Filter labels created after dated imagery

docs/BELGIUM_BUILDING_TRAINING_LOOP.md

@@ -81,3 +81,8 @@ Every failed assessment returns `continue_training_loop`. Only a report with
 `training_complete` may proceed to final human review and guarded activation.
 The orchestrator refuses to start unless the frozen dataset audit is `ok` and
 contains zero blank/low-variance positive tiles.
+For dated imagery, GRB `BEGINDATUM` and PICC `DATE_CREAT` are compared with the
+end of the imagery period. A feature created afterward is retained in the
+audit but excluded from training as `created_after_imagery_period`. UrbIS does
+not expose an equivalent feature creation field in this acquisition contract,
+so its remaining temporal relation stays an explicit sample-level limitation.

docs/CODEX_EXECUTION_LOG.md

@@ -11587,3 +11587,18 @@ Next gate:
 - The v5 tile audit passed with zero blank positive tiles. The automated CUDA
   loop started from the strongest prior candidate and will checkpoint every
   train/calibrate/test/background assessment without promoting failed models.
+
+## 2026-07-27 - Feature-level temporal mismatch filtering
+
+- Added provider-native creation-time filtering for dated training imagery:
+  GRB `BEGINDATUM` and PICC `DATE_CREAT` are parsed with explicit UTC handling.
+  Buildings created after the image period are audited and excluded rather
+  than taught as labels for structures absent from the image.
+- Frozen corpus `building-be-v6-temporal-20260727-r1` excludes 246 such temporal
+  mismatches, accepts 13,524 labels and retains all 75 independent AOIs.
+  Manifest SHA-256 is
+  `973828b453e6fbeb5c04aa567ddb615566d92825d8698654f5056d2997d382eb`.
+- Composition, spatial leakage, temporal identity and positive-imagery QA pass.
+  The v6 train/calibration/test/background exports are ready for the next loop
+  checkpoint; the running v5 iteration remains evidence but cannot supersede
+  the cleaner v6 corpus.

docs/TODO.md

@@ -955,3 +955,4 @@ This file now starts with the current implementation status. Older preparation/b
 - [ ] Resolve the v3 Flanders and Wallonia generalisation failures through additional training-only evidence and retraining.
 - [x] Replace rolling-mosaic training inputs with governed dated 2025 Flanders/Brussels and complete 2023 SPW imagery; retain exact flight-day limitations.
 - [x] Reject positive labels over blank/no-data imagery and replace partial SPW 2024 coverage with the complete dated SPW 2023 campaign.
+- [x] Exclude GRB/PICC features created after the corresponding dated imagery period while retaining auditable rejection evidence.