The claim
The owner at U41, on A125’s proposal of “an intervention kernel K_a acting
on a communicative or interpretive state”: “this feels to me like an
opportunity for Judea Pearl’s do() notation.”
A129 §1 answered with three operations (LIT-894):
- P(Y | S = friend), which “conditions on naturally occurring situations in which the speaker is a friend”;
- 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.”
The first is conditioning. The second and third are interventions on different variables. The experimental consequence of the difference between them is CLAIM-130.
A129 §1 writes them in a structural model. Z = (U, S, A, K, N, G, H) is the communicative situation, R the interpreter’s state, I “externally supplied framing information or instructions”, C a measurement protocol (TERM-039) and Y the judgement:
- R′ = f_R(R, Z, I, ε_R),
- Y = f_Y(R′, Z, C, ε_Y).
Its diagram’s caption: “Schematic causal model, not a uniquely identified causal graph.” It is to be “a separate causal layer, not a replacement of the sheaf layer.”
A129 §5 puts the layers together as a causally indexed observational system 𝔠 = (𝒮, 𝒜, ℳ, ℰ, {e_C^a}, 𝒬), with “e_C^a: empirical distributions in context C under intervention a”, and e_C^a = P(Y_C | do(a)). The review the owner relayed later the same day restates this: “make causal interventions and observational contexts separate indexed structures from the outset: e_C^{do(a)} = P(Y_C | do(a)). That would connect our recent Pearl-inspired refinement to the sheaf structure without treating observation, conditioning, and intervention as the same operation.” The refinement it names is U41 and A129.
The record already draws the distinction in words. A frame is a property of the communicative situation (TERM-030); a framing intervention is “an action that changes the conditions under which an utterance is interpreted” (TERM-018). Telling a reader that the speaker is a friend is not the same as the speaker’s being one. What the record lacks is a place for the distinction in the formalism. In the sheaf-theoretic setting a context is a set of jointly measured observables, and the empirical model assigns each a distribution. Nothing there says whether an elicitation only reads the interpreter’s state or changes it. In A129 the cover ℳ is the same for every intervention; what varies with a is the family {e_C^a}. That an intervention might also change the cover is the record’s extension, not A129’s, and nothing in the exchange argues it. Transport is then compared across a family of models indexed by intervention, and no intervention is a context inside one of them.
A129 also derives the framing kernel instead of positing it: K_a(r′ | r, z) = P(R′ = r′ | R = r, Z = z, do(I = a)). That gives TERM-026’s operators and CLAIM-052’s noncommutativity (TERM-024) a causal reading. Its equations have one step, so K_bK_a needs time-indexed states, which it does not write.
Why it matters for the open problems
This part is the record’s extrapolation, not the review’s. CLAIM-125’s second open problem is that language data signal. In the corpus data Wang et al.’s journal follow-up measured, 69 of 90 noun–verb systems signal. The manuscript’s experiments elicit judgements instead, and A35 §6 had already warned that “forcing a participant to make an explicit judgment is itself an intervention.” If an elicitation is an intervention, then reading one observable can change the interpreter’s state. That change can show up as a dependence of another observable’s marginal on context, which is signalling. The do-index gives this hypothesis a form that can be tested: models elicited under different interventions are compared separately, not pooled into one cover, and the elicited data can be checked for signalling beyond the corpus baseline.
Two routes to signalling
A129’s graph lets C enter only f_Y. So in A129’s own model an observable’s marginal can depend on its context with no change to R. That is a direct influence of context in Dzhafarov’s sense (THEORY-177). The route the extrapolation above assumes is elicitation changing the interpreter’s state. A35 §6, quoted in TERM-018, says “forcing a participant to make an explicit judgment is itself an intervention”. That route needs an arrow from C into f_R, and A129’s graph has none. Which route holds is empirical. Sequential designs, where one judgement is elicited before another, separate them (TERM-039).
What it does not say
It does not say the signalling in corpus data comes from elicitation. Corpus estimates involve no elicitation at all. It does not say interventions violate the no-signalling condition. Nor does it say Pearl’s calculus applies as it stands. A129’s model is schematic, one-step and not identified, and the record files no causal model of the interpreter.
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 contents, the second-edition preface, §§5.3.2–5.4.1 and the Epilogue. The rest of the book was not read.
- Version 2’s correction holds. In those pages observing a value is conditioning on it. The conditional bar reads “given that we see” (p. 421), E(Y | x) is “the standard conditional or observational expectation” (p. 158), and a passive observation “should not violate any of the equations” (p. 158). The Epilogue names two components of science, “Observations and interventions” (p. 421). The review’s three-way list “observation, conditioning, and intervention” is not Pearl’s.
- The gloss of do() holds. Under do(X = x) the equation for X “should be overruled and replaced with the equation X = x”, and the others are solved as they stand (p. 158; the “surgery” of p. 417).
- A sharpening the claim should carry. Surgery deletes “the equation for” the variable (p. 417). In A129’s model as quoted above, I is “externally supplied” and has no equation of its own. When an experiment assigns I and nothing else causes it, intervening on I and conditioning on its assigned value give the same distribution. That follows from the calculus’s second rule (p. 423) and is the point of Pearl’s account of randomization (p. 418); the inference is the reader’s, not a sentence of the excerpts. So for I the do() notation separates something only where I is not assigned, as in corpus data. The claim’s case for indexing e_C^a by a, and not adding a to the cover, does not rest on that difference. It concerns where a sits in the formalism, and it stands.
- Elicitation. §5.4.1 separates conditioning on a measurement taken under an intervention, P(y | do(x), z), from intervening on it, P(y | do(x), do(z)) (pp. 160–161). Whether an elicitation C is the first kind or the second is the question this claim’s “two routes to signalling” leave open.
Note of 2026-10-30: what holds it back
The defeat condition’s second branch appears met for the manuscript’s experiments. There every framing is an assigned do(I) (CLAIM-130), do(I = a) equals conditioning on I = a (the note of 2026-10-10), and with I in every context the compatibility, signalling and contextuality verdicts equal the per-a verdicts (checked on the 4-cycle at the status review of 2026-10-30). What the do-index still separates is the frame entered as separate contexts of one cover, which turns a framing effect into signalling. Until the title and defeat condition are revised to that contrast, or to unassigned framings, the claim cannot be promoted.