The claim
A topic model fitted to document–word co-occurrence recovers what a text is about. It does not thereby recover what the text does. A210 §5: “A document can remain about spending, romance, or morality while changing from affectionate complicity to reprimand.” So in a translated corpus the outcome A210 calls “especially important” is possible: “Topic associations appear stable while pragmatic or causal relationships change substantially.” It “would show that recovering thematic structure is not the same as recovering communicative-act structure.”
A214 §3 gives the reason in general form. “A two-topic model might capture the first extremely well and ignore the second”, where the first is topic (spending against relationships) and the second is stance (affectionate joking against moral condemnation). Low rank “does not tell us whether those dimensions correspond to constitutive relations.” The failure is “treating one recovered structure as exhaustive of the object.” Its boxed statement: “Relational kinds are identified relative to a family of interactions and probes. A factorization recovers structure visible through that family, not automatically every relation constituting the phenomenon.”
The record holds both halves already, in other terms.
- That renderings can share a situation model and differ in footing and force is CLAIM-061, and that a difference between translations can change the act and not only the frame is CLAIM-004. A topic is a coarser thing to share than a situation model.
- That identity is relative to the family of probes is TERM-014’s “relative to a chosen family of measurements”, and that agreement on a restricted family is not structural identity is CLAIM-142.
What this claim adds is the consequence for method. A communicative kind (TERM-040) must be tested against probe families chosen for their bearing on the act, not inferred from whichever matrix is to hand. A214 §4 names such families: document × word, realization × interpreter, realization × pragmatic judgement, realization × possible response, realization × narrative event, intervention × judgement change. CASE-040 is the test. It crosses topic with act so that the two are not confounded, and it scores each factorization on items that break their usual correlation.
The claim recurs across the stretch. A203 §37 lists it as the first thing a strong result would show: “the model distinguishes shared subject matter from shared communicative identity”. A214 §8 builds the laboratory on it. A218 makes it hypothesis H3: “Topic-preserving transformations will sometimes reduce pragmatic or causal fidelity”. It is a factorization-side form of what CLAIM-115 says survives translation, which is distinctions a receiver can act on, not wording or proposition.
Shared and private factors
A214 §4 proposes fitting the probe families jointly, with a shared and a private part, X_j ≈ Z_shared H_j + Z_private,j B_j. “The shared component is a candidate structural invariant.” And: “An English utterance and a French rendering will necessarily possess language-specific structure. Fidelity should not require erasing it.”
The proposal needs a caveat A214 does not give. The split between shared and private parts is identified only under constraints, such as the orthogonality of the shared and private subspaces in JIVE-type models. Without them, components can be traded between the shared and private terms. And with unconstrained real factors, a shared Z keeps the same general-linear gauge, now acting on all views together (CLAIM-148). Coupling the views does not remove it. So “candidate structural invariant” means nothing until the constraint is stated.
What it does not say
It does not say lexical structure is irrelevant to the act. It says that lexical structure is not sufficient evidence of the act. It does not say that a richer factorization, of task-indexed judgements or of uptake, cannot capture acts. The claim concerns what one family of relations makes visible. Nor does it say that a shared component found across probe families is an invariant of the kind.