feat: close measured model review evidence

scripts/README.md

@@ -964,6 +964,23 @@ geometry and green is a matched reference. Selection is deterministic and
 stratified by AOI and geodetic area bucket. Every CSV decision starts as
 `unreviewed`; no positive-training example is inferred.
 
+Validate the completed false-negative decisions symmetrically with the
+false-positive workflow:
+
+```bash
+docker exec geointel /opt/geointel/venv/bin/python \
+  /app/scripts/validate_detection_false_negative_review_decisions.py \
+  --review-summary /app/storage/operator-data/model-review/small-building-candidate/false-negative-visual-review/detection_false_negative_review_summary.json \
+  --decisions-csv /app/storage/operator-data/model-review/small-building-candidate/false-negative-visual-review/false_negative_review_decisions.csv \
+  --output-dir /app/storage/operator-data/model-review/small-building-candidate/false-negative-visual-review/validated \
+  --require-complete
+```
+
+The validator exits with code `2` while any row remains `unreviewed`. It emits
+only explicit `confirmed_model_false_negative` geometries; alignment,
+reference-gap and uncertain imagery decisions never become positive training
+labels.
+
 Reference features that do not intersect any persisted inference tile are not
 silently counted as reviewable model misses. They are reported separately in
 `false_negatives_outside_tile_coverage.geojson` with

scripts/run_readiness_check.sh

@@ -64,6 +64,7 @@ ${PYTHON_BIN} -m py_compile scripts/audit_detection_false_positive_evidence.py
 ${PYTHON_BIN} -m py_compile scripts/render_detection_false_positive_review_contact_sheets.py
 ${PYTHON_BIN} -m py_compile scripts/render_detection_false_negative_review_contact_sheets.py
 ${PYTHON_BIN} -m py_compile scripts/validate_detection_false_positive_review_decisions.py
+${PYTHON_BIN} -m py_compile scripts/validate_detection_false_negative_review_decisions.py
 ${PYTHON_BIN} -m py_compile scripts/activate_promoted_yolo_candidate.py
 ${PYTHON_BIN} -m py_compile scripts/cleanup_demo_artifacts.py
 ${PYTHON_BIN} -m py_compile backend/scripts/cleanup_demo_artifacts.py

