/deliverable-doctor — claim-level review of a draft

A diagnostician, not a judge: every finding says what is wrong (Why), where (an exact quote), and the change (How), and the tool gives no overall score. Design borrowed from PaperDoctor (Lin et al. 2026), adapted to MEL/SRHR grey literature. The design passed a pre-registered trial on 2026-09-28. Spec and trial result: agent-improvements/deliverable-doctor-spec.md and § Result of agent-improvements/paperdoctor-trial-preregistration-2026-09-28.md (work folder).

Split of work. Agents make judgement calls. Everything exact runs in ane_package.deliverable_doctor and is tested there (tests/test_deliverable_doctor.py): routing, merging, the noise-control gates, the comment cap, anchors, recompute, rendering. Drive it with python scripts/doctor.py <subcommand> (below, DD). If a step needs new logic, add it to the package with a test, not to this file.

Lane

• This skill: a claim-by-claim diagnosis of one draft, for the author (comments) and for Ane (HTML audit trail).
• check-deliverable: the qa-reviewer gate on a finished Ann deliverable. Pass/fail on the house standard, not a claim diagnosis.
• ane-voice: rewrites prose. This skill only audits voice.
• donor-proposal-scoring: scores a proposal against award criteria. Proposal mode hands award-criteria claims to it.
• office-review-pass: edits inside a .docx. This skill only adds comments to a copy.

Step 0 — intake (one message, then wait)