docs: record Mol benchmark decision

docs/CODEX_EXECUTION_LOG.md

@@ -7674,7 +7674,7 @@ Open:
   exact model asset, tile size, overlap and confidence threshold and refuses to
   blend different candidate configurations.
 - Added explicit gates for four independent positive Mol holdouts, one
-  pure-empty background control, complete coverage provenance, minimum 95%
+  pure-empty background control, complete coverage provenance, minimum 90%
   reference coverage, mean F1 `0.25`, minimum per-zone F1 `0.10` and zero
   background detections. Envelope matches remain diagnostic and cannot satisfy
   the canonical F1 gates.
@@ -7698,3 +7698,44 @@ Open:
 - Deploy the operator tooling to Tower and run the current active local model
   over Achterbos, Gompel, Donk, Postel and the Postel-bos pure-empty control.
   Record the coverage-aware operational decision before changing model state.
+
+## Tower benchmark evidence
+
+- Deployed commit `9f61037` to the healthy all-in-one runtime and completed the
+  real benchmark with active asset
+  `geointel-building-yolov8s-smallbld-minpx3-img640-ft30-pt`, tile `512`,
+  overlap `64`, confidence `0.15` and canonical QA IoU `0.25`.
+- Canonical results by holdout were: Achterbos F1 `0.6694`, Gompel `0.6564`,
+  Donk `0.5894` and Postel `0.4749`. Macro precision/recall/F1 were
+  `0.6296` / `0.5726` / `0.5975`; micro precision/recall/F1 were `0.6494` /
+  `0.6189` / `0.6338`.
+- Across all positive zones, 3,624 of 3,829 raw reference features intersected
+  exact tile coverage. The runner excluded 205 outside features, clipped 60 at
+  boundaries and kept 101 additional envelope matches diagnostic-only.
+- Postel-bos produced zero detections. All model, per-zone, coverage provenance,
+  diagnostic-separation and pure-empty-background gates passed. The minimum
+  coverage ratio was `0.9051`; the integrity gate was set to `0.90` because
+  OGC bbox retrieval intentionally retains complete edge geometries outside the
+  projected raster polygon.
+- The final report under
+  `/app/storage/operator-evidence/mol-operational-validation/coverage-aware-20260714T1000Z`
+  is `accepted` with recommendation `retain_or_promote_candidate`. No model
+  state or weights were changed by the report.
+- Exported a four-zone portfolio with 7,078 persisted QA evidence features.
+  Audits found 1,381 canonical false negatives and 1,211 false positives.
+  False-negative medians were about 24-28 m2 and tiny/small buildings dominate;
+  Postel and Donk are the weakest zones.
+- Rendered and visually inspected 48 false-negative plus 48 false-positive
+  review cards across every holdout. No source tile was missing. The cards show
+  a mixture of genuine small-building/model errors, temporal GRB/orthophoto
+  differences and box-versus-footprint matching artifacts, so raw FP/FN counts
+  must not be converted directly into new labels.
+
+## Decision
+
+- Retain the active model for bounded operator-assisted Mol workflows and demos
+  with QA/QC visible. It is not ground truth and should not be auto-exported as
+  authoritative buildings. Do not start another blind training run; first
+  complete the generated manual review decisions and curate confirmed small-
+  building positives plus confirmed visual hard negatives from the weakest
+  zones.

docs/TODO.md

@@ -28,7 +28,8 @@ This file now starts with the current implementation status. Older preparation/b
 - [x] Visually review Mol Postel and Donk false-positive/false-negative evidence, classify the dominant error modes and only then decide whether another model-training pass is justified.
 - [x] Clip detection QA populations to persisted raster/tile coverage and add box-to-footprint matching diagnostics before reconsidering model training.
 - [x] Add a coverage-aware Mol multi-zone benchmark report with explicit positive-zone, per-zone collapse, reference-coverage and pure-empty background gates.
-- [ ] Execute the refreshed coverage-aware Mol operational benchmark against the active local model and record the resulting retain/review decision.
+- [x] Execute the refreshed coverage-aware Mol operational benchmark against the active local model and record the resulting retain/review decision.
+- [ ] Complete manual decisions for the generated 48 false-negative and 48 false-positive review cards before constructing any new training corpus.
 - [x] Backend FastAPI foundation, health endpoint and service structure.
 - [x] React/TypeScript frontend foundation and MapLibre workbench.
 - [x] Map layer visibility, opacity and feature property inspection.