<p>Publishing has addressed the challenges posed by large language models (LLMs) through a provenance strategy: detectors, declarations, watermarks, and attestations. Provenance matters, but as a test of authorship or quality, it targets the wrong question, is unreliable for that purpose, and risks penalising non-native English writers through a ‘style tax’ on academic prose. Even a dependable instrument would indicate only that a sequence came from a specific distribution, not that the claims are reliable or that the author is accountable. The function publishing has always tracked is <i>answerability</i> – the standing to answer for the claims made – a standing no AI system is eligible to hold, and one current policy already invokes when it refuses AI authorship. Thus, in this article, I argue in favour of a role-indexed architecture specifying what answerability requires of authors, reviewers, editors, and publishers, and where editorial LLMs (eLLMs) belong within it. eLLMs sit in the workflow as recommenders, with calibrated reliance, subject to auditing, appeals, and overrides, and their deployment justified by field-specific evidence, making editors answerable as well. The shift is one of rigour, not of principle.</p>