derive cleaner min-4px YOLO corpus

2026-08-09 19:51:17 +02:00
parent 736e773fb8
commit 2de438b9cc
9 changed files with 471 additions and 25 deletions
@@ -0,0 +1,63 @@
+{
+  "schema_version": 1,
+  "evidence_id": "reviewedexp6-minpx4-derived-corpus-r1",
+  "created_at": "2026-08-09T19:45:41+02:00",
+  "status": "complete_experimental_only",
+  "claim_boundary": "Deterministic derived-corpus improvement and AI-assisted visual triage only; no human-review, governed training, evaluation, promotion or production claim.",
+  "source": {
+    "dataset_dir": "/app/storage/operator-data/yolo-building-aoi1024-reviewedexp6-minpx3vis035",
+    "summary_sha256": "7c917e31216d1df2174c0f9c736f88a81f3835aa991971f8fb8665e17ddf5c9c",
+    "tile_size": 512,
+    "stride": 256,
+    "min_label_px": 3.0,
+    "min_visible_ratio": 0.35,
+    "tile_count": 252,
+    "label_count": 79192
+  },
+  "derived": {
+    "dataset_dir": "/app/storage/operator-data/yolo-building-aoi1024-reviewedexp6-minpx4-derived-r1",
+    "manifest_sha256": "8a80d0e9d5ac2d32556e2f1ea0c6009045d48a3e56d931b3161622dca4173a67",
+    "summary_sha256": "391b0302cc8a46d15fef426d8971a4a8e594fb6d63ed1c7ce16feeee9f1406",
+    "quality_audit_sha256": "0fc9391ce09a399f51f1e0ed30c677f482c7c1b49b187926da01ecdfd9f1067e",
+    "outlier_audit_sha256": "c2453abb83aaff8aebd84f63efb14da131225192cf0434d5d1df7cf06be240ca",
+    "relationship_audit_sha256": "61660d0c448ff579272bc585de8140e82d1577c0b017e1c078b66dc093f14c2e",
+    "cross_tile_repetition_audit_sha256": "9ca22d83271707713f7f96d5409af598f834787ad5a48be14f8752a704de0da0",
+    "min_label_px": 4.0,
+    "tile_count": 252,
+    "positive_tile_count": 234,
+    "negative_tile_count": 18,
+    "label_count": 77380,
+    "removed_label_count": 1812,
+    "excluded_tile_count": 0,
+    "invalid_label_count": 0,
+    "missing_label_file_count": 0,
+    "sub_4_pixel_label_count": 0,
+    "extreme_aspect_label_count": 10,
+    "tile_edge_label_count": 4840,
+    "possible_nested_pair_count": 41,
+    "exact_duplicate_pair_count": 0,
+    "near_duplicate_pair_count": 0,
+    "cross_split_repetition_group_count": 0,
+    "training_release_eligible": false
+  },
+  "visual_review": {
+    "reviewer_type": "ai_assistant",
+    "human": false,
+    "rendered_tile_count": 252,
+    "rendered_label_count": 77380,
+    "contact_sheet_sha256s": [
+      "f197769ca68da08745277ff93331f3ea93738f6b63eb383a3b6168a9e438ee12",
+      "e9ecde3f38b4f580adb7523f647460af5d5933e2e1efe1f11f98c1656ca1cd13",
+      "497bf5878560113fcc59f21aeba09b1de167cb3eff4d0327c7584ba8caeebf70",
+      "4a8262291ed29699ae59517e1a542e513a50081a225ea1e84268afa37f99d6df"
+    ],
+    "extreme_aspect_contact_sheet_sha256": "6c08c42eb6478d1fa06745135dd588ce6522b5cec4adf0bdcd177d77cd6c78a7",
+    "conclusion": "All 252 tiles remain visually coherent; pure-background tiles remain empty, sparse contexts remain distinct, and the ten remaining extreme-aspect labels predominantly correspond to plausible elongated structures."
+  },
+  "decision": {
+    "preferred_experimental_corpus": true,
+    "replaces_source_in_place": false,
+    "reason": "It deterministically removes all 1,812 visually marginal sub-4-pixel labels while preserving every tile and all larger labels with exact lineage.",
+    "remaining_gate": "Requires independent human review and the governed release-contract inputs before training eligibility."
+  }
+}