diff --git a/artifacts/evidence/accuracy/model-training/20260809-ai-assisted-visual-review.json b/artifacts/evidence/accuracy/model-training/20260809-ai-assisted-visual-review.json
new file mode 100644
index 00000000..fe3da58d
--- /dev/null
+++ b/artifacts/evidence/accuracy/model-training/20260809-ai-assisted-visual-review.json
@@ -0,0 +1,62 @@
+{
+  "schema_version": 1,
+  "generated_at": "2026-08-09T09:57:02.8274719Z",
+  "review_kind": "ai_assisted_visual_review",
+  "reviewer_id": "openai-codex-gpt-5",
+  "human_review": false,
+  "repository_commit": "0209167cfd37c4f7865813a06e60bdd7ba5d918d",
+  "scope": "Bounded visual inspection of deterministic YOLO contact sheets; this is not a human-signed operational training release.",
+  "decision": {
+    "status": "experimental_training_only",
+    "promotion_allowed": false,
+    "protected_test_access_allowed": false,
+    "reason": "The Kempen background-aware corpus is materially better aligned than the rejected national candidates, but only a bounded contact-sheet sample was visually inspected and the mandatory human review gate remains open."
+  },
+  "accepted_experimental_input": {
+    "dataset_yaml": "/app/storage/operator-data/yolo-building-aoi1024-bgaware512r8/dataset.yaml",
+    "dataset_yaml_sha256": "6995843d63735bc248fb573d6ab3ebedb1cd1c86f1bd6dbb864068c5fb8f816d",
+    "dataset_summary": "/app/storage/operator-data/yolo-building-aoi1024-bgaware512r8/yolo_tile_dataset_summary.json",
+    "dataset_summary_sha256": "c0c94295e52de5cc4d4cdcaaabb670220f15d2139c0051fff63d05730d33bdce",
+    "contact_sheet": "/app/storage/operator-evidence/accuracy-review/20260809-bgaware512r8-r1/contact_sheet_001.png",
+    "contact_sheet_sha256": "c5112d5c76d84153f8d219699071e7f42c8b188b0624c2bf5f9406a2146e66b6",
+    "rendered_tile_count": 32,
+    "dataset_tile_count": 207,
+    "observations": [
+      "Most sampled labels visually overlap roof structures in the Kempen positive AOIs.",
+      "The training corpus contains repeated pure-empty background tiles and the validation split contains independent named positive and pure-empty AOIs.",
+      "Some dense scenes still contain footprint-versus-visible-roof ambiguity; this prevents operational acceptance."
+    ]
+  },
+  "base_model": {
+    "path": "/app/models/geointel-building-yolov8s-smallbld-minpx3-img640-ft30.pt",
+    "sha256": "a9088b8491dfae36694b53e9e9406cb4e3511d334a5712fa34f75078a47759c1"
+  },
+  "rejected_inputs": [
+    {
+      "family": "building-be-v56 national corpus",
+      "decision": "exclude_from_new_training",
+      "reason": "Representative sheets show frequent ground-footprint versus visible-roof displacement and dense-scene ambiguity; 0/180 samples have formal human acceptance."
+    },
+    {
+      "family": "building-be-v65 main/SAM candidates",
+      "decision": "quarantine",
+      "reason": "Visual samples include boxes on trees, paving and unrelated compounds after automatic refinement."
+    },
+    {
+      "family": "building-be-v66 source/alignment/SAM candidates",
+      "decision": "quarantine",
+      "reason": "Systematic misalignment remains; the swapped-axis trial uses unrelated imagery and the detector-alignment sheet contains invalid labels."
+    }
+  ],
+  "training_constraints": {
+    "device": "cuda:0",
+    "allowed_evaluation_split": "val",
+    "forbidden_splits": [
+      "test",
+      "challenge",
+      "background-test"
+    ],
+    "automatic_promotion": false,
+    "formal_release_manifest": false
+  }
+}
diff --git a/artifacts/evidence/accuracy/model-training/20260809-v67-training-decision.json b/artifacts/evidence/accuracy/model-training/20260809-v67-training-decision.json
new file mode 100644
index 00000000..76ff0379
--- /dev/null
+++ b/artifacts/evidence/accuracy/model-training/20260809-v67-training-decision.json
@@ -0,0 +1,51 @@
+{
+  "schema_version": 1,
+  "generated_at": "2026-08-09T10:01:00Z",
+  "run_id": "building-kempen-v67-ai-assisted-bgaware-r1",
+  "status": "completed_rejected",
+  "device": {
+    "requested": "cuda:0",
+    "observed": "NVIDIA GeForce RTX 4080 SUPER",
+    "torch": "2.11.0+cu128",
+    "ultralytics": "8.4.99"
+  },
+  "training": {
+    "epochs": 12,
+    "image_size": 640,
+    "batch_size": 8,
+    "workers": 0,
+    "seed": 20260809,
+    "base_model_sha256": "a9088b8491dfae36694b53e9e9406cb4e3511d334a5712fa34f75078a47759c1",
+    "candidate_best_sha256": "1337eb9fdd6bb73c525bd616e2ba961f20d585387db940576055afd38a71ca91",
+    "results_csv_sha256": "73bae1cc3e377fda8b1004a801c881f8313d6dffbbb5986555e2c5226212177d",
+    "args_yaml_sha256": "374613f927c57f9e9832cf7e0131c987dc978cbe1ea787a827030a8378385122"
+  },
+  "validation": {
+    "split": "val",
+    "image_count": 36,
+    "positive_image_count": 18,
+    "pure_background_image_count": 18,
+    "instance_count": 5686,
+    "baseline": {
+      "precision": 0.542,
+      "recall": 0.430,
+      "map50": 0.345,
+      "map50_95": 0.141,
+      "pure_background_detections_at_confidence_0_15": 0
+    },
+    "candidate_best": {
+      "precision": 0.530,
+      "recall": 0.435,
+      "map50": 0.337,
+      "map50_95": 0.135,
+      "pure_background_detections_at_confidence_0_15": 3
+    }
+  },
+  "decision": {
+    "promote": false,
+    "production_model_changed": false,
+    "reason": "The small recall gain does not compensate for lower precision and mAP plus a regression from zero to three detections on pure-background validation tiles.",
+    "active_model_sha256_after_decision": "a9088b8491dfae36694b53e9e9406cb4e3511d334a5712fa34f75078a47759c1"
+  },
+  "claim_boundary": "Experimental Kempen validation only. No protected test or challenge data was opened, no national accuracy claim is made, and formal human review remains required for any future promotion."
+}