scripts/validate_detection_false_negative_review_decisions.py

@@ -0,0 +1,208 @@
+#!/usr/bin/env python3
+"""Validate explicit operator decisions for detection QA false-negatives."""
+
+from __future__ import annotations
+
+import argparse
+import csv
+import json
+from datetime import datetime, timezone
+from pathlib import Path
+from typing import Any
+
+
+JSON_NAME = "detection_false_negative_review_validation.json"
+MARKDOWN_NAME = "detection_false_negative_review_validation.md"
+CONFIRMED_NAME = "confirmed_model_false_negatives.geojson"
+DECISIONS = (
+    "confirmed_model_false_negative",
+    "qa_alignment_mismatch",
+    "reference_gap_or_change",
+    "imagery_obscured_or_uncertain",
+    "unreviewed",
+)
+
+
+def parse_args() -> argparse.Namespace:
+    parser = argparse.ArgumentParser(
+        description="Validate manual false-negative review decisions without inferring labels."
+    )
+    parser.add_argument("--review-summary", required=True, type=Path)
+    parser.add_argument("--decisions-csv", required=True, type=Path)
+    parser.add_argument("--output-dir", required=True, type=Path)
+    parser.add_argument(
+        "--require-complete",
+        action="store_true",
+        help="Return a non-zero exit code while any selected record remains unreviewed.",
+    )
+    return parser.parse_args()
+
+
+def load_json(path: Path) -> dict[str, Any]:
+    if not path.is_file():
+        raise SystemExit(f"Review summary is not readable: {path}")
+    payload = json.loads(path.read_text(encoding="utf-8-sig"))
+    if not isinstance(payload, dict):
+        raise SystemExit(f"Review summary must be a JSON object: {path}")
+    return payload
+
+
+def load_decisions(path: Path) -> list[dict[str, str]]:
+    if not path.is_file():
+        raise SystemExit(f"Review decisions CSV is not readable: {path}")
+    with path.open(newline="", encoding="utf-8-sig") as handle:
+        reader = csv.DictReader(handle)
+        required = {"reference_feature_id", "review_decision", "review_notes"}
+        missing = required - set(reader.fieldnames or [])
+        if missing:
+            raise SystemExit(
+                "Review decisions CSV lacks required columns: "
+                + ", ".join(sorted(missing))
+            )
+        return [dict(row) for row in reader]
+
+
+def validate(
+    summary: dict[str, Any], rows: list[dict[str, str]]
+) -> tuple[dict[str, Any], dict[str, Any]]:
+    features = summary.get("selected_features") or []
+    if not isinstance(features, list) or not all(
+        isinstance(item, dict) for item in features
+    ):
+        raise SystemExit("Review summary selected_features must be a list of objects")
+
+    expected: dict[str, dict[str, Any]] = {}
+    for feature in features:
+        reference_id = str(feature.get("reference_feature_id") or "").strip()
+        if not reference_id or reference_id in expected:
+            raise SystemExit(
+                "Review summary contains a missing or duplicate reference_feature_id"
+            )
+        expected[reference_id] = feature
+
+    provided: dict[str, dict[str, str]] = {}
+    for row in rows:
+        reference_id = str(row.get("reference_feature_id") or "").strip()
+        if not reference_id or reference_id in provided:
+            raise SystemExit(
+                "Review decisions contain a missing or duplicate reference_feature_id"
+            )
+        decision = str(row.get("review_decision") or "").strip()
+        if decision not in DECISIONS:
+            raise SystemExit(
+                f"Invalid review decision for {reference_id}: {decision}. "
+                + "Allowed values: "
+                + ", ".join(DECISIONS)
+            )
+        provided[reference_id] = row
+
+    missing_ids = set(expected) - set(provided)
+    extra_ids = set(provided) - set(expected)
+    if missing_ids or extra_ids:
+        details = []
+        if missing_ids:
+            details.append("missing: " + ", ".join(sorted(missing_ids)))
+        if extra_ids:
+            details.append("unexpected: " + ", ".join(sorted(extra_ids)))
+        raise SystemExit(
+            "Review decisions do not match the selected evidence ("
+            + "; ".join(details)
+            + ")"
+        )
+
+    counts = {decision: 0 for decision in DECISIONS}
+    confirmed_features: list[dict[str, Any]] = []
+    for reference_id, feature in expected.items():
+        row = provided[reference_id]
+        decision = row["review_decision"].strip()
+        counts[decision] += 1
+        if decision != "confirmed_model_false_negative":
+            continue
+        source_properties = feature.get("properties") or {}
+        properties = (
+            dict(source_properties) if isinstance(source_properties, dict) else {}
+        )
+        properties.update(
+            {
+                "reference_feature_id": reference_id,
+                "sample_slug": feature.get("sample_slug"),
+                "area_m2": feature.get("area_m2"),
+                "area_bucket": feature.get("area_bucket"),
+                "source_tile_path": feature.get("source_tile_path"),
+                "review_decision": decision,
+                "review_notes": row.get("review_notes", "").strip(),
+            }
+        )
+        confirmed_features.append(
+            {
+                "type": "Feature",
+                "id": f"confirmed_model_false_negative:{reference_id}",
+                "geometry": feature.get("geometry"),
+                "properties": properties,
+            }
+        )
+
+    status = "complete" if counts["unreviewed"] == 0 else "review_required"
+    report = {
+        "generated_at": datetime.now(timezone.utc).isoformat(),
+        "schema_version": 1,
+        "status": status,
+        "review_summary_path": str(summary.get("portfolio_path") or ""),
+        "selected_feature_count": len(features),
+        "decision_counts": counts,
+        "confirmed_model_false_negative_count": len(confirmed_features),
+        "safety_rule": (
+            "Only explicit confirmed_model_false_negative decisions are exported; "
+            "QA false-negatives are never inferred as model training labels."
+        ),
+    }
+    return report, {"type": "FeatureCollection", "features": confirmed_features}
+
+
+def write_markdown(report: dict[str, Any], output_dir: Path) -> None:
+    lines = [
+        "# Detection false-negative review validation",
+        "",
+        f"- Status: `{report['status']}`",
+        f"- Selected records: {report['selected_feature_count']}",
+        f"- Confirmed model false-negatives: {report['confirmed_model_false_negative_count']}",
+        "",
+        "## Decision counts",
+        "",
+    ]
+    lines.extend(
+        f"- `{decision}`: {count}"
+        for decision, count in report["decision_counts"].items()
+    )
+    lines.extend(["", f"> {report['safety_rule']}", ""])
+    (output_dir / MARKDOWN_NAME).write_text("\n".join(lines), encoding="utf-8")
+
+
+def main() -> int:
+    args = parse_args()
+    summary_path = args.review_summary.expanduser().resolve()
+    decisions_path = args.decisions_csv.expanduser().resolve()
+    output_dir = args.output_dir.expanduser().resolve()
+    output_dir.mkdir(parents=True, exist_ok=True)
+
+    report, confirmed = validate(load_json(summary_path), load_decisions(decisions_path))
+    report["review_summary_path"] = str(summary_path)
+    report["decisions_csv_path"] = str(decisions_path)
+    (output_dir / JSON_NAME).write_text(
+        json.dumps(report, indent=2, sort_keys=True), encoding="utf-8"
+    )
+    (output_dir / CONFIRMED_NAME).write_text(
+        json.dumps(confirmed, indent=2, sort_keys=True), encoding="utf-8"
+    )
+    write_markdown(report, output_dir)
+    print(f"False-negative review status: {report['status']}")
+    print(f"Validation: {output_dir / JSON_NAME}")
+    print(f"Confirmed evidence: {output_dir / CONFIRMED_NAME}")
+    if args.require_complete and report["status"] != "complete":
+        print("Review remains incomplete", flush=True)
+        return 2
+    return 0
+
+
+if __name__ == "__main__":
+    raise SystemExit(main())