Document calibration evidence export results
GeoIntel CI / docs-smoke (push) Has been cancelled
GeoIntel CI / contract-smoke (push) Has been cancelled

This commit is contained in:
Codex
2026-07-07 02:35:23 +02:00
parent ea8dcb2bdd
commit e1586727d2
2 changed files with 19 additions and 2 deletions
+18 -2
View File
@@ -17,16 +17,32 @@ Validation:
- `bash -n scripts/export_detection_calibration_evidence.sh` passed.
- `bash scripts/export_detection_calibration_evidence.sh --help` passed.
- Missing-input guard printed usage and did not fetch evidence.
- `bash scripts/run_readiness_check.sh` passed: 389 backend tests, Alembic head check, frontend typecheck/build and shell syntax checks.
- Tower pulled commit `ea8dcb2` with `git pull --ff-only origin main`.
- Tower evidence export passed against `/mnt/user/appdata/geointel/artifacts/detection-calibration/20260707T002103Z/calibration_summary.json`.
- Evidence artifacts written on Tower:
- `/mnt/user/appdata/geointel/artifacts/detection-calibration/20260707T002103Z/calibration_evidence.geojson`
- `/mnt/user/appdata/geointel/artifacts/detection-calibration/20260707T002103Z/calibration_evidence_summary.json`
- `/mnt/user/appdata/geointel/artifacts/detection-calibration/20260707T002103Z/calibration_evidence_review.html`
- Evidence counts from the Geel calibration export:
- total evidence features: 2555
- false negatives: 2460
- false positives: 79
- matched detections: 8
- matched references: 8
- Copied `calibration_evidence_review.html` and `calibration_evidence_summary.json` to the local ignored `artifacts/detection-calibration/20260707T002103Z/` directory for inspection.
- Internal browser validation passed through a temporary local static server: the review page loaded with title `GeoIntel Detection Calibration Evidence`, 1 SVG, 2555 SVG paths, 4 calibration table rows and 0 console errors.
Open:
- Pull the script to Tower and export evidence for the latest Geel calibration summary.
- None for evidence export tooling.
Limitations:
- This is operator tooling only. It does not rerun inference, change API behavior, add UI behavior, fetch providers, seed demo data or download models.
- The SVG review is a lightweight geometry overview, not a replacement for full MapLibre evidence review in the workbench.
- The evidence distribution confirms the current active model misses most reference buildings on this sample. This points to model suitability and/or tiling strategy as the next bottleneck, not runtime plumbing.
Next recommended pass:
- Run the evidence export on Tower, inspect the role distribution and use the artifacts to decide whether the current model is worth further threshold tuning or should be replaced.
- Add a model-quality decision pass: compare the current evaluation model against another building/aerial model or adjusted tile/overlap settings on at least two additional local orthophoto/reference samples.
## Sprint 124 Detection calibration sweep tooling (2026-07-07)