Document multi-sample detection quality baseline

CHANGELOG.md

@@ -13,6 +13,7 @@
 - Added `scripts/run_multi_sample_detection_quality_matrix.sh` to run the existing real-data quality matrix for every prepared sample and combine the results.
 - The combined summary writes `multi_sample_quality_summary.json` with overall score/recall/precision rankings and per-sample best configurations.
 - Added readiness coverage and regression tests for the sample-preparation and multi-sample matrix contracts.
+- Ran the full 24-run Tower matrix for Geel, Mol and Turnhout. Best overall score/recall was Turnhout with `yolov8n-building-segmentation-pt`, tile `512`, overlap `64`, threshold `0.15`, 142 detections, 18 matches, 124 false positives, 755 false negatives and F1 `0.03934426229508197`; generic `yolov8n-pt` produced zero building detections across all samples.
 
 ## Sprint 126 Detection quality matrix tooling (2026-07-07)
 

docs/CODEX_EXECUTION_LOG.md

@@ -21,16 +21,37 @@ Tested:
 - `python scripts\smoke_docs.py` passed.
 - `git diff --check` passed.
 - `bash scripts/run_readiness_check.sh` passed: 393 backend tests, frontend typecheck/build, Alembic head `202606120900`, live smoke syntax checks and the new sample/multi-sample checks.
+- Pushed commit `06dfc5f` to Gitea and fast-forwarded Tower `/mnt/user/appdata/geointel` to the same commit.
+- Tower sample preparation passed by running the new helper in the live `geointel` all-in-one container through stdin:
+  - Geel reused existing runtime files: 617 GRB GBG reference features.
+  - Mol was newly prepared: 374 GRB GBG reference features.
+  - Turnhout was newly prepared: 773 GRB GBG reference features.
+  - Manifest: `/mnt/user/appdata/geointel/storage/operator-data/operator_samples_manifest.json`.
+- Tower multi-sample quality matrix passed:
+  - command: `OPERATOR_SAMPLE_MANIFEST_PATH=storage/operator-data/operator_samples_manifest.json QUALITY_MODEL_ASSET_IDS="yolov8n-building-segmentation-pt yolov8n-pt" QUALITY_TILE_SIZES="512 640" QUALITY_TILE_OVERLAPS="64" QUALITY_THRESHOLDS="0.50 0.15" bash scripts/run_multi_sample_detection_quality_matrix.sh http://192.168.10.150:1202`
+  - output: `/mnt/user/appdata/geointel/artifacts/detection-quality-matrix/multi-sample/20260707T025303Z/multi_sample_quality_summary.json`
+  - run count: 24 real persisted workflows across 3 samples.
+  - best overall score/recall: Turnhout, `yolov8n-building-segmentation-pt`, tile `512`, overlap `64`, threshold `0.15`, 142 detections, 18 matches, 124 false positives, 755 false negatives, precision `0.1267605633802817`, recall `0.02328589909443726`, F1 `0.03934426229508197`.
+  - best overall precision: Geel, `yolov8n-building-segmentation-pt`, tile `640`, overlap `64`, threshold `0.50`, 4 detections, 1 match, 3 false positives, 616 false negatives, precision `0.25`, recall `0.0016207455429497568`, F1 `0.0032206119162640897`.
+  - best Geel score: tile `512`, threshold `0.15`, 80 detections, 6 matches, 74 false positives, 611 false negatives, F1 `0.017216642754662843`.
+  - best Mol score: tile `640`, threshold `0.15`, 77 detections, 7 matches, 70 false positives, 367 false negatives, F1 `0.03104212860310421`.
+  - best Turnhout score: tile `512`, threshold `0.15`, 142 detections, 18 matches, 124 false positives, 755 false negatives, F1 `0.03934426229508197`.
+  - generic `yolov8n-pt` produced zero building detections across every sample, tile size and threshold.
+- Tower best-score evidence export passed for every sample:
+  - Geel: `/mnt/user/appdata/geointel/artifacts/detection-quality-matrix/multi-sample/20260707T025303Z/geel/calibration_evidence_review.html`, 697 features, 611 false negatives, 74 false positives, 6 matched detections and 6 matched references.
+  - Mol: `/mnt/user/appdata/geointel/artifacts/detection-quality-matrix/multi-sample/20260707T025303Z/mol/calibration_evidence_review.html`, 451 features, 367 false negatives, 70 false positives, 7 matched detections and 7 matched references.
+  - Turnhout: `/mnt/user/appdata/geointel/artifacts/detection-quality-matrix/multi-sample/20260707T025303Z/turnhout/calibration_evidence_review.html`, 915 features, 755 false negatives, 124 false positives, 18 matched detections and 18 matched references.
 
 Open:
-- Live Tower sample preparation and multi-sample matrix run still needed.
+- None for multi-sample tooling.
 
 Limitations:
 - This is operator tooling only. It does not add a live GRB provider, live orthophoto provider, application endpoint, migration, frontend feature, model download or fixture inference path.
 - The prepared sample files are runtime artifacts under appdata/storage and remain excluded from Git.
+- The three-sample benchmark confirms the current building-segmentation evaluation model is not extraction-quality for V1: best recall is only about 2.3% and false negatives dominate every sample.
 
 Next recommended pass:
-- Run the sample-preparation helper in the Tower all-in-one container, then run the multi-sample matrix from the Tower checkout and document the combined quality baseline.
+- Replace or add a stronger aerial/building model candidate and rerun the same multi-sample matrix; optionally add IoU-threshold sweeps after a model produces materially better candidate detections.
 
 ## Sprint 126 Detection quality matrix tooling (2026-07-07)
 

docs/TODO.md

@@ -99,6 +99,7 @@ This file now starts with the current implementation status. Older preparation/b
 - [x] Add calibration QA evidence export tooling for false-positive/false-negative inspection artifacts.
 - [x] Add real-data detection quality matrix tooling for model/tile/threshold comparison.
 - [x] Add reproducible Geel/Mol/Turnhout operator sample preparation and multi-sample quality matrix tooling.
+- [x] Run first Geel/Mol/Turnhout persisted detection quality baseline.
 - [ ] Calibrate confidence, IoU and model selection against persisted Geel/Mol/Turnhout detections and any additional local orthophoto/reference samples.
 
 ## Sprint 8 status