record blocked legacy model provenance migration

artifacts/evidence/accuracy/model-training/20260809-reviewedexp6-provenance-migration-audit.json

@@ -0,0 +1,102 @@
+{
+  "schema_version": 1,
+  "generated_at": "2026-08-09T13:20:00Z",
+  "audit_id": "reviewedexp6-provenance-migration-audit-r1",
+  "status": "blocked_lineage_incomplete",
+  "model": {
+    "path": "/app/models/geointel-building-yolov8s-reviewedexp6-minpx3-img640-ft20.pt",
+    "sha256": "038f1f97a6afd534f29e1f392a730a58207b928ca01e31ab8d8fed6106705820",
+    "size_bytes": 22514794,
+    "observed_mtime": "2026-07-15T05:11:29.338792783+02:00",
+    "lifecycle": "legacy_unpromoted_challenger"
+  },
+  "observed_training_artifacts": [
+    {
+      "role": "training_summary",
+      "path": "/app/storage/training/operator-yolo/geointel-building-yolov8s-reviewedexp6-minpx3-img640-ft20/training_summary.json",
+      "sha256": "961b9d119b24967977c22ba89a54f9caea51e440769bebde5641b7ee8bfae1eb"
+    },
+    {
+      "role": "ultralytics_arguments",
+      "path": "/app/storage/training/operator-yolo/geointel-building-yolov8s-reviewedexp6-minpx3-img640-ft20/args.yaml",
+      "sha256": "87e07dbed3ad12cc474e6b895147839431805b7044496277242e1538fcd91a0b"
+    },
+    {
+      "role": "training_metrics",
+      "path": "/app/storage/training/operator-yolo/geointel-building-yolov8s-reviewedexp6-minpx3-img640-ft20/results.csv",
+      "sha256": "57e72102b855c138b212efc293249952ec821e0f07201666ea4b17fb23d727a9"
+    },
+    {
+      "role": "dataset_summary",
+      "path": "/app/storage/operator-data/yolo-building-aoi1024-reviewedexp6-minpx3vis035/yolo_tile_dataset_summary.json",
+      "sha256": "7c917e31216d1df2174c0f9c736f88a81f3835aa991971f8fb8665e17ddf5c9c"
+    },
+    {
+      "role": "dataset_yaml",
+      "path": "/app/storage/operator-data/yolo-building-aoi1024-reviewedexp6-minpx3vis035/dataset.yaml",
+      "sha256": "62e12f4433508cb0640f1d90ec0ffd389cd21ebe655fac4e7d9402f04da1168e"
+    },
+    {
+      "role": "legacy_dataset_quality_audit",
+      "path": "/app/storage/operator-data/model-review/reviewedexp6-corpus/audit/operator_yolo_dataset_quality_audit.json",
+      "sha256": "42c34cddfcf7b8656d4513efa4e7249d73678164c2a1121f245be42ac56daf12"
+    }
+  ],
+  "reproduced_training_configuration": {
+    "task": "detect",
+    "epochs": 20,
+    "image_size": 640,
+    "batch_size": 4,
+    "device": "cpu",
+    "seed": 0,
+    "deterministic_requested": true,
+    "base_model_sha256": "a9088b8491dfae36694b53e9e9406cb4e3511d334a5712fa34f75078a47759c1",
+    "dataset_tile_count": 252,
+    "dataset_label_count": 79192,
+    "train_tile_count": 234,
+    "validation_tile_count": 18
+  },
+  "human_review_gate": {
+    "evaluated_with": "scripts.training_release_manifest.human_review_audit_failures",
+    "passed": false,
+    "failures": [
+      "accepted_human_review_evidence_missing",
+      "review_audit_manifest_not_immutable",
+      "review_audit_spatial_leakage_not_ok",
+      "review_complete_not_true"
+    ],
+    "interpretation": "The retained quality audit is not a checksum-bound accepted human-review ledger and cannot be upgraded into one retroactively."
+  },
+  "missing_required_original_evidence": [
+    "exact training repository commit recorded at training start",
+    "training container image digest",
+    "pinned Python, PyTorch, CUDA and Ultralytics runtime receipt for the original CPU run",
+    "dependency lock or SBOM checksum",
+    "immutable source corpus manifest and freeze-sidecar checksum",
+    "label release checksum",
+    "independent split manifest checksum and zero-under-2-km audit",
+    "accepted human review ledger with reviewer identity, per-sample decision, timestamp and reviewed artifact hash",
+    "complete training RunManifest binding all upstream artifacts before execution"
+  ],
+  "forbidden_inferences": [
+    "A nearby Git commit may not be assigned as the training commit from file times.",
+    "The current container runtime may not be copied backwards as the historical training runtime.",
+    "The dataset quality contact sheet may not be represented as accepted human review.",
+    "A structurally valid runtime sidecar may not be generated with invented UUIDs or placeholder upstream checksums."
+  ],
+  "actions": {
+    "runtime_sidecar_created": false,
+    "source_registry_rows_created": false,
+    "production_inference_executed": false,
+    "production_model_changed": false,
+    "candidate_quarantine_state": "not_active_and_not_exposed_by_production_catalog",
+    "active_model_sha256": "a9088b8491dfae36694b53e9e9406cb4e3511d334a5712fa34f75078a47759c1"
+  },
+  "decision": {
+    "migration_allowed": false,
+    "promotion_allowed": false,
+    "new_training_allowed": false,
+    "reason": "GeoIntel's execution contract explicitly forbids retroactively completing legacy RunManifest evidence. The current accuracy-program ledger also keeps training and promotion locked until P2-00 through P2-07 and the governed Phase-4 product gates pass.",
+    "safe_next_gate": "Complete P2-00 through P2-07, obtain representative human review, freeze a new independently split corpus, then train a new pre-registered GPU candidate with complete provenance from process start."
+  }
+}

