From c2d7dd48a85721ee5de9fd2a1126fcf8200bd9de Mon Sep 17 00:00:00 2001 From: Codex Date: Wed, 15 Jul 2026 03:40:59 +0200 Subject: [PATCH] docs: record live model review proof --- CHANGELOG.md | 3 +++ docs/CODEX_EXECUTION_LOG.md | 16 ++++++++++++++++ 2 files changed, 19 insertions(+) diff --git a/CHANGELOG.md b/CHANGELOG.md index be06e524..69cee555 100644 --- a/CHANGELOG.md +++ b/CHANGELOG.md @@ -24,6 +24,9 @@ trained, downloaded, activated or reconfigured. - Clarified map quality output with strict match counts and the existing diagnostic reference-envelope match, without changing canonical QA metrics. +- Live Mol proof measured 282 candidates, 209 strict matches, precision 74.1%, + recall 68.5% and F1 71.2%; 212 diagnostic envelope matches made the remaining + three possible box/footprint differences explicit. ## Sprint 197 Measured detection accuracy and durable review (2026-07-15) diff --git a/docs/CODEX_EXECUTION_LOG.md b/docs/CODEX_EXECUTION_LOG.md index a14471a6..39563742 100644 --- a/docs/CODEX_EXECUTION_LOG.md +++ b/docs/CODEX_EXECUTION_LOG.md @@ -8227,6 +8227,22 @@ Known limitation: independently collected training-only AOIs and complete tile labels; holdout review evidence must not be recycled into training. +Live operational proof: +- Deployed commits `919e102` and packaging fix `8758721` to the Tower + all-in-one runtime. The container became healthy, PostGIS reported 3.6, + Alembic remained at `202607150001`, and the frontend/API/icon proxy smoke + passed. +- Both review validators ran inside the rebuilt image against persistent + operator data with `status=complete`: 5 confirmed model false positives and + 10 confirmed model false negatives, with zero unreviewed cards. +- A new 15.31 ha browser-drawn Mol selection persisted 282 candidates and 209 + strict GRB matches. It measured precision `0.741`, recall `0.685`, F1 `0.712`, + 73 false positives and 96 false negatives. The diagnostic reference-envelope + pass found 212 matches and only 3 possible box/footprint differences. +- Browser checks at 1280x720 and 3440x1440 showed the strict and diagnostic + values without horizontal overflow. The ultrawide map used about 2,408 px; + the browser console contained no warnings or errors. + Next: - Collect a new training-only small-building/background evidence pack outside all operational holdouts, then train an inactive candidate only if the pack