Record YOLOv8s e60 candidate evaluation

docs/CODEX_EXECUTION_LOG.md

@@ -1,3 +1,44 @@
+## Sprint 145 YOLOv8s hardneg r8 e60 full candidate evaluation (2026-07-08)
+
+Changed:
+- Verified that the Tower-local YOLOv8s hardneg r8 training run completed all 60 requested CPU epochs.
+- Finalized the completed model artifact:
+  - `/app/models/geointel-building-yolov8s-hardneg160r8e60.pt`
+  - model asset id `geointel-building-yolov8s-hardneg160r8e60-pt`
+  - SHA256 `d7daea04bd51a54a06944f0d4bf1961fe453dceb1fb34ef51daa73f6901fca81`
+- Wrote `/app/storage/training/operator-yolo/geointel-building-yolov8s-hardneg160r8e60/training_summary.json`.
+
+Tested:
+- Training summary reports `status=ok`, requested epochs `60`, completed epochs `60`.
+- Ran 7-AOI positive matrix:
+  - output `/mnt/user/appdata/geointel/artifacts/detection-quality-matrix/multi-sample/yolov8s-hardneg160r8e60-positive-20260708/multi_sample_quality_summary.json`
+  - sample count `7`
+  - run count `21`
+  - best result: Westerlo threshold `0.15`, F1/score `0.20588235294117646`, precision `0.26666666666666666`, recall `0.16766467065868262`, detections `210`, false positives `154`, false negatives `278`.
+- Ran hard-negative matrix:
+  - output `/mnt/user/appdata/geointel/artifacts/detection-hard-negatives/yolov8s-hardneg160r8e60-live/hard_negative_matrix_summary.json`
+  - Postel-bos detections `0/0/0` at thresholds `0.05/0.15/0.25`
+  - Lommel-heide detections `0/0/0`
+  - Kasterlee-bos detections `30/11/9`
+- Assembled evidence portfolio:
+  - output `/mnt/user/appdata/geointel/artifacts/detection-calibration-portfolio/yolov8s-hardneg160r8e60-positive-20260708/output/calibration_evidence_portfolio.json`
+  - sample count `7`
+  - evidence features `15714`
+  - role counts `false_negative=10468`, `false_positive=4198`, `match_candidate=524`, `match_reference=524`
+- Ran promotion report:
+  - output `/mnt/user/appdata/geointel/artifacts/detection-model-promotion/yolov8s-hardneg160r8e60-20260708/detection_model_promotion_report.json`
+  - evaluated `3` candidate thresholds
+  - recommended candidate `none`
+  - threshold `0.05`: mean F1 `0.07825931710275633`, max background detections `30`, rejected for `positive_mean_f1_below_gate` and `background_false_positive_pressure`
+  - threshold `0.15`: mean F1 `0.07870592446136859`, max background detections `11`, rejected for the same reasons
+  - threshold `0.25`: mean F1 `0.06111741186503092`, max background detections `9`, rejected for the same reasons
+- Comparison baseline: `geointel-building-yolov8n-expanded160e50-pt|640|64|0.15` had mean F1 `0.19718007234060198` and max background detections `46`.
+
+Open:
+- The e60 YOLOv8s hardneg r8 candidate improves hard-negative behavior at threshold `0.15` compared with `expanded160e50`, but its positive-AOI mean F1 is too low for operational use.
+- Do not activate this model as default.
+- Next model pass should focus on improving positive recall/fit without losing the hard-negative gains, likely through better labels, more positive AOIs, or a different architecture/training strategy rather than simply extending this same run.
+
 ## Sprint 144 YOLOv8s hardneg r8 partial candidate evaluation (2026-07-08)
 
 Changed:

docs/TODO.md

@@ -432,6 +432,7 @@ This file now starts with the current implementation status. Older preparation/b
 - [x] Prevent same-threshold calibration evidence responses from overwriting each other in multi-model portfolios.
 - [x] Add a model promotion decision report that combines positive-AOI score with hard-negative false-positive pressure.
 - [x] Train and reject a YOLOv8s hard-negative r8 partial candidate after 12 CPU epochs through the full positive/background promotion gate.
+- [x] Finish and reject the full YOLOv8s hard-negative r8 e60 candidate through the same promotion gate.
 - [ ] Add more AOIs after the tile-level baseline so the next local model attempt is not limited to Geel/Mol/Turnhout.
 - [ ] Add negative/background AOIs so the next tile dataset is not all positive tiles.
-- [ ] Finish or resume a long-running higher-capacity YOLO training pass outside the 1-hour interactive command window, then rerun the same promotion gate.
+- [ ] Improve positive training coverage/label quality before the next higher-capacity model attempt; simply extending the same hardneg r8 run is not enough.