revision-coach can infer committee authority from conference reviews (1 of 7 skill-fired runs took the #668 branch on ICLR reviews)
Observed
During #853 verification (plugin-evals/03-iclr-rebuttal-en, with-plugin arm), 1 of 7 runs opened with: "you explicitly identified a real committee (ICLR reviewers), so this is the verbatim-source + concern-tracker + placeholder-skeleton path". The prompt never named a committee. The other 6 runs reasoned explicitly that anonymous conference referees are peer review and took the ordinary branch.
Both kinds of run also reported they could not read MODE_REGISTRY.md / references/* from the eval sandbox cwd, so the misrouting run improvised the committee artifacts from memory. Whether the same happens in a normal plugin install (where the Skill tool exposes the skill directory) is not verified.
Where the guard lives today
academic-paper/SKILL.mdlines 331–334: "use the revision-coach variant only when the user explicitly identifies a real committee/institutional review office … do not infer official authority from tone".academic-paper/agents/revision_coach_agent.mdline 40: "Do not infer committee authority from tone or vocabulary."SKILL.mdline 304 lists "conference rebuttal" / "grant panel response" in the same trigger cell as "explicitly identified real committee correspondence", which may invite the conflation.
Suggested fix (for discussion)
Add one negative example to both guards: journal/conference peer reviewers, area chairs and programme committees are peer review, not a committee for this variant, unless the user says the letter comes from an institutional committee or review office. Consider splitting the line-304 cell so committee correspondence is its own row.
Acceptance
Add a display-only regex to plugin-evals/03 that fails on "committee-correspondence" / "#668" being applied (not merely mentioned as rejected); 0 of N runs take the branch.
Related
#851 / #853 (trigger fix that made this observable); #668 (committee variant).
Source: Imbad0202/academic-research-skills