26  Model Cards / ML Governance

Audience: Kav AI ML team + technical-diligence reviewers + auditors. Purpose: document each detection model’s scope, evaluation, calibration, and drift so the PRD’s accuracy claims are governable, not just asserted. Reference: Google “Model Cards for Model Reporting”; Hugging Face model-card format. Pairs with the validation methodology doc in ../../evaluation/.

26.1 One card per model (RGB, thermal, OGI, cross-source engine, OOD detector)

• Model details — name, version ID, owner, date, architecture.
• Intended use — modality, asset classes, what it must NOT be used for.
• Training data — campaigns/sources, size, labeling process, known gaps.
• Evaluation — TPR/FPR vs. the PRD targets, on which benchmark, with NDT ground truth.
• Calibration — confidence-calibration curve; how/when regenerated (PRD pilot Weeks 9–10).
• OOD behavior — how out-of-distribution inputs are flagged/routed.
• Drift & retraining cadence — triggers, monitoring (ties to FR-AI-04).
• Limitations & ethical/safety notes — HITL requirement; Filter Skill > Stage 3.5 > cross-source.
• Versioning/rollback — per PRD AI Engine section.

These cards are how the TPR > 98% / FPR < 2% claims become defensible. The numbers must match the validation methodology doc and the PRD Performance table.