nucleation

CLAIM-111: A text supports a distribution over possible communicative situations, and a translation can be faithful by preserving that distribution and its response to further evidence

3 min read

Properties10
StatusProposed
Rolethesis
Warrantinterpretive
Tagsphilosophy-of-language, representation-learning
Linepragmatic-transport
AnswersQUESTION-021 — Is there a frame-independent communicative act that translations approximate, or is the act's identity constituted by a family of frame-relative realizations?
GroundsTHEORY-172 — In Frank and Goodman's one-shot reference games, listeners' mean bets matched a parameter-free listener inverting a speaker who chooses among true utterances by informativeness, with a separately measured salience prior, and not salience alone; and in Bergen, Goodman and Levy's signalling games with novel symbols, listeners read a signal as excluding what the available alternative would have conveyed
Objected byCLAIM-236 — Under the rational-speech-act model the distribution-over-situations thesis grounds on, a form's interpretation depends on what else the speaker could have said, so matching source and target posteriors needs a correspondence of alternative sets that translation changes and no entry names
ComplementsCLAIM-214 — The distribution-over-situations criterion compares each language's own listener posterior, so it needs no correspondence of alternative sets; that a literal rendering can fail it is the criterion detecting a lost implicature, and a target language may admit no rendering that passes
UsesTERM-011 — frame, as inferred from evidence

Graph View

The claim

A45 §6: P_o(z | c_o, u_o) ≈ P_t(τ_z(z) | c_t, u_t), “where τ_z maps corresponding latent communicative frames across linguistic environments. This suggests an interesting middle position between our original relativistic account and the later dynamical account. Rather than assuming that the text has one fully specified communicative identity, we represent it as supporting a distribution over possible communicative situations. Different wording and delivery contexts change that distribution. The translator’s objective can be framed as preserving some relevant structure of the distribution, perhaps including its response to further evidence.”

It answers QUESTION-021 between its horns. There is no single act, and identity is not merely a family of realizations either: it is a distribution over acts.

Where it went

A48 §4.2 kept the factorization as a “candidate explanatory model” but not as a fidelity criterion, and A52 did the same. Nothing argued against it. The mechanistic-interpretability programme A45 attached to it went at the same time: “Can one identify internal representations that systematically track inferred speaker footing or audience alignment? Can interventions on those representations reproduce the effects of changing the linguistic framing context?”

What it does not say

That z exists: “a modeling assumption. In an arbitrary neural network, no unique, identifiable latent variable z need exist.” It is compatible with CLAIM-046 if the posterior is read as a summary of observations.

A worked model

Rational speech act models (THEORY-172) treat the listener’s output as exactly this: a posterior over what the speaker meant, given the utterance and its alternatives. Frank and Goodman (LIT-821) give it for referents. Goodman and Frank’s extended model (LIT-816) gives a joint posterior over the world and the speaker’s topic, knowledge or lexicon, which is an utterance supporting a distribution over the situation of its saying.

Note of 2026-10-10: the replies

CLAIM-214 (2026-10-10): the criterion compares each language’s own posterior and needs no correspondence of alternative sets. Two consequences: a rendering with the source’s literal content can fail it, and a target language may admit no rendering that passes.

Note of 2026-10-30: what holds it back

It offers an answer to QUESTION-021, which is Open with two rival offered answers, CLAIM-137 and CLAIM-151, both Proposed. Promoting it would answer that question, which needs a showing that the distribution criterion predicts faithfulness judgements better than one act or a family of realizations does. The comparison in its defeat condition, renderings whose inferred-situation posteriors match the source’s against renderings whose posteriors differ, has not been run.

Backlinks