The claim
The owner at U41, on A125’s intervention kernel: “this feels to me like an
opportunity for Judea Pearl’s do() notation.” A129 §1 wrote the structural
model (CLAIM-127) and separated three operations. The one this claim is
about is the third:
- P(Y | S = friend), which “conditions on naturally occurring situations”;
- P(Y | do(S = friend)), which “describes a hypothetical intervention replacing the structural mechanism governing the speaker’s role”;
- P(Y | do(I = “speaker is a friend”)): “This changes what the interpreter has been told. It need not change who actually spoke.”
In Pearl’s calculus (LIT-894) do(S = s) replaces S’s structural equation and leaves the others in place. do(I = a) replaces I’s. The two act on different variables, and nothing in the model makes their effects on Y equal. A129 drew the experimental consequence: “a model prompted to interpret a line as affectionate teasing is not necessarily responding to the same causal manipulation as a reader told that the speaker is an intimate friend.”
A191 §6 turned the distinction on the ontology. Its table compares four interventions, each with its question:
| Intervention | Ontological question |
|---|---|
| Change wording while retaining role and norms | Does the communicative act persist despite altered realization? |
| Change speaker authority while retaining wording | Does the act’s normative character change? |
| Change the reader’s information while retaining the actual situation | Are judgments changing, or has the communicative event itself changed? |
| Change conversational history | Are prior commitments constitutive of the present act? |
Of the third row: “If we merely tell participants that an utterance is a reprimand, and they subsequently classify it as a reprimand, we have learned something about interpretation under instruction. We have not necessarily demonstrated that the underlying speech act changed.” A203 §16: “A stronger test would manipulate actual speaker authority, conversational commitments, or the circumstances under which the utterance is performed.” The distinction recurs at A151 §3, A214 §9 and A218 Ch10.4 and Study 4.
So the effect of a cue is evidence about how interpretation responds to instruction. It bears on which relations are constitutive of an act (TERM-038; QUESTION-030) only through a do(S). The record had the words already: a frame is a property of the situation (TERM-030), and “Telling a reader that the speaker is a friend is an intervention” (TERM-018). This claim draws the consequence for what an experiment can show. It complements CLAIM-151, which says the act is partly constituted by its relations to participants and norms: only an intervention on those relations tests that.
What it does not say
It does not say the two effects differ in fact. That is the defeat condition. It does not say the structural model is identified. A129’s caption: “Schematic causal model, not a uniquely identified causal graph.” It does not say that attributed-speaker studies are useless. They measure interpretation, which is what CLAIM-115’s thesis is about.
The sharp edge is that do(S) is rarely available. In a vignette or rating study every manipulation is a change in what the reader is told. That includes the “speaker authority, interpersonal history, or the norms governing the exchange” that A214 §9 proposes to manipulate. Under TERM-018 each is a framing intervention, a do(I). A214’s “intervening on the communicative situation”, which “may change the performed act”, is not available in such a design. CASE-021’s factorial study manipulates “attributed speaker identity”, which is a do(I). A do(S) needs an interactive or field design, where the speaker’s role or the conversation’s history is actually changed, or an interpreter that is a model whose context is the situation itself.
Note of 2026-10-10: the book read in part
LIT-894 has now been skimmed in four excerpts the author posts (NOTE-696). The rest of the book was not read.
- The gloss holds. “do(S = s) replaces S’s structural equation and leaves the others in place” is what pp. 158 and 417 say.
- “do(I = a) replaces I’s” needs a qualification. The surgery deletes “the equation for” a variable (p. 417), and A129’s model, as CLAIM-127 quotes it, gives I no equation: I is an exogenous input. So do(I = a) sets I. It does not replace a mechanism unless the model says how I is chosen. The claim’s distinction is untouched by this. do(I) and do(S) still act on different variables, and nothing makes their effects on Y equal.
- Why a vignette study is a do(I). On Pearl’s account a randomized experiment is a surgery that cuts a variable’s usual link and gives it a new mechanism, a coin (p. 418). A rating study that assigns what participants are told does that to I, and to I only. That supports the claim’s “sharp edge”. A do(S) needs the same cut made on the situation itself.
Note of 2026-10-12: evidence now linked
THEORY-138 (Woodward) now grounds this claim. An intervention alone sets X and affects Y only through X, which is the principled reason do(I = a) differs from do(S = s).
Note of 2026-10-14: a caution inside do(S)
THEORY-282, filed from LIT-968 (read in NOTE-782), adds a caution inside do(S), beside this claim’s gap between telling an interpreter and changing the situation. A subspace intervention set to values off the data distribution can reproduce the output effect of changing a variable without localizing it, through a direction that cannot reach the output on the data. So a do(S) can create the very relation it is meant to detect, unless S is set to values S can actually take. It does not bear on do(I).
Note of 2026-10-28: promoted
The status review of this date promoted this claim (ADR-057). It was held back by one premise, CLAIM-127, which is still Proposed. But what this claim needs from CLAIM-127 is only that the three operations differ, and that is Pearl’s surgery (THEORY-419) and Woodward’s account of an intervention (THEORY-138), both Active. CLAIM-127’s open content is a proposal about the formalism, that an intervention should index the empirical model rather than enter its cover. This claim does not need that, so CLAIM-127 is a complement here, not a premise.
QUESTION-030 was split under ADR-049. This claim answers the part it settles, QUESTION-296, and bears on what stays open.
Note of 2026-10-28: split under ADR-061
The title’s second half, that “the effect of a cue is evidence about interpretation under instruction, not about whether the act changed”, was a verdict on what a study can show. It is now ARG-012, from CASE-021, the factorial study of attributed speakers, to CLAIM-151, Rejected. This claim keeps the distinction between do(I) and do(S), which is about the subject matter, and it undercuts that argument.