IBIB: A Protocol for Measuring Enterprise AI Systems by Serving Route, Not Model Identifier

AI Digest - ArXiv AI

IBIB: A Protocol for Measuring Enterprise AI Systems by Serving Route, Not Model Identifier

Enterprises deploy systems, not checkpoints. Usable capability depends jointly on weights, serving route, precision, output contract, and harness, yet all 18 audited benchmarks score advertised model identifiers. We treat this as measurement error and give a protocol that makes it reportable. It has three parts. A gold-blind capability-binding preflight verifies that a route can execute the evaluation contract before any task reaches it; a reliability-inclusive first-pass scoring rule keeps fail

Source: ArXiv AI