How does the way information reaches a transformer -- as symbolic tokens, a clean per-factor "oracle" code, or an entangled perceptual vector -- shape whether it binds that information compositionally? We study ~6-10K-parameter transformers on finite factored worlds enumerated exhaustively, so every measurement covers the whole input space and the informative routes are information-matched (exact Bayes ceiling 1.0). Four findings. (1) Endpoint invariance: on held-out binding queries no informative route reaches converged zero-shot composition -- each ends at or below chance, so the failure reflects inductive bias under a lookup-sufficient objective, not missing information. It also holds with the state in its own sequence position rather than added to every token. (2) Readability governs few-shot binding: when a small fraction of the held-out type is leaked into training, sample efficiency tracks code readability. The least readable entangled code is worst (-0.147 pooled, under both schemes); a dimension-matched control and a graded readability sweep isolate readability from input dimension. (3) The injection site, not the pathway: v1 read the remaining ordering among readable routes as pathway parameter sharing; that is withdrawn. The compared pathways are block-decomposable, hence one function class, so sharing was never manipulated; the ordering comes from the initialization scale of the broadcast-added state and vanishes when it gets its own position (+0.057 -> -0.006, 20 paired seeds/cell). (4) Failure anatomy: symbolic routes lose the answer at the readout; index routes mis-bind -- the answer stays decodable, yet an input intervention shows the output tracks the wrong slot. The central positive claim is readability; the others bound what a pathway comparison can establish. Code, manifests, and per-seed logs are released.
翻译:暂无翻译