The claim
A97 §3: a Schrödinger bridge Q* = argmin KL(Q ‖ P) with fixed endpoints treats translation as “a path through a space of progressively reconstructed representations. The path itself becomes observable. We can investigate which constraints survive early transformations, which are introduced later, and whether particular communicative relations emerge only through successive refinements.”
Where it went
Manuscript §7: “Schrödinger bridges offer a complementary path-space viewpoint … Such constructions can model guided stochastic transport and conditional simulation (Shi et al., 2022).” Nothing is observed along the path. It is the diffusion-side counterpart of the pragmatic phase change (CLAIM-114). It is not the same as historical drift (CLAIM-071).
The crystallized argument’s §17 gave the rationale its last form: “the bridge perspective is useful because intermediate states may matter. A generative process can gradually establish scene geometry, objects, attributes, and relationships. We can investigate the trajectory by which a relational configuration emerges or fails.” The manuscript keeps the bridge and the disanalogy of times, and drops the reason, so bridges enter §7 with no stated purpose.
Note of 2026-10-30: what holds it back
Its defeat condition is met as written: given the final state, an intermediate state carries no information about the final relational content beyond what that endpoint carries, in any chain whatever. What the claim means is that the order in which relations are fixed along the path is informative, and no result in the record measures that for a reconstruction chain. THEORY-948’s early-phase layout control is image generation under intervention, not an observation of a translation path. Until the condition is restated at that level, which is the owner’s call as for CLAIM-029, it is not promoted.