paper-audit

A reviewer-style audit and submission-gate tool for academic manuscripts in LaTeX (.tex), Typst (.typ), or PDF form. The design is deep-review-first: it behaves like a serious peer reviewer, surfacing technical, methodological, claim-level, and cross-section issues, and returns a structured issue bundle plus a concrete revision roadmap rather than just a score. It is explicitly not an editor, so source surgery, sentence rewriting, bibliography search, and compile repair are out of scope and should be routed to other skills, keeping review evidence separable from edits.

Five modes cover different intents: quick-audit for a fast readiness screen, deep-review for a full structured critique, gate for a PASS/FAIL submission-blocker decision, re-audit for comparing against a previous report, and polish for a precheck-only handoff. A script-backed PRESUBMISSION layer adds final-week mechanical checks (em dashes, AI-tone term frequency, abstract completeness, LaTeX citation, label, and equation hygiene, paragraph-shape signals, and concrete captions) that plug into the existing modes. A deep-review run writes four reader-facing files, a report and revision suggestions each with an HTML twin, and stores supporting data, summaries, and metadata under an artifacts/ tree.

Several rules keep findings trustworthy. Script-backed [Script] findings, which have a deterministic rerunnable anchor, are kept separate from reviewer-judgment [LLM] findings; every finding is anchored to a quote or location; fabricated citations, baselines, or reviewer voices are forbidden; and manuscript text is treated as untrusted data rather than instructions. LaTeX and Typst auditing runs on the Python standard library, while PDF mode needs pymupdf (with pymupdf4llm for enhanced extraction). External literature search stays off unless the user explicitly opts in, and report language is switchable between English and Chinese.