"A reviewer may point; only the authoring role may write the replacement."
Separate the powers
Diagnosis and authorship are different powers. A reviewer can notice stiff cadence, borrowed syntax, or an odd idiom. It may explain the problem and point to a locus, but it does not silently replace the translation.
Let state name the role
Persist the reviewer's diagnosis in translations.review_notes while leaving translations.revised_text untouched. A later authoring route may read those notes and choose revised wording.
The convenient overreach
A convenient reviewer rewrite collapses accountability. When wording changes, nobody can tell whether the translator chose it or the critic installed it. The same model may perform both roles at different times, but the job boundary must remain.
Follow one sentence through four acts
Test “The Reviewer Holds No Pen” by tracing original authorship, translation authorship, blind review, and author revision. For each act list inputs, outputs, visible evidence, writable fields, and the proof handed forward. One brain may perform two roles at different times; the job and authority must still remain distinct.
Make the judgment
Define a review-note schema that can be actionable without containing replacement prose. If you cannot answer who chose the final wording in one sentence, provenance is too thin. A note that silently edits text or a stet whose reason disappears changes who holds the pen.
Narrow roles sharpen quality
Giving a reviewer every truth is not always an improvement. Restricting evidence and action makes the naturalness question answerable. The author uses that answer as material while retaining responsibility for the final choice.
Prove authority in the write set
The current translation row gives the reviewer one writable field, review_notes, and reserves revised_text for the authoring role. Role prompts explain the boundary; the store-level allowlist enforces it.
The same model may return later as author, but that later job carries different authority and provenance.