarded promoted V1 building detector activation.
-- [ ] Expand focused small-building training evidence only after reviewing persistent false negatives from the promoted model; do not start another blind training run.
+- [x] Expand focused small-building training evidence after reviewing persistent false negatives, train one inactive candidate and pass it through positive, pure-empty and fixed-reference promotion evidence before guarded activation.
+- [ ] Review the remaining 5,838 persistent false negatives and the increased false-positive load before any further model training; do not start another blind run.
 
 ## Sprint 8 status
 
@@ -494,7 +495,7 @@ This file now starts with the current implementation status. Older preparation/b
 - [x] Regenerate Tower AOI1024 operator samples with paged GRB references, then re-export and audit labels before any new training attempt.
 - [x] Improve AOI1024 label quality before retraining: `yolo-building-aoi1024-cleanpx12vis035` now audits `ok` with 14,632 labels, `min_label_px=12`, `min_label_visible_ratio=0.35`, median normalized box area `0.001373291015625` and small-box share `0.0`.
 - [x] Train and reject `geointel-building-yolov8s-aoi1024cleanpx12vis035e50-pt` through the positive/background promotion gate.
-- [ ] Keep every local YOLO candidate inactive until positive-AOI and hard-negative promotion reports recommend default activation.
+- [x] Keep every local YOLO candidate inactive until positive-AOI and hard-negative promotion reports recommend default activation.
 
 # Sprint 171 - Positive AOI expansion and small-building recovery
 
@@ -505,4 +506,6 @@ This file now starts with the current implementation status. Older preparation/b
 - [x] Refresh the full AOI1024 operator manifest on Tower and fetch only missing AOIs.
 - [x] Export and audit a low-minimum-label dataset without changing the active model.
 - [x] Render and inspect a sample-balanced label contact sheet before training.
-- [ ] Finish the inactive expanded-minpx4 candidate and run the full promotion gate.
+- [x] Finish the inactive expanded-minpx4 candidate and run the full promotion gate.
+- [x] Export the focused 23-sample minpx3 corpus with independent Vosselaar/Grobbendonk validation and external Turnhout/Retie/Westerlo holdouts.
+- [x] Train, audit and guarded-activate the focused small-building candidate only after all persisted promotion gates passed.

frontend/README.md

@@ -124,7 +124,7 @@ AI Lab run controls explicitly explain when no raster dataset is available, inst
 - Detection Lab now exposes the `yolo-configured` capability reported by the backend.
 - When `yolo-configured` is selected, users can provide an existing raster tile manifest path.
 - Detection Lab lists local model assets from `GET /api/v1/detection/model-assets` so operators can choose an existing mounted model file instead of editing only one hidden `YOLO_MODEL_PATH` slot.
-- Detection Lab exposes explicit operator profiles for the local AOI1024 building detector: balanced review at threshold `0.15` remains candidate-only, while conservative review at threshold `0.35` is marked as the promoted profile after the split-background pure-empty gate passed.
+- Detection Lab exposes explicit operator profiles for mounted local building detectors. The focused small-building model is the recommended recall-balanced `0.15` profile; the previous expanded-AOI `0.15` model remains available for higher precision, and the background-aware `0.35` model remains the conservative review choice. Applying a profile never downloads weights, changes runtime environment or starts inference automatically.
 - Applying a profile deliberately selects the local model asset and threshold for the browser-run request; runtime default activation remains a separate guarded `.env` operation through `scripts/activate_promoted_yolo_candidate.py`.
 - Detection Lab includes a read-only YOLO runtime preflight panel with backend status, dependency visibility, local model configuration, `torch`/`ultralytics` versions, CUDA state and `YOLO_CONFIG_DIR`.
 - The UI still does not download models or create fake detections; backend status and error codes remain the source of truth.

