The claim
A43: the model “lets us separate changes in communicative context from changes in the underlying interpretive system”. A44 §4 gave two axes, context c and interpreter parameters θ′_c = A(θ, c). A45 §7 named “three sources of apparent interpretive change: changes in evidence, changes in transient computation, and changes in the interpreter’s parameters”, with the interpreter state s_t = (c_t, h_t, θ_t).
Manuscript §5: “In-context learning changes the effective decoding context without necessarily changing model weights. Test-time training or learned fast-state adaptation modifies the decoder itself. We thus distinguish improving the signal from improving the interpreter.”
What it does not say
A44 §3B: “One should not identify this with literal weight updating in an ordinary transformer”, and “TTT terminology needs care”. The consequence for fidelity (CLAIM-074) is not in the manuscript.
What the in-context-learning readings say
Read by what each change carries:
- Per-sequence learning is the signal. In unnormalized causal linear attention, conditioning on a context is one gradient step of an inner learner (Sun et al., LIT-801; von Oswald et al., LIT-792; THEORY-169), and the learned state is discarded with the sequence. The identity is between two descriptions of what the current context supplies. Nothing in it is carried into the next sequence, so it does not merge context and interpreter.
- Two uses of one context differ. Akyürek et al. (LIT-796, THEORY-439) update the weights on the demonstrations, discard the update after the task, and beat conditioning alone (BIG-Bench Hard 50.5% to 57.8%). Both are uses of the current context, so this shows that how a signal is used matters, not that the interpreter changed.
An exactly solved case of the separation
Mainali and Teixeira (LIT-859, read in NOTE-658) solve the training of one linear-attention layer on in-context regression exactly, when tasks and inputs share an eigenbasis. Training ends at a known fixed point: a predictor that applies a preconditioner, the inverse of the training input covariance (exactly so in the long-context limit; the paper’s finite-context correction holds as stated only for isotropic inputs, THEORY-481), to each context’s own empirical covariance (their Eq. 11). In the long-context limit each eigenmode of the training covariance is learned on its own slow, logistic timescale.
That bears on the “against” item above. THEORY-169 makes conditioning one gradient step of an inner learner. This case shows what builds that inner learner: the outer weights, which change only over training, store a preconditioner fitted to the training distribution. The context then supplies only its own covariance and task. So even in the one class of system where conditioning is a learning step, there are two separate sources of change, on two timescales: the context, which changes the inner learner’s data, and the weights, which change what the inner learner is. That is the claim’s separation, at the level of the layer.
It does not settle the claim. The case is one linear layer, with an assumption (weights staying in the shared eigenbasis) that is not derived, and the paper’s evidence for non-linear transformers rests on timing coincidences (NOTE-658). And the interpreter it shows is the one training left, not one adapted after training and kept.
Note of 2026-10-29: restated by what each history carries
The title told context from interpreter by mechanism, and THEORY-169 holds that in unnormalized causal linear attention conditioning and per-sequence weight updates “differ in what state is kept, not in kind”. The claim is now stated by what each change carries, as CLAIM-074 is (A44’s “contextual” and “adaptational” history): a weight update computed from the current context and discarded after it is part of what the context supplies. THEORY-169’s identity then falls inside the signal, and its own scope says the same (“Not about persistent change”). THEORY-439 leaves corroborated_by: its test-time update is discarded after the task, so it shows that two uses of one context differ, not that the interpreter changed. THEORY-481 stays: the outer weights carry a preconditioner fitted to the training distribution across contexts, and each context supplies only its own covariance.
Note of 2026-10-29: what holds it back
Its promotion waits on the owner’s ruling that test-time training whose update is discarded after the task is part of the signal. Manuscript §5 says test-time training “modifies the decoder itself”, and the restatement departs from that sentence, though A44 §3B says “TTT terminology needs care”. No premise or objection is on file, and nothing in the record meets its defeat condition.
Note of 2026-10-29: promoted
The blocker above was the owner’s ruling on discarded test-time training, which the owner left to the reviewing session’s judgement: it is part of the signal. Where the manuscript (§5) says test-time training “modifies the decoder itself”, that is a fact about where the update sits, and this claim tells the two sources apart by what is carried into later contexts, not by where a change sits. THEORY-169’s identity of conditioning and per-sequence weight updates is then a fact about the signal side and does not tell against the claim.