The inference
CASE-040 is the design A218 “names as the first experiment to run”. It crosses topic (spending, relationships) with act (teasing, reprimand), and holds out critical items that “break the familiar correlation”. CLAIM-129 is defeated if, on those items, “a topic-only factorization of lexical co-occurrence recovers the act distinction as well … as factorizations fitted to interpretive-judgement or uptake data do”. So a pass, with judgement factorizations recovering the act and topic-only NMF failing, would be evidence that a factorization recovers only what its relations make visible.
Critical questions
- Would the evidence be expected if the hypothesis were false? Yes, and this is the question that fails. CASE-040 “makes the act a manipulated factor rather than an outcome”, and “every cell is an example label, not a stimulus”. Whoever writes the items decides which are teasing and which are reprimand, so a factorization of judges’ responses that recovers the act “recovers the writer’s intent as the judges read it. That is close to a manipulation check.” And a joking item written with admonitory words has, by design, the vocabulary of a reprimand, so topic-only NMF must get it wrong. CLAIM-203: “CLAIM-129’s defeat condition can then scarcely be met, and a pass shows that the items were built as intended.”
- Does it tell the hypothesis apart from its rivals? Not against the rival a critic would field. The baseline is “topic-only NMF”. CLAIM-139’s own defeat condition names “topic-factor, embedding or surface similarity”, and A218’s Chapter 15, as CLAIM-078 quotes it, asks for prediction “beyond standard lexical, embedding, and factorization baselines”. A supervised text classifier or a language model reads wording in context, and is not in the design.
- Does the measure measure what it is said to? The act recovered is the item writer’s assignment, not a structure that the judgements reveal and the words hide. CASE-040’s “What it cannot show” names two other limits, that it has not been run and that its interventions manipulate a described speaker (CLAIM-127, CLAIM-130). It does not name these.
Rejected as an inference from CASE-040 as filed, with constructed items and a topic-only baseline. The record conceded on 2026-10-10 that for CLAIM-129, “a claim of possibility, constructed items are a legitimate existence test”. CLAIM-129 v2 adds naturally occurring items labelled by raters who did not select them, and CLAIM-204 places the strong text-model rival in the central tests. As a first pilot the design still shows whether judges and models can be made to separate topic from act at all. This argument does not say CLAIM-129 is false. A topic is plausibly not an act.
What it reaches. CASE-040 illustrates CLAIM-129 and CLAIM-135, and CLAIM-135 is a premise of CLAIM-139’s epistemic clause. The same verdict would hold for an argument from CASE-040 to CLAIM-135. It is not filed separately.
The lesson in general form. It recurred three times in the exchange “Experimental Infrastructure Review” (T21 to T23), about latent-structure models of an organization’s members (QUESTION-276). T22: “If we start generating employees from the IBP and then successfully recover the IBP structure, we’ve mostly demonstrated that inference works under matched assumptions.” The remedy was to fit the model to a mechanistic simulator that does not assume it, and to keep planted-structure recovery as a check on inference only (CASE-165). The hLDA paper has the published form of that check (THEORY-1010). Recovering a structure from data generated under its own prior tests the inference, not the structure.
What would answer it
- Acts fixed independently of the item writers: naturally occurring items whose act is labelled by raters who did not write them, or by observed uptake.
- Critical items whose vocabulary is not chosen to mislead the baseline, or a report of how the baseline does on items where the lexical cue is neutral.
- A baseline at least as strong as a supervised text classifier or a language model, alongside topic-only NMF.