The claim
A24: “we evaluate whether the translation preserves the pattern of context-dependent communicative judgments, rather than whether it preserves one canonical vector of meaning”, with D(τ) = Σ_C w_C d(e_C^source, e_τ(C)^target) (TERM-021). Manuscript §6: L_obs(T) = Σ_C w_C d_C[(K_C)#e_C^o, e_τ(C)^t]. §10 Case I’s success criterion is its defeat condition turned round: “incremental prediction of held-out human communicative-fidelity judgments”.
The form that reached the manuscript is A78’s D_sheaf (U29): Σ_C w_C d_C(T_C# e_i(C), e_(i+1)(τ(C))), “a graded, directed measure of preservation of local observational structure”, stated over the sheaf scenario’s contexts (CLAIM-105).
What it does not say
A24’s caveats: “the outcome categories must be made comparable across contexts, and the correspondence between contexts cannot be assumed.” It is not the whole of fidelity: the manuscript adds a structural term and a decision term, and the dynamical requirement is CLAIM-117.
Note of 2026-10-10: the contexts must be as fine as the pairing
A181 §4, after the owner’s U50 pointer to FID: a generator may produce “the correct overall proportions of smiling and stern-looking faces, but assign[] them to the wrong prompts”, and “Even a perfect measure of the unconditional image distribution cannot detect a generator that preserves the overall distribution while breaking the association between prompts and images”. An unconditional distance cannot see a transformation that keeps the output distribution and breaks the pairing of inputs with outputs.
D(τ) is context-indexed, so it is not unconditional. But the failure returns inside a context. Within any context C that pools several source items, a target that permutes their renderings leaves e_τ(C)^target unchanged. So the contexts must be at least as fine as the pairing that fidelity is meant to track. This is elementary, and the manuscript’s L_obs inherits it. The case beside it is CASE-041.
Note of 2026-10-10: the replies
After the audit (CLAIM-207, CLAIM-230, CLAIM-256): for one item, L_obs can be made zero by a constant kernel, so it is restated over a class of items with one kernel family, fitted on some items and scored on others (CLAIM-228), and over both covers, with the target’s new questions asked of the source too (CLAIM-257). Its distinctive test against a static pragmatic profile is CLAIM-202. The manuscript’s text quoted above is unchanged.
Note of 2026-10-29: narrowed
The title now carries CLAIM-207 and CLAIM-256 instead of leaving them to the note of 2026-10-10. For one pair the kernel can make the match perfect whatever the translation did, and what a rendering adds is in no term. The contrast with one canonical meaning vector stands.
Note of 2026-10-30: what holds it back
The defeat condition still cannot tell this claim from its rivals. Version 3 fixed a margin and a power before the data (CLAIM-238), a named population of judges (CLAIM-226) and one outcome correspondence across items (CLAIM-207), but it still sets the distance only against embedding or propositional similarity. Two rivals the record holds, Nida’s dynamic equivalence and a proposition with its illocutionary force, also predict a pass (CLAIM-233, narrowed on 2026-10-30), and so does Gutt’s interpretive resemblance (CLAIM-501, Active). CLAIM-244 (Proposed, unanswered) objects that the condition takes a task-free holistic judgement as ground truth. Restating it is put to the owner, with the other defeat conditions queued for restatement.