feat: add coverage-aware Mol benchmark

2026-07-14 12:27:48 +02:00
parent 4f0d2891dd
commit 9f61037794
12 changed files with 682 additions and 10 deletions
@@ -7659,3 +7659,42 @@ Open:
  from envelope diagnostics. The next safe model step is an evidence-led
  footprint-label/matching review, followed by a fresh bounded Mol multi-zone
  benchmark before any activation or retraining decision.
+
+# Sprint 185 - Coverage-aware Mol operational benchmark
+
+## Implementation
+
+- Extended the existing real-data quality matrix summaries with the exact
+  persisted tile-coverage population, raw/evaluated/excluded/clipped candidate
+  and reference counts, CRS/tile provenance and diagnostic-only
+  box-to-footprint match gap.
+- Preserved manifest-backed Mol municipality, operational-zone, validation
+  split, WGS84 bounds and source reference counts in the multi-sample summary.
+- Added `build_mol_operational_benchmark_report.py`. It groups evidence by the
+  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%
+  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.
+- Wired the report into the persistent Mol operator runner, readiness gate and
+  all-in-one image. The report does not activate, replace, download or train a
+  model and does not change APIs, migrations or inference behavior.
+
+## Local validation
+
+- Focused Sprint 126/127/178/184/185 regression set passed: `14` tests.
+- Full readiness passed with `512` backend tests, one Alembic head
+  `202606120900`, 81 audited API operations, frontend typecheck and Vite 7.3.6
+  production build.
+- `npm audit --audit-level=high` reported zero vulnerabilities.
+- Shell syntax checks passed for the single-, multi-sample and Mol operational
+  runners. Fixture tests prove both an accepted four-zone result and explicit
+  rejection for missing coverage, one-zone collapse and background pressure.
+
+## Next pass
+
+- 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.