SecureCoating Meta

SecureCoating Meta — Know When AI Should Not Decide SecureCoating Meta is an evidence-grounded AI inspection prototype for battery-electrode coating defects. It investigates a critical industrial AI problem: how should an inspection system respond when its models produce incomplete, conflicting, or insufficient evidence? The project extends a disclosed, pre-existing YOLO-based inspection prototype with a new Meta SAM 3.1 integration, segmentation evidence processing, detector–segmenter reconciliation, and auditable decision reports. A key finding emerged from real experiments: successful API execution does not guarantee useful defect coverage. Across three completed Meta SAM 3.1 requests, one box-conditioned prediction achieved a full-frame IoU of 0.0071 against the published reference mask, while two concept-prompted predictions returned empty masks despite nonempty defect annotations. These failures exposed a potentially unsafe decision path: when both the detector and segmenter return empty results, their agreement alone must not be interpreted as proof of a defect-free surface. We implemented a conservative mitigation in the new Meta inspection pipeline: dual-empty or insufficient segmentation evidence results in HOLD rather than automatic clearance, including under simulated factory readiness. Historical inference results are preserved for reproducibility. The project includes saved real Meta API evidence, an offline replay workflow, provenance-linked reports, regression tests, and a reproducible demo. Factory actuation remains disabled. The earlier Tsinghua-associated prototype and a separate OpenCV 2026 competition entry are explicitly disclosed. We do not claim improved segmentation accuracy, calibrated uncertainty, or production readiness.