anti-autoresearch

Plugin anti-autoresearch from marketplace wenhaochai in wenhaochai/claude-plugins. Commands read from the repo's manifest — 6 days ago.

Skills (13)

• adversarial-case-builder

  Builds the strongest rejection argument only from the evidence ledger and other auditors' findings, then tests each point; the paper may survive. Memo-only, no verdict weight (the adjudicator decides); run last. Triggers: \"strongest objection\", \"rejection memo\", \"kill argument\", \"最强拒稿点\".

• ai-style-impressions

  Lists located AI writing-style impressions in a paper (LLM phrase tics, defensive hedging, jargon-stuffing, invented codenames, bullet overuse) with zero verdict weight; not an authorship classifier. Run after evidence-ledger. Triggers: \"AI style\", \"vibe check\", \"vibe paper\", \"writing fingerp

• anti-autoresearch

  Full integrity forensic sweep of a paper (dir, PDF or arXiv id), especially autoresearch/AI-Scientist output: evidence ledger, all auditors, report with a deterministic verdict. Detect-only; audits integrity, not authorship. Triggers: \"integrity audit this paper\", \"forensic review\", \"audit a su

• baseline-comparison-audit

  Audits a paper's baseline comparisons: missing SOTA/floor baselines, undertuned or unequal-budget ones, \"outperforms\" within error bars or without seeds, cross-row gains miscomputed. Needs evidence-ledger. Triggers: \"missing baselines\", \"weak baseline\", \"is the comparison fair\", \"SOTA earne

• citation-forensics

  Checks each cited reference against DBLP/arXiv/DOI: exists, metadata right, not retracted, supports the citing sentence. Run after evidence-ledger on LaTeX source; a PDF-only ledger has no citations. Triggers: \"check the references\", \"verify references\", \"hallucinated citations\", \"wrong-conte

• consistency-audit

  Checks a paper against itself: numbers clashing across sections, miscomputed gains, impossible p-values/SDs, best seed shown as the mean, described vs evaluated method, overclaims, causal leaps, unisolated ablations. Works from the PDF alone; run evidence-ledger first. Triggers: \"self-consistency\"

• eval-design-forensics

  Audits whether a paper's evaluation measures what it claims: train/test leakage, conflicted or unvalidated LLM judges, selective reporting (dropped conditions, switched metrics). Works PDF-only, after evidence-ledger. Triggers: \"data leakage\", \"evaluation validity\", \"LLM judge bias\", \"cherry-

• evidence-ledger

  Builds the evidence ledger every other Anti-Autoresearch auditor reads: from a paper dir, PDF or arXiv id, derives the observability level and extracts span-anchored, hashed claims (claims.json). No findings or verdict; run FIRST. Triggers: \"build the ledger\", \"extract claims\", \"prep for integr

• experiment-forensics

  Checks paper numbers and method against released code and results: fake ground truth, self-normalized scores, phantom or too-clean results, placeholder data, missing code/prompts. Needs evidence-ledger first; without repo + results, info notes only. Triggers: \"audit the results\", \"check the eval

• novelty-duplication-advisory

  Retrieves uncited prior work and lays it beside a paper's contribution claims so a reviewer can weigh A+B+C stapling or duplicate publication. Memo-only, never rules novelty, not a plagiarism checker. Run after evidence-ledger. Triggers: \"prior-work overlap\", \"duplicate submission\", \"is this st

• presentation-signals

  Flags surface tells in a paper: duplicate tables, leftover chatbot/pipeline/template strings, too few or LLM-looking figures, page padding. Auxiliary, capped at minor; AI prose style goes to ai-style-impressions. Run after evidence-ledger. Triggers: \"presentation signals\", \"surface check\", \"dup

• proof-derivation-forensics

  Detect-only check that a third party's written proofs and derivations establish their theorems: gaps, circularity, invalid steps, symbol drift, smuggled assumptions. Needs the evidence ledger; verdicts need LaTeX source (PDF-only: info). Triggers: \"check this proof\", \"audit the math\", \"证明审计\",

• reviewer-discipline

  Filters an already-drafted weakness list through venue norms (CVPR reviewer errors, banned vague phrasings, ARR rewrites, COLM/TMLR guards); finds no weaknesses itself. For reviews, self-reviews, meta-reviews, rebuttals. Triggers: \"reviewer comments\", \"weakness list\", \"审稿\", \"是不是审稿意见写太虚\".