Files
geointel/artifacts/evidence/accuracy/model-training/20260810-v73-flanders-remediation/outcome.json
T
Jens 98188e0a44
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
evaluate fresh Flemish remediation training
2026-08-10 03:47:51 +02:00

61 lines
2.3 KiB
JSON

{
"schema_version": 1,
"status": "completed_candidates_rejected",
"claim_boundary": "Experimental training and independent calibration only; no production release claim.",
"training_portfolio": {
"region": "flanders",
"sample_count": 12,
"unique_non_overlapping_tile_count": 48,
"label_count": 1710,
"negative_tile_count": 3,
"excluded_overlapping_edge_cover_tile_count": 60,
"source_authority": "GRB",
"source_manifest_sha256": "afa0edc5f229e0451d4a3630bb0d716b0ac1d446a6f8c2f22807d4f83b818525",
"human_review_complete": false,
"ai_assisted_visual_review_tiles": 48
},
"spatial_independence": {
"status": "independent",
"minimum_required_distance_m": 2000.0,
"observed_minimum_distance_m": 9203.454080711017
},
"frozen_v72_calibration": {
"tile_count": 24,
"instance_count": 348,
"active": {
"model_sha256": "a9088b8491dfae36694b53e9e9406cb4e3511d334a5712fa34f75078a47759c1",
"precision": 0.1576550586330707,
"recall": 0.11781609195402298,
"map50": 0.05677865740374553,
"map50_95": 0.02175035531261916
},
"aggressive_e20": {
"decision": "rejected",
"model_sha256": "c4d9995ca040dff4ca327f6022f30d679335998ef1968c3061102ae6ed4105ac",
"precision": 0.012361111111111111,
"recall": 0.2557471264367816,
"map50": 0.01885292400464415,
"map50_95": 0.00378065636983694
},
"controlled_e10": {
"decision": "rejected",
"model_sha256": "e4f5ffc3080dc5781868f7add07176fa7b59cc50fb947ebb9b87495f3d9f20b2",
"precision": 0.017933658902666374,
"recall": 0.32471264367816094,
"map50": 0.0222135694107064,
"map50_95": 0.005679536478252313
}
},
"production_decision": {
"active_model_unchanged": true,
"reason": "Both fresh-data candidates materially regress precision and mAP on independent V72 calibration.",
"promotion_allowed": false
},
"remaining_evidence_limits": [
"Accepted independent human label review is still pending.",
"V72 contains only 24 calibration tiles and is not a national accuracy claim.",
"Wallonia and Brussels remain diagnostic-only until PICC and UrbIS authority contracts are governed for training.",
"The current active detector remains low-accuracy on fresh regional imagery; fail-closed product messaging is required."
]
}