Anti-Autoresearch
Don't trust an autoresearch paper at face value. Reviewer-side integrity forensics (self-consistency + fabrication), deterministic verdict. 61 signals: 46 integ
What is Anti-Autoresearch?
Don't trust an autoresearch paper at face value. Reviewer-side integrity forensics (self-consistency + fabrication), deterministic verdict. 61 signals: 46 integrity hack-patterns (families A–H, verdict-bearing) + 13 zero-weight AI writing-style impressions (AIS) + 2 advisory. Not an opaque AI-text classifier. The dual of ARIS.
Overview of Anti-Autoresearch
What it is
Anti-Autoresearch is a reviewer-side integrity forensics tool for autoresearch papers. It audits a submission for self-consistency and fabrication, producing a deterministic verdict based on 61 signals: 46 integrity patterns (families A–H, verdict-bearing), 13 zero-weight AI writing-style impressions (AIS), and 2 advisories. It is not an opaque AI-text classifier and does not assign authorship probabilities or an “AI-written” verdict. The tool is built on the ARIS foundation (~12.5k stars, HuggingFace Daily Papers #1) and points ARIS’s production audit DNA outward to audit a third party’s submission.
Key features
- 46 integrity patterns across 8 families (A–H): numeric self-consistency, method & scope, baseline integrity, experiment integrity, citation integrity, presentation & surface, proof & derivation, evaluation design & validity.
- 13 zero-weight AI writing-style impressions (AIS) reported in a separate section with no impact on the integrity verdict.
- 2 advisories: adversarial case builder (strongest rejection paragraph) and novelty duplication advisory (prior-work overlap).
- Deterministic spine: an evidence ledger → LLM auditors that propose findings → a rules-only reporter that lists every proposal and summarizes.
- 8 patterns eval-gated end-to-end (GRIM, GRIMMER, statcheck, delta arithmetic, hedge-density, etc.) with CI.
- Delivered as 11 skills + 1 orchestrating workflow.
- Not an opaque AI-text classifier; no authorship probabilities or “AI-written” verdict.
Use cases
Reviewers, area chairs, and honest authors can use Anti-Autoresearch to verify that a paper’s science holds together and reflects its own evidence. Common use cases include: checking whether tables match the text, whether baselines are present and fairly configured, whether code reproduces the reported numbers, whether citations are fabricated or misattributed, and whether derivations are valid. The tool surfaces discrepancies a human reviewer should investigate, without judging misconduct.
Getting started
bash git clone https://github.com/wanshuiyin/Anti-Autoresearch.git ./Anti-Autoresearch/tools/install_anti_autoresearch.sh claude mcp add codex -- codex mcp-server claude mcp list claude > /anti-autoresearch ~/papers/submission The run writes REPORT.md, report.json, claims.json, and per-skill *.findings.json into the paper directory.
License and pricing
The project is licensed under MIT and is open source.
Based on the project's public README and repository metadata.