Three matched-control gates localize long-context retrieval failure in sparse attention to a single locus- delivery. Attention redirection acts as transport, while stored-state correction is not reusable. Delivery works reliably, is content-addressable (no positional decay), and achieves a 27-byte attribute→code binding. Mechanism: RAG micro-hints.

Umbrella review: four pre-registered matched-control gates (attention, delivery, synthesis, latent state) show that 1B–3B LLMs use long context only via external buffer. No latent computation or composition. A map of the boundary.

We optimize a compact latent state (frozen weights) to force failed multi-hop chains to output the missing answer D. 5 pre-registered controls show it simply injects D: carries it without the code-fact, leaves intermediates invisible, inert to hop corruption, and doesn’t transfer. No latent composition at 3B (Llama-3.2-3B, Qwen2.5-3B).

Synthesis should break the “everything is transport” hypothesis: D must be assembled from absent facts. Cross-model tests (Llama-3.2-3B, Qwen2.5-3B) show it doesn’t. Implicit multi-hop reduces to retrieval + elimination. True composition appears only as externalized CoT. At 3B scale there is no internal computation of absent facts.