frontend/src/components/detection/detectionProfiles.ts

@@ -15,9 +15,26 @@ export interface DetectionOperatorProfile {
 }
 
 export const DETECTION_OPERATOR_PROFILES: DetectionOperatorProfile[] = [
+  {
+    id: 'small-building-balanced-review',
+    displayName: 'Recommended small-building review',
+    modelAssetId: 'geointel-building-yolov8s-smallbld-minpx3-img640-ft30-pt',
+    confidenceThreshold: 0.15,
+    defaultApproved: true,
+    promotionRecommendation: 'promote_candidate',
+    precision: 0.5898197517793451,
+    recall: 0.576992100419565,
+    f1: 0.5824578631584316,
+    positiveSampleCount: 7,
+    maxBackgroundDetections: 0,
+    description:
+      'Recommended recall-balanced profile for building review, with improved small-building coverage across seven Kempen AOIs.',
+    limitationMessage:
+      'The pure-empty gate passed and persisted QA found 1,571 fewer false negatives than the previous balanced profile; expect a higher false-positive review load.',
+  },
   {
     id: 'expanded-balanced-review',
-    displayName: 'Recommended balanced review',
+    displayName: 'Legacy expanded balanced review',
     modelAssetId: 'geointel-building-yolov8s-aoi1024expandedminpx4vis035e50-pt',
     confidenceThreshold: 0.15,
     defaultApproved: true,
@@ -27,9 +44,9 @@ export const DETECTION_OPERATOR_PROFILES: DetectionOperatorProfile[] = [
     f1: 0.5432865390636915,
     positiveSampleCount: 7,
     maxBackgroundDetections: 0,
-    description: 'Recommended expanded-AOI profile for balanced building review across the validated Kempen samples.',
+    description: 'Legacy expanded-AOI profile for review sessions where precision matters more than the newest recall gain.',
     limitationMessage:
-      'Default-approved after the pure-empty gate passed; persistent small-building misses still require operator QA.',
+      'The pure-empty gate passed; this profile has fewer false positives but more persistent small-building misses than the recommended profile.',
   },
   {
     id: 'conservative-review',

scripts/README.md

@@ -531,8 +531,27 @@ Current Tower audit status:
   `0.000694274766`, small-box share `0.3832694151486098`, no invalid labels and
   no missing label files. The balanced visual pass rendered 40 tiles across all
   19 source samples that retained at least one tile, with no invalid labels,
-  missing images or low-variance selections. A new candidate may be trained,
-  but remains inactive until positive and split-background promotion gates pass.
+  missing images or low-variance selections. Its promoted model remains the
+  higher-precision legacy `0.15` operator profile.
+- `yolo-building-aoi1024-smallbld-minpx3vis035`: focused small-building corpus
+  exported from an explicit 23-sample subset. Beerse, Rijkevorsel, Hoogstraten
+  and Vorselaar extend training; Vosselaar and Grobbendonk are validation-only;
+  Turnhout, Retie and Westerlo remain external operation-level holdouts. The
+  Tower export retained 198 tiles and 58,820 labels. Its small-object-aware
+  audit passed with no invalid/missing labels, and the 48-tile balanced visual
+  review contained no missing, invalid or low-variance selections. The trained
+  `geointel-building-yolov8s-smallbld-minpx3-img640-ft30-pt` candidate passed
+  seven positive-AOI and three pure-empty background gates at tile `512`,
+  overlap `64`, threshold `0.15` and QA match IoU `0.25`. Mean F1 is `0.5825`
+  and all pure-empty samples remain at zero detections. Persisted comparison
+  found 1,571 fewer false negatives than the previous balanced model, with a
+  lower mean precision and therefore a higher operator review load.
+
+Use `--samples` or `OPERATOR_YOLO_SAMPLES` to make an experimental corpus
+membership explicit. Dataset summaries preserve the complete manifest count,
+selected sample slugs and excluded sample slugs. Split validation still applies
+after filtering, so a manifest-backed holdout cannot be selected as training by
+omitting it from `--val-samples`.
 
