Files
geointel/artifacts/evidence/accuracy/model-training/20260830-independent-ai-visual-review.json
T

85 lines
4.5 KiB
JSON

{
"schema_version": 1,
"review_id": "20260830-independent-ai-visual-review",
"reviewed_at": "2026-08-30T00:00:00+02:00",
"reviewer_kind": "independent_ai_visual_inspection",
"human_reviewer": false,
"claim_boundary": "Visual inspection evidence only. This is not a human acceptance ledger, protected-test result, accuracy measurement or model-promotion approval.",
"inputs": [
{
"path": "artifacts/evidence/accuracy/model-training/20260809-v72-fresh-calibration/contact_sheet_001.png",
"sha256": "f7734c5a9db8837adeda7f8e0f1a2942148330abfcb588d41dc245017ec2435a"
},
{
"path": "artifacts/evidence/accuracy/model-training/20260810-v73-flanders-remediation/contact_sheet_001.png",
"sha256": "4e092afac6bbff820908f73acd2a29b6d327d7476c3b2ae5001302a0bde4a24f"
},
{
"path": "artifacts/evidence/accuracy/model-training/20260810-v73-flanders-remediation/contact_sheet_002.png",
"sha256": "ef4e4a9663b4ac1920bee76bf0ca40c69d6232ddeb40d75c6963a17540b8d211"
},
{
"path": "artifacts/evidence/accuracy/model-training/20260810-v74-root-cause/active-error-contact-sheet.png",
"sha256": "bb5e4809802643cc8b673a86f3027ed06aa0430fbe79b1d33718c21d5feb269d"
},
{
"path": "storage/operator-data/model-review/reviewedexp6-corpus/full-ai-review-r1/contact_sheet_001.png",
"sha256": "688b97a6e82e6b906b7966ed3e2b43685786e89d50549188bfdcc3837d042eb2"
},
{
"path": "storage/operator-data/model-review/reviewedexp6-corpus/nested-pairs-review-r2/contact_sheet_001.png",
"sha256": "2a6bdd542b8ac49bcd5ac4d9c84999982461463ddd3c2fe0dc7332ceeb7a5e60"
},
{
"path": "storage/operator-data/model-review/reviewedexp6-corpus/tile-edge-review-r1/contact_sheet_001.png",
"sha256": "5d96c5a0a732077ed16b22d33bed998b37bd33b5536b1491e778999a2a6638dd"
}
],
"findings": [
{
"severity": "blocker",
"category": "reference_geometry_semantics",
"observation": "Many yellow reference rectangles are coarse axis-aligned extents rather than roof or building footprints; several include vegetation, roads, fields or multiple structures.",
"impact": "A detector trained against these boxes is penalized for geometrically correct roof localization and rewarded for oversized detections."
},
{
"severity": "blocker",
"category": "dense_and_nested_labels",
"observation": "Dense urban and industrial samples contain strongly overlapping, nested and grouped boxes without an unambiguous single-object labelling rule.",
"impact": "Label conflict and inconsistent object granularity make precision, recall and NMS behaviour unreliable."
},
{
"severity": "critical",
"category": "tile_edge_and_visibility",
"observation": "Multiple labelled structures are clipped by tile boundaries or only partly visible, while edge inclusion rules are not consistently evident.",
"impact": "This creates avoidable false-negative and duplicate-detection pressure at inference tile seams."
},
{
"severity": "critical",
"category": "hard_negative_coverage",
"observation": "Industrial rails, paved surfaces, tree canopy and tree-shadow remain recurrent false-positive contexts in the active-error sheet; representative pure-background coverage remains insufficient, especially for Brussels.",
"impact": "The candidate has no defensible low-false-positive operating point across Belgium."
},
{
"severity": "critical",
"category": "model_error_balance",
"observation": "The active-error review shows simultaneous false positives and false negatives across industrial, vegetated and sparse-rural contexts.",
"impact": "Threshold tuning alone cannot repair the observed error pattern."
}
],
"decision": {
"corpus_accepted": false,
"training_gate": "blocked",
"promotion_allowed": false,
"human_review_satisfied": false,
"required_before_next_governed_training": [
"re-extract building labels from task-appropriate authoritative geometry with an explicit object-granularity rule",
"resolve dense and nested labels and publish a checksum-bound review ledger",
"apply deterministic tile-edge visibility and ownership rules",
"add independent pure-background and hard-negative strata for Flanders, Wallonia and Brussels",
"prove AOI and derived-image independence across train, validation, test and challenge splits",
"obtain representative human acceptance before opening protected evaluation"
]
}
}