docs/CODEX_EXECUTION_LOG.md

@@ -12498,3 +12498,33 @@ Open:
 - The active model remains review-required and Kempen-scoped. The export gate
   prevents overstated results; it does not replace the missing independent,
   human-reviewed national accuracy corpus or improve model weights by itself.
+
+## 2026-08-09 - Reviewedexp6 provenance migration audit
+
+### Inspected
+
+- Recovered and checksummed the original challenger weights, training summary,
+  Ultralytics arguments, results, dataset YAML, 252-tile/79,192-label summary
+  and legacy dataset-quality audit.
+- Reproduced the recorded seed `0`, deterministic request, CPU device, twenty
+  epochs, active base-model hash and train/validation counts.
+- Ran the current human-review eligibility validator against the retained
+  audit; it failed on missing accepted review evidence, manifest immutability,
+  spatial-leakage status and incomplete review state.
+
+### Decision
+
+- Refused to invent the absent historical training commit, container digest,
+  runtime receipt, corpus/label/split manifests, reviewer decisions or source
+  registry UUIDs.
+- Created no runtime sidecar, database provenance row or production model
+  activation. The active scoped model and healthy GPU runtime remain unchanged.
+- Retained the complete fail-closed audit under
+  `artifacts/evidence/accuracy/model-training/20260809-reviewedexp6-provenance-migration-audit.json`.
+
+### Required next gate
+
+- Complete P2-00 through P2-07 and the governed Phase-4 blockers, obtain real
+  representative human review, freeze a new independent corpus and only then
+  train a pre-registered GPU candidate with provenance captured from process
+  start. The legacy checkpoint cannot be made compliant retroactively.

docs/TODO.md

@@ -1141,3 +1141,9 @@ This file now starts with the current implementation status. Older preparation/b
   image/model/config hashes, shadow evidence and tested rollback.
 - [ ] P2-12: monitor drift into a human reviewqueue without automatic
   self-training or silent promotion.
+- [x] Audit whether the July `reviewedexp6` checkpoint can be migrated into the
+  current runtime provenance contract without invented evidence; retain the
+  checksum-bound blocked report.
+- [ ] Do not generate a production sidecar for `reviewedexp6`: its exact
+  training commit/runtime, immutable corpus-label-split manifests and accepted
+  human-review ledger were not retained and cannot be reconstructed safely.

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

@@ -62,3 +62,28 @@ the only safe production choice.
 
 Machine-readable hashes, metrics, paths and the exact decision are in
 `artifacts/evidence/accuracy/model-training/20260809-v68-checkpoint-and-threshold-review.json`.
+
+## Provenance migration follow-up
+
+The original July training directory was audited after the threshold review.
+It retains exact weights, Ultralytics arguments, result curves, dataset YAML,
+dataset summary and a structural quality audit. This recovers useful facts,
+including seed `0`, deterministic mode, the base-model hash and all dataset
+counts.
+
+It does not retain the complete evidence required to construct a current
+production sidecar truthfully. In particular, the exact training commit,
+container digest, dependency/runtime receipt, immutable corpus and label
+release manifests, independent split audit and accepted human-review ledger
+are absent. The current review validator reproduces four failures:
+
+- `accepted_human_review_evidence_missing`;
+- `review_audit_manifest_not_immutable`;
+- `review_audit_spatial_leakage_not_ok`; and
+- `review_complete_not_true`.
+
+No UUID, upstream checksum, historical runtime or reviewer decision was
+invented. Consequently no runtime sidecar or database source snapshot was
+created, and the challenger remains unavailable to production inference. The
+full machine-readable audit is retained in
+`artifacts/evidence/accuracy/model-training/20260809-reviewedexp6-provenance-migration-audit.json`.