evaluate models on fresh regional calibration AOIs

2026-08-09 22:36:27 +02:00
parent 116b8e291e
commit b068a5e065
28 changed files with 5053 additions and 18 deletions
@@ -0,0 +1,61 @@
+{
+  "schema_version": 1,
+  "run_id": "building-be-v72-fresh-calibration-r1",
+  "status": "completed_diagnostic_no_production_change",
+  "claim_boundary": "Fresh non-protected calibration diagnostics only. Wallonia and Brussels remain evaluation-only while regional authority harmonisation contracts are pending; no protected-test, release or national accuracy claim.",
+  "portfolio": {
+    "region_count": 3,
+    "aoi_count": 6,
+    "tile_count": 24,
+    "label_count": 348,
+    "tile_size": 512,
+    "stride": 512,
+    "overlapping_edge_cover_tiles_excluded": 30,
+    "training_allowed": false,
+    "release_claim_allowed": false,
+    "minimum_model_lineage_distance_m": 4869.625822725313,
+    "model_lineage_sample_count": 31,
+    "model_lineage_spatial_status": "independent"
+  },
+  "visual_review": {
+    "reviewer_type": "ai_assisted",
+    "human": false,
+    "reviewed_tile_count": 24,
+    "finding": "All retained tiles contain real imagery. Havelange is a positive village context despite its pre-registered background-candidate name; Assenede and Laeken are sparse/hard-negative contexts, not pure-empty backgrounds. No pure-background claim was made."
+  },
+  "gpu": "NVIDIA GeForce RTX 4080 SUPER",
+  "models": {
+    "active": {
+      "sha256": "a9088b8491dfae36694b53e9e9406cb4e3511d334a5712fa34f75078a47759c1",
+      "precision": 0.1576550586330707,
+      "recall": 0.11781609195402298,
+      "map50": 0.05677865740374553,
+      "map50_95": 0.02175035531261916,
+      "f1_at_0_25_iou_0_25": 0.2315270935960591
+    },
+    "challenger": {
+      "sha256": "038f1f97a6afd534f29e1f392a730a58207b928ca01e31ab8d8fed6106705820",
+      "precision": 0.13343411909876413,
+      "recall": 0.1235632183908046,
+      "map50": 0.04470477656341424,
+      "map50_95": 0.01600413321513577,
+      "f1_at_0_25_iou_0_25": 0.23588456712672523
+    }
+  },
+  "threshold_decision": {
+    "best_aggregate_active_f1_threshold": 0.3,
+    "best_aggregate_active_f1": 0.25035561877667145,
+    "production_threshold_changed": false,
+    "reason": "The aggregate optimum reduces the already weak Flemish subgroup F1, and regional authority/human-review gates remain closed."
+  },
+  "evidence_sha256": {
+    "operator_samples_manifest": "66bfaacaee11806ed035d3ac5725f7b3967dfc82e7a12ecf603dc3836bc26ad7",
+    "yolo_tile_dataset_summary": "78e16f46fa2a313a147404a19f2ce99417d2cb384b5db5f6ba33e934ff99c631",
+    "spatial_independence": "4971887d0f2356bfba674b0f0a4606530920b30d458765a1fec5b3086ac47e72",
+    "contact_sheet": "f7734c5a9db8837adeda7f8e0f1a2942148330abfcb588d41dc245017ec2435a",
+    "checkpoint_matrix": "ad178fac5968c7e1d6b1c79b192624e4cd2956f054d5cf70270275a0be0b80c5",
+    "active_stratified": "eefa9bab20674c30427f02ca19571f869f728eccf2dca376acf02ae4ef0e7678",
+    "challenger_stratified": "9c22f6cdc30555616f2d1b595fdfde544474be396f2bd86ff22f684d3ddd6e58",
+    "active_threshold_extension": "7752f2e2bd84306f3863fb32eb141b30d473581c9ef0dbb5a6970a1d3c73a0ec"
+  }
+}