Gate filtered YOLO candidate and harden provenance

docs/CODEX_EXECUTION_LOG.md

@@ -1,3 +1,32 @@
+## Sprint 169 Filtered YOLO candidate gate and operator hardening (2026-07-12)
+
+Changed:
+- Trained inactive local model asset `geointel-building-yolov8s-aoi1024cleanpx12vis035lowvar512e50-pt` from the visually audited AOI1024 dataset after low-variance negative filtering.
+- Added SHA256 provenance fields to future `training_summary.json` output for `dataset.yaml`, the dataset summary, the local base model and the copied trained model.
+- Kept long workbench context names compact with matching native tooltips and clamped readiness values to two lines.
+- Archived the training, positive-AOI, split-background and promotion evidence under `artifacts/model-review/aoi1024cleanpx12vis035lowvar512e50` locally and matching Tower artifact directories.
+
+Runtime evidence:
+- Training completed for 50 CPU epochs with YOLOv8s, image size `512`, batch `4`; best-model validation ended at precision `0.454`, recall `0.491`, mAP50 `0.368` and mAP50-95 `0.149`.
+- Model catalog SHA256: `75345767b51cc66692a9d2c2cd6577b7feecaf5971c8762365e8b610b9dfde8e`; catalog status `available`, `active=false`, `will_download_models=false`.
+- Local model-load preflight passed with Torch `2.13.0` and Ultralytics `8.4.92`; CUDA is unavailable and no model download occurred.
+- Seven-AOI persisted QA matrix produced 28 runs. Best single result was Westerlo at threshold `0.15`, F1 `0.3002114164904862`.
+- Mean positive F1 was `0.153943` at threshold `0.05`, `0.153872` at `0.15`, `0.128891` at `0.25` and `0.099011` at `0.35`.
+- Strict pure-empty background evidence was clean at thresholds `0.15`, `0.25` and `0.35`; threshold `0.05` produced one Postel-bos detection.
+- Split-aware promotion report recommended `none`: every threshold failed the positive mean-F1 gate, and `0.05` also failed background false-positive pressure.
+- The existing active `geointel-building-yolov8s-aoi1024bg512r3e50-pt` remains materially stronger at its promoted `0.35` profile with mean F1 `0.320866`; no model default or `.env` value was changed.
+
+Tested:
+- Red step: `python -m pytest backend/tests/test_sprint169_long_context_name_readability.py -q` failed before context tooltips and readiness clamping were present.
+- `python -m pytest backend/tests/test_sprint169_long_context_name_readability.py backend/tests/test_sprint22_workbench_status_strip.py backend/tests/test_sprint161_widescreen_workbench.py -q` (`5 passed`).
+- Red step: `python -m pytest backend/tests/test_sprint129_operator_yolo_training_dataset.py::test_operator_yolo_train_smoke_script_contract -q` failed before training hashes were recorded.
+- `python -m pytest backend/tests/test_sprint129_operator_yolo_training_dataset.py -q` (`3 passed`).
+- `bash -n scripts/train_operator_yolo_detector.sh` and frontend typecheck passed.
+- Full readiness: `bash scripts/run_readiness_check.sh` (`461 passed`; one Alembic head; frontend typecheck/build and shell syntax gates passed).
+
+Next:
+- Do not retrain the same architecture blindly. Inspect per-AOI false-negative evidence and improve label geometry/class balance or add targeted positive samples for the weakest AOIs before the next candidate.
+
 ## Sprint 164 Tower AI deploy env hardening (2026-07-11)
 
 Changed:

docs/TODO.md

@@ -133,7 +133,9 @@ This file now starts with the current implementation status. Older preparation/b
 - [x] Add deterministic visual YOLO label QA contact sheets before spending more CPU on another training run.
 - [x] Filter no-data/low-variance pure-empty negative tiles from operator YOLO exports before the next training run.
 - [x] Regenerate the AOI1024 cleanpx YOLO dataset with low-variance negative filtering and rerun visual contact-sheet QA before training.
-- [ ] Train one inactive candidate from the filtered AOI1024 cleanpx YOLO dataset and gate it through the positive-AOI plus split-background promotion workflow.
+- [x] Train one inactive candidate from the filtered AOI1024 cleanpx YOLO dataset and gate it through the positive-AOI plus split-background promotion workflow; reject it because mean positive F1 remains below gate.
+- [x] Add deterministic dataset/base/trained-model SHA256 provenance to future operator training summaries.
+- [ ] Review per-AOI false-negative evidence for the weakest AOIs and improve positive sample/label geometry coverage before another training candidate.
 - [ ] Apply promoted V1 default building detector only after explicit operator review of the emitted `.env` updates, followed by rebuild/restart and browser/runtime smoke.
 
 ## Sprint 8 status