Record v31 iteration three assessment

docs/CODEX_EXECUTION_LOG.md

@@ -1,3 +1,23 @@
+## 2026-07-29 - V31 closed-loop iteration 3 assessment and semantic correction
+
+- Iteration 3 stopped deterministically after epoch 21 with epoch 3 retained as
+  the best checkpoint. Calibration selected threshold `0.10` and produced
+  aggregate precision `0.460359`, recall `0.737302` and F1 `0.566811`.
+- The candidate remained rejected: Flanders measured precision `0.340852`,
+  recall `0.403960` and F1 `0.369733`; Wallonia measured precision `0.456198`,
+  recall `0.786742` and F1 `0.577518`. Brussels passed its regional minimums.
+  Protected test and background-test evidence remained unopened.
+- Oostende remained the dominant protected calibration failure with precision
+  `0.082569`, recall `0.391304` and F1 `0.136364`. The next train-only sampling
+  manifest therefore targets Flemish coastal/port/dunes, industrial/rail and
+  ribbon/farmland/forest contexts, plus Walloon rural/architecture hard-negative
+  families. It contains 3,600 entries (Flanders 2,340, Wallonia 972, Brussels
+  288), removes 1,278 excess repeats to retain the 65% regional cap, and records
+  zero protected samples in training.
+- Iteration 4 started automatically on CUDA with the iteration 3 candidate and
+  exact failure-driven train-list SHA-256
+  `66f510ee563ae6ef8499c018cc1b12b175497dd1180b9c4eb73e9f62fb7d75fd`.
+
 ## 2026-07-29 - Context-aware Belgian training correction
 
 - Audited the active closed-loop calibration trend and confirmed that the