Why it is a question
A3: “Can an activation intervention reproduce the task-relative behavioral constraints induced by a textual instruction across context shifts and conflicting instructions? For example, can either intervention reliably preserve the distinction between describing a rejected claim and endorsing it?” Evaluation is over task-relevant decisions, as CLAIM-050 requires, “not merely output embeddings or token distributions”.
CLAIM-029 treats a model’s response to a context as a parameterized pragmatic frame. A3 separated three observations its defeater runs together: a prompt changes the output, a prompt changes an internal representation, two interventions induce equivalent pragmatic behaviour over a family of contexts (see the note of 2026-10-15 on CLAIM-029). This question asks the third directly.
A3’s counter goes with it: “Behavioral equivalence cannot establish mechanistic equivalence.” The steering papers A1 and A3 cite are machine learning and belong to the anthology. None is registered here.