 After rebuilding the all-in-one image, the operator scripts are available inside
 the container at `/app/scripts/...`. Before rebuilding, use the host checkout or
@@ -765,7 +784,10 @@ python scripts/audit_detection_false_negative_evidence.py \
 ```
 
 The audit reports false-negative rates and area buckets per AOI/model, plus
-reference buildings missed by every compared portfolio. Stable
+reference buildings missed by every compared portfolio. It writes the combined
+`persistent_false_negatives.geojson`, records geodetic persistent-miss area and
+adds persistent area buckets so operators can inspect the shared misses on a
+map instead of relying only on counts. Stable
 `source_feature_id` values are preferred; a normalized geometry fingerprint is
 used only when source IDs are absent. Invalid or missing geometry fails the
 audit instead of being silently skipped. The tools do not run inference,

scripts/audit_detection_false_negative_evidence.py

@@ -119,6 +119,7 @@ def audit_feature_collection(payload: dict[str, Any], geod: Any, shape: Any) ->
     reference_ids: set[str] = set()
     false_negative_areas: list[float] = []
     matched_reference_areas: list[float] = []
+    reference_records: dict[str, dict[str, Any]] = {}
     bucket_counts = {
         label: {"false_negative": 0, "matched_reference": 0, "total_reference": 0, "false_negative_rate": None}
         for label, _, _ in AREA_BUCKETS
@@ -144,6 +145,11 @@ def audit_feature_collection(payload: dict[str, Any], geod: Any, shape: Any) ->
         bucket = area_bucket(area_m2)
         reference_id = stable_reference_id(feature, geometry)
         reference_ids.add(reference_id)
+        reference_records[reference_id] = {
+            "area_m2": area_m2,
+            "area_bucket": bucket,
+            "feature": feature,
+        }
         bucket_role = "false_negative" if role == "false_negative" else "matched_reference"
         bucket_counts[bucket][bucket_role] += 1
         bucket_counts[bucket]["total_reference"] += 1
@@ -161,6 +167,7 @@ def audit_feature_collection(payload: dict[str, Any], geod: Any, shape: Any) ->
     return {
         "false_negative_ids": false_negative_ids,
         "reference_ids": reference_ids,
+        "reference_records": reference_records,
         "false_negative_count": len(false_negative_areas),
         "matched_reference_count": len(matched_reference_areas),
         "total_reference_count": total_reference,
@@ -285,6 +292,7 @@ def run_audit(portfolio_args: list[str], output_dir: Path) -> tuple[Path, Path]:
             )
 
     sample_reports: list[dict[str, Any]] = []
+    persistent_evidence_features: list[dict[str, Any]] = []
     for sample_slug in sorted(expected_slugs):
         portfolio_rows = []
         false_negative_sets = []
@@ -297,7 +305,7 @@ def run_audit(portfolio_args: list[str], output_dir: Path) -> tuple[Path, Path]:
                 {
                     key: value
                     for key, value in raw.items()
-                    if key not in {"false_negative_ids", "reference_ids"}
+                    if key not in {"false_negative_ids", "reference_ids", "reference_records"}
                 }
             )
         if any(reference_ids != reference_sets[0] for reference_ids in reference_sets[1:]):
@@ -309,16 +317,61 @@ def run_audit(portfolio_args: list[str], output_dir: Path) -> tuple[Path, Path]:
                 f"Sample {sample_slug} has different reference populations across portfolios ({counts})"
             )
         persistent_ids = sorted(set.intersection(*false_negative_sets))
+        reference_records = portfolio_samples[parsed_portfolios[0][0]][sample_slug]["reference_records"]
+        persistent_areas = [float(reference_records[reference_id]["area_m2"]) for reference_id in persistent_ids]
+        persistent_bucket_counts = {
+            label: {"count": 0, "share": 0.0}
+            for label, _, _ in AREA_BUCKETS
+        }
+        for reference_id in persistent_ids:
+            record = reference_records[reference_id]
+            persistent_bucket_counts[record["area_bucket"]]["count"] += 1
+            source_feature = record["feature"]
+            properties = dict(source_feature.get("properties") or {})
+            properties.update(
+                {
+                    "qa_evidence_role": "persistent_false_negative",
+                    "sample_slug": sample_slug,
+                    "persistent_reference_id": reference_id,
+                    "area_m2": record["area_m2"],
+                    "area_bucket": record["area_bucket"],
+                    "compared_portfolios": labels,
+                }
+            )
+            persistent_evidence_features.append(
+                {
+                    "type": "Feature",
+                    "id": f"{sample_slug}:{reference_id}",
+                    "properties": properties,
