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…
We haven't written up this one. arXiv cs.AI has the full story — the link below goes straight to it.
