The claim
U17: “I’m wondering if leveraging an LLM (and how it responds to a context, i.e. functionally a parameterized frame) as a motivating example, maybe even just for notational purposes. there might also be valuable connections or even experiments in the ICL/TTT literature we could draw from.”
A43 made the model “the primary formal example”, with Henley as “the motivating humanistic example”, and the model as “a parameterized interpretive instrument, rather than being treated as a definitive judge of meaning”. A49’s c_f = Γ(f, c_0) (TERM-025) is the claim in notation.
Where it went
A48 kept the model as running case B, an “operational proxy”. A52 kept it as running case C. Its use as evidence is limited by CLAIM-085. The manuscript keeps language models as one kind of interpreter in its empirical programme (Abstract, §10) and keeps one consequence (CLAIM-118), but no longer presents a model’s context-conditioning as a frame. C6 §6 (U32) demoted it: ICL and TTT “are compatible with our theory but not required for its mathematical definitions”, and the notation P_θ(y | c, u) is gone from the manuscript.
What the in-context-learning readings say
- For the frame reading. Min et al. (LIT-814) find a context’s effect is carried by input distribution, label space and format rather than by correct labels. Xie et al. (LIT-797) give the formal sense in which a context fixes a posterior over a latent concept.
- Toward the defeater. Lu et al. (LIT-833) find that order alone, with content fixed, moves accuracy from near chance to near the state of the art. Good orders do not transfer across model sizes, so the “frame” is sensitive to surface form.
- A qualification of that defeater. Kirsanov et al. (LIT-946, read in NOTE-789; THEORY-251) find that across draws of examples the class representations barely move while accuracy swings widely: the sensitivity to example selection sits in the vocabulary readout. If the frame is the representation a context sets, it is more stable than the accuracy suggests. Order was not varied, only which examples were drawn, so this qualifies Lu et al.’s point without settling their case.
Prior statement
Shanahan, McDonell and Reynolds’s role play (LIT-460, read in NOTE-370) states this idea most clearly before the manuscript. The prompt selects a role from a superposition of characters, and each reply narrows the set. THEORY-490 states that framing as part of its account of what a chat agent’s “I” refers to.
Note of 2026-10-15: three observations the defeater runs together
A3 of the exchange of 2026-10-14 (workbench entry of 2026-10-15) found this claim and separated three observations: a prompt changes the output distribution; a prompt changes an internal task or interpretive representation; two prompts, or a prompt and an activation intervention, induce equivalent pragmatic behaviour over a stated family of contexts. “The first does not entail the second, and neither entails the third.”
Its point about defeated_if is right. Failure of paraphrase invariance
and transfer would defeat a stronger hypothesis, that the model’s
representation is invariant to surface form, not the reading of a
model’s response as a parameterized frame. Kirsanov et al.’s stable class
representations, above, already pull the two apart. The defeater is left
as it stands; restating it is the owner’s call. A3’s prompt-equivalence
claim, equivalence over decision consequences, is CLAIM-050 applied to
prompts and is not filed. Its question about activation steering is
QUESTION-106.
Note of 2026-10-29: what holds it back
Its defeat condition is met as written by an Active reading. On Lu et al.’s results (THEORY-683), reordering a fixed set of examples moves accuracy between near chance and near the state of the art, and the good orders do not transfer across model sizes. That is a prompt effect failing invariance to surface form and failing transfer. The note of 2026-10-15 holds that this defeats a stronger hypothesis than the claim. The one finding that separates the two, Kirsanov et al.’s stable class representations (THEORY-251), is Proposed and did not vary order. Until the defeat condition is restated at the level the claim means and that separation is held, the claim is not promoted.