Local Interface Descriptions Do Not Compose
Representation-Conditioned Access to a Non-Local Predicate
88% structured vs 46% flat accuracy; the same tool gets different hiddenness judgments depending on composition partner; models revert to lexical proxies without relational framing
A frontier LLM achieving exact hiddenness computation (100% across prompt conditions) from schema inspection alone, or a flat prompt format matching structured accuracy
Abstract
We identify a class of compositionally relevant predicates — hidden conventions — that are not recoverable from local interface descriptions alone. The coherence fee measures the number of independent convention dimensions lost when tools are composed. We prove that hiddenness is a graph-relative predicate: the same tool has different hidden sets depending on its composition partner. In a controlled synthetic benchmark, frontier LLMs recover this non-local object with 88% accuracy under structured relational prompts but collapse to 46% under flat representations, reverting to lexical heuristics. The effect is representation-conditioned: it requires both relational framing (+33pp) and convention-specific task language (+33pp). Cross-model replication reveals divergent failure modes. Bulla computes the non-local object exactly across all prompt conditions.