# Skill: QA/QC Review

Use this skill for comparing predicted geospatial outputs against reference layers such as GRB buildings.

## Read first

- `docs/QA_QC_SPECIFICATION.md`
- `docs/QA_QC_ENGINE.md`
- `docs/qaqc/ANTI_REGRESSION_RULES.md`
- `docs/API_EXAMPLE_RESPONSES.md`

## QA/QC rules

- Reference data and predicted data must remain separate.
- Matching thresholds must be explicit.
- False positives and false negatives must be exportable.
- Metrics must be reproducible from stored inputs.
- Do not hide low-confidence results; classify or flag them.

## Core metrics

- IoU.
- Precision.
- Recall.
- F1.
- Matched count.
- Unmatched reference count.
- Unmatched prediction count.

## Implementation checklist

- [ ] Define matching method.
- [ ] Implement geometry overlap calculation.
- [ ] Calculate metrics.
- [ ] Persist or return QA results.
- [ ] Create GeoJSON outputs for matched/unmatched classes.
- [ ] Add fixture-based tests.

## Completion report fields

- Matching method.
- Thresholds.
- Metrics returned.
- Fixtures used.
- Known QA limitations.