AutoVerifier

AutoVerifier is an LLM-based agentic Claude Agent Skill that rigorously verifies the claims in scientific and technical artifacts — papers, preprints, technical reports, and "breakthrough" announcements. Rather than asking a model for a one-shot opinion, it drives a structured six-layer verification pipeline: it builds a corpus, decomposes the work into provenance-tagged claim triples, checks each claim for internal consistency and overclaiming, triangulates against independent external sources, corroborates with financial / conflict-of-interest / supply-chain signals, and finally assembles a hypothesis matrix with a technology-maturity rating and alpha-signal detection. The result is a transparent, auditable assessment of whether a result is reproducible, genuinely novel, independently corroborated, or overclaimed.

Note: This skill packages and implements the method introduced in the paper AutoVerifier: An Agentic Automated Verification Framework Using Large Language Models (arXiv:2604.02617). If you use this skill in your work, please cite the paper — not this repository (see Citation).

What it does

AutoVerifier runs a six-layer verification protocol:

1. Corpus construction — assemble the target artifact together with the supporting and competing literature needed to evaluate it.
2. Entity & claim extraction — decompose the work into (subject, predicate, object) claim triples, each tagged with a 5-level provenance rating.
3. Intra-document verification & overclaim detection — check each claim for internal consistency and flag claims that outrun their own evidence.
4. Cross-source verification — triangulate claims against external sources, weighting each by source independence.
5. External-signal corroboration — corroborate against financial, conflict-of-interest (COI), and supply-chain signals.
6. Hypothesis matrix — synthesize findings into a hypothesis matrix with a technology-maturity rating and alpha-signal detection.

Install

Claude Code (CLI)

Personal / global scope — available in every project:

git clone git@github.com:yuntaod/AutoVerifier.git ~/.claude/skills/autoverifier

Project scope — available only inside one repository. Clone the skill anywhere, then copy the folder into the project's .claude/skills/ directory:

git clone git@github.com:yuntaod/AutoVerifier.git /tmp/autoverifier
mkdir -p .claude/skills
cp -R /tmp/autoverifier .claude/skills/autoverifier

Either way, the skill is loaded automatically the next time Claude Code starts.

Claude.ai / Desktop app

1. Download autoverifier.skill from the latest GitHub Release.
2. Open Settings → Skills.
3. Upload the autoverifier.skill file.

Citation

This skill is based on the paper below. If you use AutoVerifier, please cite the paper (not this repository):

Yuntao Du, Minh Dinh, Kaiyuan Zhang, and Ninghui Li. "AutoVerifier: An Agentic Automated Verification Framework Using Large Language Models." arXiv:2604.02617 (2026).

@article{du2026autoverifier,
  title   = {AutoVerifier: An Agentic Automated Verification Framework Using Large Language Models},
  author  = {Du, Yuntao and Dinh, Minh and Zhang, Kaiyuan and Li, Ninghui},
  journal = {arXiv preprint arXiv:2604.02617},
  year    = {2026}
}

A machine-readable citation is also provided in CITATION.cff.

License

• Skill (this repository's code and skill definition): Apache-2.0 — Copyright 2026 Yuntao Du.
• Paper ("AutoVerifier: An Agentic Automated Verification Framework Using Large Language Models," arXiv:2604.02617): CC BY 4.0.