+                    "geometry": source_feature["geometry"],
+                }
+            )
+        if persistent_ids:
+            for values in persistent_bucket_counts.values():
+                values["share"] = values["count"] / len(persistent_ids)
         sample_reports.append(
             {
                 "sample_slug": sample_slug,
                 "reference_population_count": len(reference_sets[0]),
                 "persistent_false_negative_count": len(persistent_ids),
                 "persistent_reference_ids": persistent_ids,
+                "persistent_false_negative_area_m2": area_stats(persistent_areas),
+                "persistent_area_buckets": persistent_bucket_counts,
                 "portfolios": portfolio_rows,
             }
         )
 
+    output_dir = output_dir.expanduser().resolve()
+    output_dir.mkdir(parents=True, exist_ok=True)
+    persistent_geojson_path = output_dir / "persistent_false_negatives.geojson"
+    persistent_geojson_path.write_text(
+        json.dumps(
+            {"type": "FeatureCollection", "features": persistent_evidence_features},
+            indent=2,
+            sort_keys=True,
+        ),
+        encoding="utf-8",
+    )
     report = {
         "generated_at": datetime.now(timezone.utc).isoformat(),
         "schema_version": 1,
@@ -328,10 +381,9 @@ def run_audit(portfolio_args: list[str], output_dir: Path) -> tuple[Path, Path]:
         "portfolios": portfolio_meta,
         "sample_count": len(sample_reports),
         "samples": sample_reports,
+        "persistent_evidence_geojson_path": str(persistent_geojson_path),
         "recommendations": build_recommendations(sample_reports),
     }
-    output_dir = output_dir.expanduser().resolve()
-    output_dir.mkdir(parents=True, exist_ok=True)
     json_path = output_dir / "detection_false_negative_audit.json"
     json_path.write_text(json.dumps(report, indent=2, sort_keys=True), encoding="utf-8")
 
@@ -345,18 +397,26 @@ def run_audit(portfolio_args: list[str], output_dir: Path) -> tuple[Path, Path]:
         "",
         "## AOI comparison",
         "",
-        "| AOI | Persistent misses | "
+        "| AOI | Persistent misses | Persistent tiny/small | Persistent median m2 | "
         + " | ".join(f"{label} FN rate" for label, _ in parsed_portfolios)
         + " |",
-        "|---|---:|" + "---:|" * len(parsed_portfolios),
+        "|---|---:|---:|---:|" + "---:|" * len(parsed_portfolios),
     ]
     for sample in sample_reports:
         rates = [
             f"{(row['false_negative_rate'] or 0.0):.3f}"
             for row in sample["portfolios"]
         ]
+        persistent_buckets = sample["persistent_area_buckets"]
+        persistent_small_count = sum(
+            persistent_buckets[key]["count"]
+            for key in ("tiny_lt_25_m2", "small_25_100_m2")
+        )
+        persistent_median = sample["persistent_false_negative_area_m2"]["median"]
+        persistent_median_text = f"{persistent_median:.1f}" if persistent_median is not None else "n/a"
         lines.append(
             f"| {sample['sample_slug']} | {sample['persistent_false_negative_count']} | "
+            f"{persistent_small_count} | {persistent_median_text} | "
             + " | ".join(rates)
             + " |"
         )

scripts/export_operator_yolo_tile_dataset.py

@@ -26,7 +26,14 @@ PURE_EMPTY_BACKGROUND_CATEGORY = "pure_empty_negative"
 SPARSE_BACKGROUND_CATEGORY = "sparse_building_context"
 LOW_VARIANCE_NEGATIVE_SKIP_REASON = "low_visual_variance_negative"
 DEFAULT_VALIDATION_SAMPLE_SLUGS = frozenset(
-    {"turnhout", "retie", "westerlo", "arendonk_heide"}
+    {
+        "turnhout",
+        "retie",
+        "westerlo",
+        "arendonk_heide",
+        "vosselaar_center",
+        "grobbendonk_center",
+    }
 )
 DEFAULT_VALIDATION_SAMPLES = ",".join(sorted(DEFAULT_VALIDATION_SAMPLE_SLUGS))
 rasterio: Any = None
@@ -69,12 +76,20 @@ def parse_args() -> argparse.Namespace:
     )
     parser.add_argument("--tile-size", type=int, default=int(os.environ.get("OPERATOR_YOLO_TILE_SIZE", "256")))
     parser.add_argument("--stride", type=int, default=int(os.environ.get("OPERATOR_YOLO_TILE_STRIDE", "128")))
+    parser.add_argument(
+        "--samples",
+        default=os.environ.get("OPERATOR_YOLO_SAMPLES", ""),
+        help=(
+            "Optional comma/space separated manifest sample slugs to export. "
+            "An empty value keeps every manifest sample."
+        ),
+    )
     parser.add_argument(
         "--val-samples",
         default=os.environ.get("OPERATOR_YOLO_VAL_SAMPLES", DEFAULT_VALIDATION_SAMPLES),
         help=(
             "Comma/space separated sample slugs assigned to validation. "
-            "Defaults to the documented Turnhout, Retie, Westerlo and Arendonk-heide holdouts."
+            "Defaults to all documented validation samples."
         ),
     )
