correct overlapping checkpoint evaluation

docs/CODEX_EXECUTION_LOG.md

@@ -12671,3 +12671,36 @@ Open:
 - The min-4px corpus is preferred for future experimental work, but remains
   explicitly training/release-ineligible until real human review and governed
   release evidence exist. The active GPU model was not changed.
+
+## 2026-08-09 - Non-overlapping checkpoint correction
+
+### Corrected
+
+- Traced the prior 36-image checkpoint matrix to four AOIs tiled at 512/256.
+  The same interiors were repeated and nine Arendonk-heide tiles were visually
+  blank no-data, despite being counted as background observations.
+- Added a deterministic non-overlap evaluation-view builder. It selects four
+  complete 512 px grid tiles per 1024 px AOI, excludes low-variance no-data
+  with reason codes, creates empty train directories and writes a hashed
+  `NO_TRAINING.json` marker.
+- Added a fail-fast marker check to the operator YOLO training wrapper before
+  release verification, model loading or CUDA allocation.
+- Corrected checkpoint overlap semantics: zero pixel overlap is reported, but
+  statistical independence is explicitly not established for adjacent tiles.
+
+### GPU evidence
+
+- The clean r2 view contains 12 images: eight positive Turnhout/Westerlo tiles
+  and four real Postel backgrounds. It has 2,551 labels, 2,384 unique interior
+  objects, 167 edge labels and zero repeated interior objects.
+- Active: precision 0.542632, recall 0.435317, mAP50 0.342034, mAP50-95
+  0.141316. Challenger: 0.581715, 0.462300, 0.368364 and 0.155318.
+- At confidence 0.15 the active model has zero Postel detections and the
+  challenger one; at 0.25 both have zero. The threshold-sensitive relative
+  challenger advantage is confirmed without changing production.
+
+### Remaining limitations
+
+- 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.

docs/TODO.md

@@ -1170,6 +1170,13 @@ This file now starts with the current implementation status. Older preparation/b
 - [ ] Compare the existing 0.35 edge-visibility policy against a separately
   versioned stricter ablation from original source geometry before any new
   governed training; do not approximate this by dropping every edge label.
+- [x] Replace the overlapping checkpoint ranking with a checksum-bound,
+  training-disabled non-overlap view: 12 real tiles, 2,384 unique interior
+  objects, zero repeated interiors and explicit removal of blank Arendonk
+  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.
 - [ ] 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

@@ -187,3 +187,32 @@ visible; it was not suppressed or relabelled as success.
 This min-4px version is the preferred experimental successor to the legacy
 min-3px corpus. It explicitly remains ineligible for governed training until
 real human review and all release-contract evidence exist.
+
+## Non-overlapping checkpoint re-evaluation
+
+The 36-tile checkpoint matrix was traced to a 512 px validation corpus with
+stride 256. It contained nine views each of Turnhout, Westerlo, Postel-bos and
+Arendonk-heide. Visual inspection then exposed that all Arendonk-heide views
+are blank/no-data imagery, not meaningful pure-background observations. The
+historical claim of 18 background images is therefore corrected: nine were
+blank no-data and nine represented Postel under overlap.
+
+A training-disabled validation view now covers each 1024 px AOI with four
+non-overlapping 512 px tiles. Four blank Arendonk representatives are excluded
+with explicit reason codes. The resulting set contains eight positive tiles
+from Turnhout/Westerlo, four real Postel background tiles and 2,551 labels.
+Exact reconstruction finds 2,384 unique interior objects, 167 edge labels and
+zero repeated interior objects. The view has empty train directories, a
+checksum-bound `NO_TRAINING.json`, and the training wrapper rejects that marker.
+
+On the Tower RTX 4080 SUPER, the active checkpoint measures precision
+`0.542632`, recall `0.435317`, mAP50 `0.342034` and mAP50-95 `0.141316`.
+The reviewedexp6 challenger measures `0.581715`, `0.462300`, `0.368364` and
+`0.155318` respectively. At background confidence 0.15 the active checkpoint
+has zero Postel detections and the challenger has one. At 0.25 both have zero.
+Thus the refined evidence confirms the challenger's relative metric advantage
+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.