How to use the Integrity Forensics Audit agent skill

What the integrity-forensics skill makes the agent do

An agent skill is a folder of instructions that an AI agent loads when your request matches it. This skill tells the agent to run a fixed, version-pinned external audit tool, Anti-Autoresearch, against a paper. The audit uses automated statistical checks plus GPT-based reviewer agents to list possible issues.

The skill is a launcher. It clones and runs the upstream tool rather than doing the analysis itself. It pins one upstream commit, validates that pin with the upstream eval gate, and then converts the upstream verdict into a policy vocabulary of BLOCK, WARN or NO_NEW_BLOCKER, plus an append-only obligations ledger.

• Evidence ledger: claims are anchored to spans in the paper.
• Auditor proposals: GPT auditors propose findings across nine dimensions.
• Rules-only reporter: lists every proposal together with what the auditor said about it.
• Typed gate: the upstream verdict is mapped to BLOCK, WARN or NO_NEW_BLOCKER.

Install the integrity-forensics skill

Download the skill ZIP from this page and unzip it. Put the resulting integrity-forensics folder in one of two places: ~/.claude/skills/ to make it available across your projects, or .claude/skills/ inside a single project to limit it to that project.

If you use the Claude apps instead, upload the ZIP under Settings → Capabilities → Skills.

• Personal install: ~/.claude/skills/
• Project install: .claude/skills/ in your project folder
• Claude apps: Settings → Capabilities → Skills, then upload the ZIP

Trigger the skill with a natural request

You do not need a special command. The skill's description lists phrases that should activate it. You can ask in English or Chinese. Name the paper by folder, PDF or arXiv ID, since the skill's argument hint accepts a paper directory, a PDF or an arXiv ID.

The phrase 'anti-autoresearch' triggers the skill only when the upstream repo's own skills are not installed. The skill is also invoked by /paper-writing (submission self-forensics, default ON), /peer-review (forensic appendix) and /resubmit-pipeline.

• 'Integrity forensics on ./my-paper'
• 'Forensic audit this paper: paper.pdf'
• 'Run an integrity check on arXiv 2401.00001 before I submit' (example ID for format only)
• '投稿前自查诚信' or '审这篇论文的诚信'

Requirements and what happens on first run

You need git, Python 3, network access to GitHub to clone the pinned repository, and Codex / GPT access through the mcp__codex__codex tool. The audit cannot run without them.

On first run the skill clones the repository into ~/.aris/anti-autoresearch and checks out the pinned commit. It then runs the upstream eval gate, which the source describes as 8 injected-defect classes with 100% recall and zero clean false positives. The gate must pass before the pin can produce a verdict. Later runs reuse the validated pin, so they can work offline.

• git
• Python 3
• Network access to GitHub (first clone and any fetch of a new pin)
• Codex / GPT access via mcp__codex__codex

Read the verdict correctly

The upstream tool produces a verdict of CLEAN_GIVEN_EVIDENCE, SOFT_FLAGS, HARD_FLAGS or REVIEW_UNAVAILABLE. The skill maps these to policy. HARD_FLAGS and REVIEW_UNAVAILABLE become BLOCK. SOFT_FLAGS becomes WARN. CLEAN_GIVEN_EVIDENCE becomes NO_NEW_BLOCKER.

NO_NEW_BLOCKER is not a pass. It means no flag was found in the evidence at hand, not an acquittal. A BLOCK is not a finding of misconduct either. The tool summarizes evidence and does not adjudicate, so every flag is a prompt for human review.

Read the list of dimensions that never ran, too. A WARN can sit on top of a sweep where verdict-bearing dimensions did not run.

• HARD_FLAGS or REVIEW_UNAVAILABLE → BLOCK
• SOFT_FLAGS → WARN
• CLEAN_GIVEN_EVIDENCE → NO_NEW_BLOCKER
• Any open critical obligation → BLOCK; any open obligation → at least WARN

Safety notes: scripts and network access

This skill runs scripts and uses the network. It runs Bash commands, clones a GitHub repository, runs the upstream eval/run_eval.py with Python 3, and calls the Codex / GPT tool for auditor calls. The launcher forces a clean tree at the pin before running, so local edits to the clone are discarded rather than run.

The launcher exposes no reviewer model or effort settings. The pinned upstream decides which model and effort level it uses. Do not expect to change that from the skill.

• Clones https://github.com/wanshuiyin/Anti-Autoresearch.git at a fixed commit
• Runs eval/run_eval.py and forensics_gate.py
• Sends paper content to Codex / GPT auditor calls, as the audit requires
• Refuses to produce a verdict if the eval gate fails

Tips for getting useful results

Use the skill as a checklist, not a target. The skill itself forbids iterating edits until the tool stops flagging, because that would game the verdict. Review each proposal, decide whether it is valid, and record the decision.

Give the agent an absolute path to the paper. The upstream tool finds its files from the repository root, so a relative path can resolve to the wrong place.

If you want the verdict as a standard gate, use the typed policy output and the obligations ledger together rather than the verdict alone.

• Do not loop the audit with edits until it stops flagging.
• Use absolute paths for the paper.
• Keep the report.json and REPORT.md outputs for your records.
• Treat every BLOCK as a question for a human reader.