     parser.add_argument(
@@ -152,6 +167,31 @@ def split_slugs(raw: str) -> set[str]:
     return {value.strip().lower() for value in raw.replace(",", " ").split() if value.strip()}
 
 
+def select_manifest_samples(
+    samples: list[dict[str, Any]],
+    requested_slugs: set[str],
+) -> tuple[list[dict[str, Any]], list[str]]:
+    manifest_slugs = {
+        str(sample.get("sample_slug") or "").strip().lower()
+        for sample in samples
+        if str(sample.get("sample_slug") or "").strip()
+    }
+    if not requested_slugs:
+        return samples, []
+    unknown = requested_slugs - manifest_slugs
+    if unknown:
+        raise SystemExit(
+            "YOLO sample selection references unknown samples: " + ", ".join(sorted(unknown))
+        )
+    selected = [
+        sample
+        for sample in samples
+        if str(sample.get("sample_slug") or "").strip().lower() in requested_slugs
+    ]
+    excluded = sorted(manifest_slugs - requested_slugs)
+    return selected, excluded
+
+
 def validate_validation_split(samples: list[dict[str, Any]], val_slugs: set[str]) -> set[str]:
     sample_slugs = {
         str(sample.get("sample_slug") or "").strip().lower()
@@ -533,9 +573,13 @@ def main() -> int:
     ensure_yolo_directories(args.output_dir)
 
     manifest = json.loads(args.manifest_path.read_text(encoding="utf-8-sig"))
-    samples = manifest.get("samples") or []
-    if not samples:
+    manifest_samples = manifest.get("samples") or []
+    if not manifest_samples:
         raise SystemExit("Operator sample manifest contains no samples")
+    samples, excluded_sample_slugs = select_manifest_samples(
+        manifest_samples,
+        split_slugs(args.samples),
+    )
     val_slugs = validate_validation_split(samples, split_slugs(args.val_samples))
     exported_tiles: list[dict[str, Any]] = []
     for sample in samples:
@@ -584,7 +628,12 @@ def main() -> int:
         "min_label_visible_ratio": args.min_label_visible_ratio,
         "drop_low_variance_negatives": args.drop_low_variance_negatives,
         "blank_range_threshold": args.blank_range_threshold,
+        "source_manifest_sample_count": len(manifest_samples),
         "source_sample_count": len(samples),
+        "selected_sample_slugs": sorted(
+            str(sample.get("sample_slug") or "").strip().lower() for sample in samples
+        ),
+        "excluded_sample_slugs": excluded_sample_slugs,
         "validation_sample_slugs": sorted(val_slugs),
         **validation_coverage,
         "tile_count": len(kept_tiles),

scripts/prepare_operator_real_data_samples.py

@@ -28,8 +28,20 @@ SPARSE_BACKGROUND_CATEGORY = "sparse_building_context"
 TRAINING_EXPANSION_SAMPLE_SLUGS = frozenset(
     {"olen_center", "lille_center", "oud_turnhout_center", "kasterlee_center"}
 )
+SMALL_BUILDING_TRAINING_SAMPLE_SLUGS = frozenset(
+    {"beerse_center", "rijkevorsel_center", "hoogstraten_center", "vorselaar_center"}
+)
+SMALL_BUILDING_VALIDATION_SAMPLE_SLUGS = frozenset(
+    {"vosselaar_center", "grobbendonk_center"}
+)
 DEFAULT_VALIDATION_SAMPLE_SLUGS = frozenset(
-    {"turnhout", "retie", "westerlo", "arendonk_heide"}
+    {
+        "turnhout",
+        "retie",
+        "westerlo",
+        "arendonk_heide",
+        *SMALL_BUILDING_VALIDATION_SAMPLE_SLUGS,
+    }
 )
 requests: Any = None
 rasterio: Any = None
@@ -122,6 +134,42 @@ SAMPLES: dict[str, OperatorSample] = {
         center_lon=4.9678120,
         center_lat=51.2407915,
     ),
+    "beerse_center": OperatorSample(
+        slug="beerse_center",
+        display_name="Beerse center small-building training expansion",
+        center_lon=4.8534,
+        center_lat=51.3192,
+    ),
+    "rijkevorsel_center": OperatorSample(
+        slug="rijkevorsel_center",
+        display_name="Rijkevorsel center small-building training expansion",
+        center_lon=4.7604,
+        center_lat=51.3487,
+    ),
+    "hoogstraten_center": OperatorSample(
+        slug="hoogstraten_center",
+        display_name="Hoogstraten center small-building training expansion",
+        center_lon=4.7609,
+        center_lat=51.4002,
+    ),
+    "vorselaar_center": OperatorSample(
+        slug="vorselaar_center",
+        display_name="Vorselaar center small-building training expansion",
+        center_lon=4.7731,
+        center_lat=51.2020,
+    ),
+    "vosselaar_center": OperatorSample(
+        slug="vosselaar_center",
+        display_name="Vosselaar center small-building validation",
+        center_lon=4.8899,
+        center_lat=51.3095,
+    ),
+    "grobbendonk_center": OperatorSample(
+        slug="grobbendonk_center",
+        display_name="Grobbendonk center small-building validation",
+        center_lon=4.7358,
+        center_lat=51.1907,
+    ),
     "postel_bos": OperatorSample(
         slug="postel_bos",
         display_name="Postel forest background candidate",