SOURCE-LINKED INTELLIGENCE

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

arXiv · AI, language, vision and robotics · article · Sep 9, 2026 · UTC

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 failure in the score while keeping unsupported capability out; and adjudication is structurally score-bl

Read original source ↗ Open in workspace

recordType

paper

region

Global