Skip to content
Get packs

Guide · Agent skill

How to use the Integrity Forensics Paper Audit agent skill

Updated · written with AI from the project's own documentation and checked against our catalogue data

Integrity Forensics is an agent skill that audits a research paper for numerical integrity problems, such as inconsistent statistics (GRIM, GRIMMER, statcheck) and arithmetic deltas. It runs the rule-based part of the Anti-Autoresearch project and turns the result into a gate labelled BLOCK, WARN or NO_NEW_BLOCKER, plus a ledger of findings. This guide explains what an agent skill is, how to install this one, how to ask for an audit, and what its limits are.

Integrity Forensics Paper Auditwanshuiyin/Auto-claude-code-research-in-sleep · MIT

Quick start

  1. 1Download and unzip. Download the ZIP from this page and unzip it. You should end up with an integrity-forensics folder containing SKILL.md.
  2. 2Place the skill folder. Move the integrity-forensics folder into ~/.claude/skills/ or into .claude/skills/ inside your project. Alternatively, upload it under Settings, Capabilities, Skills.
  3. 3Confirm git and python3. Make sure git and python3 are installed and available to the agent on your machine.
  4. 4Ask for an audit. Tell the agent to run integrity forensics on your paper folder, PDF or arXiv ID.
  5. 5Let the first run set up the pin. On first use the agent clones the Anti-Autoresearch repository at the pinned commit and runs its eval script before proceeding.
  6. 6Read the gate. Check whether the result is BLOCK, WARN or NO_NEW_BLOCKER and note each flag in the report.
  7. 7Resolve findings in the ledger. Record a typed resolution for each finding, or have a human waive it. Re-run the gate afterwards.

What an agent skill is and what this one does

An agent skill is a folder containing a SKILL.md file with instructions that an AI coding agent reads when a matching request comes up. This skill is a thin launcher: it tells the agent how to fetch a specific version of the external Anti-Autoresearch tool, run its deterministic checks on a paper, and record the outcome.

The audit target can be a paper folder, a PDF or an arXiv ID. The deterministic mode covers the numeric forensic core (GRIM, GRIMMER, statcheck and delta arithmetic) and a rules-only reporter.

  • Checks numerical consistency of reported statistics and arithmetic
  • Produces a typed gate: BLOCK, WARN or NO_NEW_BLOCKER
  • Keeps an append-only obligations ledger of findings and how each was resolved

What this mode cannot do

The upstream project's nine semantic review dimensions are written as Claude Code contracts. This pack runs only the deterministic-only mode, so those semantic reviews do not run. The skill can raise flags, but it can never return a clean result such as CLEAN_GIVEN_EVIDENCE.

Because of this, a run that finds no flags still reports REVIEW_UNAVAILABLE, and the gate blocks it. Treat the output as a first pass for numeric problems, not a certification of integrity.

Install the skill

Download the ZIP from this page and unzip it. Then place the integrity-forensics folder in one of two locations: ~/.claude/skills/ for all your projects, or .claude/skills/ inside a single project folder.

Alternatively, upload the skill in the Claude apps under Settings, then Capabilities, then Skills.

  • Personal: ~/.claude/skills/integrity-forensics/
  • Project only: .claude/skills/integrity-forensics/ inside your project

Check the requirements before the first run

The skill needs git and python3 on the machine where the agent runs. It also needs network access, because it clones the Anti-Autoresearch GitHub repository at a pinned commit and runs that repository's eval/run_eval.py script the first time it sees the pin.

  • git
  • python3
  • Network access to GitHub
  • Upstream Anti-Autoresearch repository, cloned to ~/.aris/anti-autoresearch

Trigger the skill with a natural request

The skill activates on phrases such as 'integrity forensics', 'forensic audit this paper', or the Chinese phrase 投稿前自查诚信. Give the agent the paper location in the same message. Each of these is a reasonable request:

  • Run integrity forensics on ./my-paper
  • Forensic audit this paper: /path/to/manuscript.pdf
  • Do an integrity forensics check on arXiv 2401.00001 before I submit

Read the gate result and fix the findings

HARD_FLAGS produce BLOCK, and REVIEW_UNAVAILABLE also produces BLOCK. SOFT_FLAGS produce WARN. The source lists NO_NEW_BLOCKER as a pass-capable decision alongside WARN.

Each finding goes into an append-only ledger. You resolve a finding by recording a typed receipt (corrected-from-results, claim-narrowed, claim-withdrawn or citation-replaced) with verified-by set to human:<name>, checker:<tool> or cross-family-review:<thread-id>. Only a human can waive a finding. Numeric findings route to the result files, and other findings route to the matching audit skill or the author.

Do not loop by editing, re-running and repeating until the flags disappear. The skill names this as its one forbidden pattern.

  • BLOCK: fix the findings or record resolutions, then re-run
  • WARN: review each flag and record how it was handled
  • Any change to the ledger deletes the standing gate.json

Safety notes before you run it

This skill clones a repository and runs code from it, and it uses the network. The bootstrap resets the clone to the pinned commit and removes untracked files in that clone directory, so local edits to that clone are discarded. The pin is fixed in the skill and is only changed through a reviewed update.

Only audit papers you have permission to process, because the tool may send data over the network.

  • Use only papers you are allowed to process
  • Expect the clone directory to be overwritten at the pinned commit
  • Do not treat a pass as proof that a paper is clean

What people use it for

  • Authors checking the statistics in a manuscript before journal submission
  • Reviewers or students making a quick first pass for numeric inconsistencies
  • Research teams keeping a repeatable, logged integrity checklist for each paper

Frequently asked questions

Is the Integrity Forensics skill free to use?

The skill is published under the MIT licence. The upstream Anti-Autoresearch repository it clones is a separate project, and you should check its own licence and terms before relying on it.

What do I need installed to use it?

You need git, python3, and network access to clone the Anti-Autoresearch GitHub repository at the pinned commit. The skill runs that repository's eval script on first use.

Can this skill certify that my paper has no integrity problems?

No. The skill can flag problems but can never return a clean result. In this deterministic-only mode the semantic review dimensions do not run, so many issues will not be caught.

Does it check GRIM, GRIMMER and statcheck?

Yes. The numeric forensic core covers GRIM, GRIMMER, statcheck and delta arithmetic. It does not run the upstream semantic review dimensions.

Why does a clean-looking paper still get BLOCK?

When no flags are found, the deterministic-only run reports REVIEW_UNAVAILABLE because the semantic checks never ran. The gate treats that as BLOCK so that nothing waves the paper through.

Where do I put the skill folder for Claude Code?

Put it in ~/.claude/skills/ for all projects, or in .claude/skills/ inside one project. In the Claude apps, you can upload it under Settings, Capabilities, Skills.

New agent skills, repos and AI tools — every week

The best new skills and open-source projects, the fastest-growing repos, and one free prompt. Five minutes to read. Unsubscribe anytime.

Our verdict

Use this skill if you want a repeatable first pass for numeric inconsistencies in a paper you are allowed to process, and a log of how each finding was handled. Pick a different tool, or a full review, if you need a clean integrity certification, because this mode never produces one and skips the semantic checks.

Download the skill