benchmark-methodology

Evaluation governance methodology for domain-specific AI doctrinal benchmarking. Derived from 7 AI safety publications. SAICRED v2 reference implementation.

Protocol outweighs method: patch sampling, evaluation scope, consensus level and threshold dominate architecture and supervision-target choice in lung nodule segmentation on LIDC-IDRI.

Local-first AI governance gateway for controlled agent/provider execution, audit receipts, cost signals, and public QBS benchmark methodology.