Files
geointel/artifacts/evidence/accuracy/model-training/20260809-v69-nonoverlap-checkpoint-review.json
T
Jens 48084799c9
GeoIntel release gates / Compile, test, contracts and builds (push) Canceled after 0s
GeoIntel release gates / Python and npm vulnerability policy (push) Canceled after 0s
GeoIntel release gates / GIS image, SBOM and container scan (push) Canceled after 0s
correct overlapping checkpoint evaluation
2026-08-09 20:41:54 +02:00

73 lines
3.5 KiB
JSON

{
"schema_version": 1,
"run_id": "building-kempen-v69-nonoverlap-checkpoint-review-r2",
"generated_at": "2026-08-09T20:35:51+02:00",
"status": "completed_no_production_change",
"claim_boundary": "Non-protected two-positive-AOI and one-background-AOI validation refinement only; tiles do not overlap, but statistical independence, protected-test performance, national accuracy and promotion eligibility are not established.",
"correction_to_prior_evidence": {
"prior_validation_tile_count": 36,
"prior_tile_overlap_pixels": 256,
"prior_background_tile_count": 18,
"invalid_background_sample": "arendonk_heide",
"invalid_background_reason": "all nine source tiles are visually blank/low-variance no-data; four non-overlap representatives were excluded with explicit reason codes",
"effect": "The prior matrix remains a historical overlapping-tile ranking, but its 18-image background count must not be interpreted as 18 real background observations."
},
"evaluation_view": {
"path": "/app/storage/operator-data/yolo-building-aoi1024-bgaware512r8-nonoverlap-eval-r2",
"manifest_sha256": "69f555e57fa52d8b7cdfc5632167bde96b64030ee37e9bf49ff2d9891d66574a",
"summary_sha256": "c17a8c47eaecc2dfb38778ff5a612f9e6c4d67c9e90e31d3a6a66d8857adec4c",
"cross_tile_audit_sha256": "b0a49d7f5983ccaa07754a1e3d6d8f7b10345d3c57b9f613bb856de9d6f70b74",
"contact_sheet_sha256": "f5954a2f6ad3dc59642bf3a51d98ca94343491f522968ab38d60c999ed21c797",
"tile_size": 512,
"stride": 512,
"tile_count": 12,
"positive_tile_count": 8,
"real_background_tile_count": 4,
"positive_aoi_count": 2,
"background_aoi_count": 1,
"label_count": 2551,
"unique_interior_object_count": 2384,
"edge_label_count": 167,
"repeated_interior_object_count": 0,
"excluded_low_variance_tile_count": 4,
"excluded_sample_slug": "arendonk_heide",
"training_prohibited": true,
"statistical_independence_established": false
},
"runtime": {
"device": "cuda:0",
"gpu": "NVIDIA GeForce RTX 4080 SUPER",
"torch": "2.11.0+cu128",
"ultralytics": "8.4.99"
},
"checkpoint_metrics": {
"active": {
"model_sha256": "a9088b8491dfae36694b53e9e9406cb4e3511d334a5712fa34f75078a47759c1",
"precision": 0.5426324241859346,
"recall": 0.4353171786761234,
"map50": 0.342034346792601,
"map50_95": 0.1413160008545276,
"background_detections_confidence_0_15": 0,
"background_detections_confidence_0_25": 0
},
"challenger": {
"model_sha256": "038f1f97a6afd534f29e1f392a730a58207b928ca01e31ab8d8fed6106705820",
"precision": 0.5817150200182469,
"recall": 0.462299518515736,
"map50": 0.3683635371499511,
"map50_95": 0.15531818619443444,
"background_detections_confidence_0_15": 1,
"background_detections_confidence_0_25": 0
},
"confidence_0_15_matrix_sha256": "915c8523fe71a2d88b44bcb75d9070d25940f3969213692cc121f1239b31489b",
"confidence_0_25_matrix_sha256": "173d8af94d5d3d805b67a1a99d36e196255166e2e32ecabbef40ebf065c92138"
},
"decision": {
"relative_metric_ranking": "challenger",
"strict_background_ranking_at_0_15": "active",
"background_tie_at_0_25": true,
"production_model_changed": false,
"reason": "The challenger remains stronger on this refined non-overlapping ranking and clears Postel at 0.25, but one real background AOI and two adjacent positive AOIs are insufficient for release; historical provenance and human-review gates also remain closed."
}
}