block training-seen checkpoint evaluation

docs/CODEX_EXECUTION_LOG.md

@@ -12704,3 +12704,36 @@ Open:
 - Two adjacent positive AOIs and one real background AOI are not representative
   release evidence. Human review, broader independent geography and exact
   challenger provenance remain blocking.
+
+## 2026-08-09 - Training/evaluation membership gate
+
+### Corrected
+
+- Compared the non-overlap evaluation sample membership against the exact
+  active and challenger training summaries. Turnhout and Westerlo are absent
+  from both train splits; Postel-bos is present in both.
+- Reclassified the Postel detection counts as training-seen regression checks,
+  not independent background validation. No existing metric or checkpoint was
+  promoted on that evidence.
+- Extended the checkpoint evaluator with a strict multi-corpus membership gate.
+  Governed runs now persist a blocked manifest and stop before PyTorch import,
+  model loading and GPU inference if any evaluation AOI is training-seen.
+
+### Reproduced on Tower
+
+- Active training summary SHA-256:
+  `49b2a07d2105d08356431757b83eafc1498eaf1fb76965b1efe05b776824942a`.
+- Challenger training summary SHA-256:
+  `7c917e31216d1df2174c0f9c736f88a81f3835aa991971f8fb8665e17ddf5c9c`.
+- Evaluation summary SHA-256:
+  `c17a8c47eaecc2dfb38778ff5a612f9e6c4d67c9e90e31d3a6a66d8857adec4c`.
+- The strict run blocked on `postel_bos`, with
+  `model_loading_attempted=false` and `gpu_inference_attempted=false`.
+- Blocked-manifest SHA-256:
+  `62e8a9377ac87552936f472b0acf55d1d659ebf27f84ba8c4dbcefe7a374e1da`.
+
+### Remaining limitation
+
+- Independent real-background AOIs are still required. Literal 100% model
+  correctness is not demonstrated; the system now refuses to misrepresent the
+  known overlap as proof.

docs/TODO.md

@@ -1176,7 +1176,11 @@ This file now starts with the current implementation status. Older preparation/b
   no-data imagery.
 - [ ] Add multiple independent real-background AOIs and more geographically
   separated positive validation AOIs before interpreting non-overlap ranking
-  as release evidence; Postel alone is not representative.
+  as release evidence. Postel occurs in both compared train splits and is only
+  a training-seen regression check, not independent validation.
+- [x] Add a fail-closed checkpoint gate that checks every evaluation AOI
+  against the exact train split of every supplied model corpus before PyTorch
+  import, model loading or GPU inference.
 - [ ] Convert the AI-assisted ledger into no stronger claim than experimental
   triage; a real human must independently review and sign the frozen artifacts
   before the governed training wrapper may unlock.

docs/reviews/2026-08-09-ai-assisted-checkpoint-review.md

@@ -216,3 +216,24 @@ and its threshold sensitivity, but still does not authorize promotion.
 Non-overlap is not overstated as statistical independence: the positive tiles
 remain adjacent and come from only two AOIs, edge objects can remain split, and
 there is only one real background AOI. The production model remains unchanged.
+
+## Training/evaluation membership correction
+
+The evaluation AOIs were subsequently checked against the exact tile summaries
+of both compared checkpoints. `turnhout` and `westerlo` do not occur in either
+train split. `postel_bos`, however, occurs in both: the active corpus contains
+20 training AOIs and the challenger corpus 26, with Postel included in each.
+The Postel image is real rather than blank/no-data, but it is training-seen.
+
+Consequently the positive Turnhout/Westerlo metrics remain a non-protected,
+adjacent-AOI candidate ranking; the Postel detection counts are only
+training-seen sanity/regression observations. They are not independent
+pure-background validation and cannot support release, threshold or
+generalisation claims. The earlier wording about a "real background AOI" must
+be read with this correction.
+
+The checkpoint evaluator now accepts exact training summaries and, in governed
+mode, fails before PyTorch import, model loading or GPU inference when any
+evaluation AOI overlaps any supplied train split. The reproduced gate blocked
+on `postel_bos` for both checkpoints. Its immutable machine-readable record is
+`artifacts/evidence/accuracy/model-training/20260809-v70-evaluation-independence-gate.json`.