The case
A214 §8 crosses two factors: topic T ∈ {spending, relationships} and pragmatic act P ∈ {teasing, reprimand}. “The important point is to avoid perfectly correlating topic and pragmatic function. Otherwise, a topic model may appear to recover pragmatic identity merely because the dataset makes the two inseparable.”
The four cells, as A214 labels them:
| Teasing | Reprimand | |
|---|---|---|
| Spending | Friends joking about an empty wallet | An authority figure criticizing overspending |
| Relationships | Friends affectionately teasing about romance | An authority figure condemning a relationship decision |
The data have four sources: lexical co-occurrence, interpretive judgements (“tone, stance, solidarity, and authority”), expected uptake (choices among continuations) and intervention responses (“changes under experimentally manipulated speaker roles, histories, or framing information”). Five models are fitted: topic-only NMF, a factorization of the pragmatic responses, a joint lexical and pragmatic factorization, a shared-and-private factorization, and a richer context-sensitive model. The critical items are held out and break the familiar correlation: “test whether the model correctly distinguishes a joking use of admonitory vocabulary from a serious reprimand.”
There are four outcomes: reconstruction error, latent recovery, cross-realization transfer and held-out decision performance. “These outcomes should be reported separately.”
A218 Part III makes it the first experiment: “hold topic and pragmatic act independently variable, collect both lexical observations and human response judgments, and compare models on cases where thematic similarity and communicative identity disagree.” A218’s Study 2 adds “perturbations that break correlations between topic and pragmatic act”.
The cross-corpus version. A210 §5, after the owner’s U58, runs the same contrast through translation. Topic models are fitted to a source corpus and to its translation, paraphrase or adaptation, and the two structures are compared after each is independently anchored. There are three possible outcomes. Topics change coordinates or labels but keep their anchored associations. Topics “merge, split, or become less distinguishable”. Or “Topic associations appear stable while pragmatic or causal relationships change substantially.” The third would show “that recovering thematic structure is not the same as recovering communicative-act structure.” Raw factors of the two corpora are not comparable as they stand (A210 §4: “Their raw NMF factors are not directly comparable”). Comparing them needs paired anchors (THEORY-175), and without those a flexible alignment can make any factors match (CLAIM-077).
How it differs from CASE-021
CASE-021 crossed wording, attributed speaker and the order of judgement questions, with Henley’s renderings as stimuli. This design drops the order factor. It makes the act a manipulated factor rather than an outcome, where CASE-021 manipulated the attributed speaker and measured the judged act. It adds collaborative-filtering baselines over interpreters. Its spending row is CASE-004’s conditions A and B (shared complicity, normative authority) in text rather than image, again without condition C.
What it can show
If run, whether a structure identified from lexical co-occurrence carries the act (CLAIM-129), and how far identification falls short of a kind (CLAIM-135). A synthetic form, with planted latent factors that are recoverable, ambiguous or nonidentifiable (A214 §12, Experiment A), gives a ground truth for QUESTION-027.
What it cannot show
It has not been run, and every cell is an example label, not a stimulus. Its “intervention responses” manipulate a described speaker or history. In a judgement study that is a framing intervention (TERM-018) on the interpreter (CLAIM-127, CLAIM-130), not a change to the performed act. So the design can show how interpretation responds to information about the situation, and not whether the act itself changed. Nor does a clean result bear on the ontology. A218: “That doesn’t prove OSR, but it establishes the kind of empirical result an OSR-inspired theory should explain.”
Note of 2026-10-10: the replies
After CLAIM-203: the constructed critical items show possibility, which is what CLAIM-129 claims; whether judges recover the act from naturally occurring text needs the naturally occurring items CLAIM-129 v2 adds. A report of the models on items whose lexical cue is neutral should accompany the critical items.