Classify locally
A bag-of-words neural classifier, trained in your browser with TensorFlow.js, maps the claim to one of eight audit intents. No server, no API key, no data leaving the tab.
Local-first AI · block-pinned evidence
Turn on-chain claims into receipts.
A small open-source AI agent reads a claim in plain English, routes it to the right on-chain check, and produces a MATCH / MISMATCH report with every answer pinned to a block.
- TensorFlow.js in your browser
- No API keys
- MIT licensed
How it works
An agent loop you can inspect.
Check the chain
The agent calls the matching tool: ERC-8004 owner and wallet reads, tokenURI and registration files, registration receipts, payout Transfer logs, or token facts.
Pin the verdict
Every check is tied to a named block and the raw calldata behind it. You get a claim-by-claim MATCH / MISMATCH table and a Markdown report you can re-run.
Live audit
FlapJax Q31, checked in your browser.
This is a real audit of ERC-8004 agent 364913 on BNB Smart Chain. The model routes each claim; the chain supplies the evidence. Nothing is mocked.