Every framework that minimizes uncertainty — Shannon's channel coding, active inference, Bayesian decision theory — presupposes that sender and receiver share a codebook. This paper proves that no quantity computable from Shannon's model detects codebook misalignment, and shows that this gap extends to every downstream framework. When systems cross a boundary without verifying their codebooks, the receiver proceeds with zero uncertainty about an interpretation that may be wrong — a condition we call ignorance of incoherence. We provide the instrument that fills this gap: a coherence protocol that decomposes the codebook into three comparison facets — Meaning, Structure, and Data — and a resolution layer, Context, that accumulates what was missing, surfaced, and resolved across the other three. The protocol is grounded in an irreducible five-column canonical claim form onto which any system can project its codebook at the moment of contact, requiring no prior agreement on terms. The protocol produces three actions — Halt, Ask, and Act — where Ask is a measurement in the formal sense: it acquires information that no computation on the received data can produce. The cost is O(1) per facet comparison per boundary crossing. The substrate reduces to binary and admits tensor decomposition, opening a path toward self-referential inference on the same surface that performs measurement. This paper serves as the theoretical foundation and operational north star for the W3C Context Graphs Community Group.