How accurate is it, and how we measured

Does it quote your contract correctly?

This is the failure that matters most. A tool that invents a clause, or quietly rewords a real one, is worse than no tool: you would go into a negotiation arguing about a term your contract does not contain, or sign past one it does. So every finding carries the exact words it was drawn from, and those words get searched for in your document before the finding reaches you.

That search is the plainest check there is. It takes the quote, takes your file, and looks for one inside the other, allowing for the ways a PDF or a Word export rewraps text and nothing else. No model is asked whether it did a good job. Each quote in your results is marked when it was found word for word, and marked differently when it was not, so you are never left guessing which kind of statement you are reading.

The figures below are that same check, run over a corpus of real contracts, which is how we can tell you what to expect before you upload anything.

The 2 slips in that run were transcription errors: a quote that did not match the source character for character, close enough to look right and wrong enough to break the verbatim promise. Neither invented an obligation. Inventing a clause is the far more dangerous failure, which is why its count sits in the figure above rather than in a footnote.

How much of a contract does it catch?

Not all of it, and this is the number most tools do not publish. Semantic completeness measures how much of what a careful reader would flag the review actually flagged.

That number needs the standard it was measured against, or it says something worse than the truth. The benchmark is every point a careful reader with no deadline would raise, marked up clause by clause in advance, and the score is how much of that list the review reached on its own. Against that, it surfaced a bit over half.

So: a strong first pass that catches most of the obvious problems and some of the subtle ones, and not a substitute for a lawyer reading your contract. That is why the disclaimer at the bottom of every page says what it says, and why the section above matters as much as this one. What it does hand you is exact: the clauses it raises are really in your document, quoted word for word, and checked.

We publish this because a number that only flatters us is marketing, and because improving it is the current priority. When it improves, the date above changes and the old figure goes away.

Why there is only one run here

There used to be two, and the older one was retired on purpose. It was measured against a test corpus we replaced hours later the same day: the documents got roughly twice as long and the answer keys were rewritten against the new text. Our own record says in writing that scores from before that change cannot be compared to scores after it, so quoting the older, friendlier figures beside the newer ones would have implied a comparison that does not exist.

The retired figures were also better looking than these. That is precisely why they had to go: a number kept because it flatters is not a measurement.

The run that survives is published in full, defects included, at the eval record for 2026-08-04.

Why the numbers on this page expire

Each claim carries the date its measurement ran. A build check fails if any of them goes more than 180 days without being re-measured, so a stale number cannot quietly survive here. Every measured figure lives in one file and is imported into this page rather than typed into it, which means the page cannot say something the measurement does not.

This matters because the alternative is common. An accuracy figure that never changes is a figure nobody is re-checking.

What we will not claim

• A percentage accuracy for the review as a whole. There is no single number that honestly describes it, which is why there are several above.
• That a second model checks the first. That was considered and never built, and a build check now fails if the claim reappears anywhere in the site.
• Customer counts, rankings, or endorsements. If we cannot link the source, it does not go on the page, and every source above is a link to the record itself rather than a file path you would have to take our word for.

Check it against your own contract

The reason to publish a measurement is so that you do not have to take it on faith. Ours ran on documents we chose. The run that settles anything for you is the one on the document you actually have to sign.

Every finding comes back with the words it was drawn from, already searched for in your file, and marked with whether it was found. You do not have to take that marking on faith either: the quote is right there, and your document is on your screen. Read the completeness section above first, and treat what comes back the way it asks to be treated: a first pass, not the last word. No account, no email.