small print

Paste a Terms of Service. See who each clause favours, and the exact words that say so.

The document

Paste the terms

Or open an exhibit

Loading exhibits…

Three fictional companies, with terms written for this demo and readings written by hand in the exact format the model returns. They work with no AI key at all.

Read before in this browser

Kept only in this browser’s storage.

Not legal advice

small print is a reading aid. Labels and one-line readings come from a fixed set of hand-written rules (or, on a server with an AI key, a model; for the exhibits, a person) and can be wrong or miss context. The highlighted words are the evidence: read them yourself, and ask a qualified lawyer before a decision that matters.

j / k next and previous clause · Esc clear · click a clause to select it

Who it favours

How it reads

Pasted text is read in your browser, by Python. Pyodide (CPython compiled to WebAssembly) runs this project’s Python in a background worker: it splits the text into clauses, then tries a fixed library of hand-written rules on each one, for forced arbitration, class-action and jury waivers, changes to the terms, auto-renewal, account closure, liability caps, indemnity, giving up claims, selling and sharing data, AI training on your content, tracking, licences over your content, law and courts, refunds and fees, age limits, and the clauses that protect you. Each rule says who its kind of clause favours.

The words a rule matched become the quote, and every quote is checked against your text, character for character, inside its own clause, by the same verifier a model’s quotes go through. A rule’s “match” says how specific the wording was, not how sure anyone is about the law. No AI reads your text here.

Your text is split into clauses by a small Python program running in a Cloudflare Worker. The clauses go to Claude Opus 5.5, which returns, for each one, who it favours, a one-line reading, a confidence and a quote. The Worker then checks every quote against your text, character for character, and drops any reading whose quote is not inside its own clause. The page tells you how many were dropped.

If Opus 5.5 declines, another Claude model may answer; each part shows which model read it.

What is kept

Nothing leaves your device: the text is read in this browser tab, and the Python runtime comes from this site, not a third party. Your last five documents are kept in this browser’s storage only.

The same page can also send text to a Python Worker that asks Claude, when its owner deploys one with an API key. This site has no key, so it does not.

The Worker caches each analysis in which at least 80% of the clauses got a checked reading, under a SHA-256 fingerprint of the text, so the same document is not paid for twice. The cache holds offsets, labels and readings, not your text, and can only be looked up by someone who already has the exact text. Nothing you paste is logged.

Made by

Ramen Protocol (ramenprotokol), built with AI assistance from Claude. The exhibits are fictional; no real company’s terms are quoted here. Not legal advice.

Runs on Pyodide (MPL 2.0) and CPython (PSF License), served from this site. Third-party notices.