94 lines
4.8 KiB
JSON
94 lines
4.8 KiB
JSON
{
|
|
"schema_version": 1,
|
|
"generated_at": "2026-08-09T12:30:00Z",
|
|
"run_id": "building-kempen-v68-checkpoint-and-threshold-review-r1",
|
|
"status": "completed_no_production_change",
|
|
"review_type": "ai_assisted_visual_and_metric_review",
|
|
"claim_boundary": "Non-protected Kempen validation only. This is not a formal human sign-off, protected-test result, national accuracy claim or authorization to infer outside the configured geographic scope.",
|
|
"runtime": {
|
|
"device": "cuda:0",
|
|
"gpu": "NVIDIA GeForce RTX 4080 SUPER",
|
|
"torch": "2.11.0+cu128",
|
|
"ultralytics": "8.4.99"
|
|
},
|
|
"checkpoint_matrix": {
|
|
"newer_family_manifest": {
|
|
"path": "/app/storage/training/building-kempen-v68-checkpoint-matrix-r1/checkpoint-matrix.json",
|
|
"sha256": "a17640f8190915bd38bed98b2b6d086e1020987842b48bffadec2e20163f8647"
|
|
},
|
|
"original_family_manifest": {
|
|
"path": "/app/storage/training/building-kempen-v68-checkpoint-matrix-original-r1/checkpoint-matrix.json",
|
|
"sha256": "face91b3b0cb4d332dff1f176051dc46dd931275fceb91fe465ad010bbd675eb"
|
|
},
|
|
"active": {
|
|
"model_sha256": "a9088b8491dfae36694b53e9e9406cb4e3511d334a5712fa34f75078a47759c1",
|
|
"precision": 0.541626,
|
|
"recall": 0.429757,
|
|
"map50": 0.345426,
|
|
"map50_95": 0.140719,
|
|
"pure_background_detections_at_confidence_0_15": 0
|
|
},
|
|
"best_validation_challenger": {
|
|
"model_sha256": "038f1f97a6afd534f29e1f392a730a58207b928ca01e31ab8d8fed6106705820",
|
|
"precision": 0.573691,
|
|
"recall": 0.454274,
|
|
"map50": 0.368177,
|
|
"map50_95": 0.155438,
|
|
"pure_background_detections_at_confidence_0_15": 2
|
|
}
|
|
},
|
|
"threshold_review": {
|
|
"split": "val",
|
|
"image_count": 36,
|
|
"positive_image_count": 18,
|
|
"pure_background_image_count": 18,
|
|
"match_iou": 0.25,
|
|
"active_manifest": {
|
|
"path": "/app/storage/training/building-kempen-v68-reviewedexp6-threshold-r1/active-thresholds.json",
|
|
"sha256": "3462c3f4be480e34976084901c8dff1141168d818e1c6a74f86d26ae5872391"
|
|
},
|
|
"challenger_manifest": {
|
|
"path": "/app/storage/training/building-kempen-v68-reviewedexp6-threshold-r1/reviewedexp6-thresholds.json",
|
|
"sha256": "c810aa0372a2767d538c9454f6f9f7a54338e61650756e35fdd1750a57684756"
|
|
},
|
|
"confidence_0_25": {
|
|
"active": {"tp": 3830, "fp": 2584, "fn": 1856, "precision": 0.597131, "recall": 0.673584, "f1": 0.633058, "pure_background_detections": 0},
|
|
"challenger": {"tp": 4005, "fp": 2774, "fn": 1681, "precision": 0.590795, "recall": 0.704362, "f1": 0.642599, "pure_background_detections": 0}
|
|
},
|
|
"confidence_0_30": {
|
|
"active": {"precision": 0.713823, "recall": 0.543088, "f1": 0.616860, "pure_background_detections": 0},
|
|
"challenger": {"precision": 0.697588, "recall": 0.595146, "f1": 0.642308, "pure_background_detections": 0}
|
|
}
|
|
},
|
|
"visual_review": {
|
|
"active_contact_sheet": {
|
|
"path": "/app/storage/training/building-kempen-v68-reviewedexp6-visual-r1/active/candidate_error_contact_sheet.png",
|
|
"sha256": "609cffab8d5f47d93ec482a3ac1b960dac35645f65906ce4344cd50e9742b5f3"
|
|
},
|
|
"challenger_contact_sheet": {
|
|
"path": "/app/storage/training/building-kempen-v68-reviewedexp6-visual-r1/reviewedexp6/candidate_error_contact_sheet.png",
|
|
"sha256": "65ac7a24ee2b51bcbaa83d5972639602b4b4d5fd71a0a3aab2ef99510b5465f6"
|
|
},
|
|
"observations": [
|
|
"The challenger reduces missed-building pressure in both dense Turnhout and lower-density Westerlo examples.",
|
|
"Dense urban overlays still contain many low-IoU disagreements caused by visible roof position and shape versus GRB ground-level building footprints.",
|
|
"AI-assisted visual inspection cannot convert ambiguous roof-footprint disagreements into accepted human labels."
|
|
]
|
|
},
|
|
"production_gate": {
|
|
"attempted_action": "Run the existing seven-zone production-API benchmark with challenger confidence 0.25.",
|
|
"result": "blocked_fail_closed_before_inference",
|
|
"reason": "The historical challenger predates the mandatory immutable runtime-provenance sidecar and governed database snapshot binding.",
|
|
"unsafe_bypass_used": false,
|
|
"production_model_changed": false,
|
|
"active_model_sha256_after_decision": "a9088b8491dfae36694b53e9e9406cb4e3511d334a5712fa34f75078a47759c1",
|
|
"active_image_after_decision": "geointel-all-in-one:0209167cfd37-wip25e7de62cde9-ai",
|
|
"server_health_after_decision": "healthy"
|
|
},
|
|
"decision": {
|
|
"v67_candidate": "rejected",
|
|
"reviewedexp6_at_confidence_0_25": "retain_as_unpromoted_validation_challenger",
|
|
"reason": "The challenger is measurably stronger on the non-protected validation split and clears its local background controls at 0.25, but production promotion remains blocked until its original training lineage is migrated into an exact sidecar and governed source snapshot and the existing seven-zone workflow is rerun."
|
|
}
|
|
}
|