CappinCheck
CappinCheck is a Gemini-powered claim-auditing tool for dense AI model reports and launch posts.
Core trick
Extracts risky claims from a document, dispatches verifier agents to check each against references and live web evidence, and writes a ledger of verdicts.
Description
CappinCheck is a Gemini-powered claim-auditing tool for dense AI model reports and launch posts. It extracts risky factual claims, checks them against provided references and live web evidence, and produces grounded verdicts plus defensible rewrites.
README excerpt
# HonestLaunch Grounded adversarial claim audit for dense expert documents. HonestLaunch reads an AI model report, paper, technical blog post, or other dense expert document, extracts the riskiest factual claims, then dispatches specialist verifier agents to produce an audited launch page plus a structured evidence ledger. It is not a paper summarizer. The primary output is an audited page view that preserves the source structure, substitutes overstated copy inline, and shows the original wording plus citations on hover or tap. The secondary surface is the claim ledger: original wording, formal verdict, evidence contrast against references, collapsible agent steps, supporting evidence found, contradictions or narrowing evidence, missing context, computed checks when relevant, and the strongest defensible rewrite. ## Current Archive Shape The current public-facing direction is a reviewed static archive of audited launch pages, not an open live-generation product. Browse the archive locally: ```bash cd /Users/rupert/engineering/honestlaunch python3 -m http.server 8765 ``` Then open: - `http://127.0.0.1:8765/` - `http://127.0.0.1:8765/examples/index.html` - `http://127.0.0.1:8765/examples/methodology.html` The first canonical archive set is: - Google - Anthropic - OpenAI - xAI Archive maintenance commands: ```bash source .venv/bin/activate python scripts/rebuild_archive.py python scripts/evaluate_archive.py ``` This rebuilds the featured HTML reports, refreshes arc