All articles

Research

·

9 min read

What we learned measuring hallucinations at scale

We graded two million answers by hand and by machine. The results changed how we train, evaluate and talk about accuracy.

Mira Castellanos

Research Lead

Accuracy is easy to claim and hard to measure. Over six months we graded more than two million answers across contracts, research papers and financial filings, combining expert reviewers with automated checks.

Errors hide in confident sentences

The most dangerous mistakes were not wild fabrications. They were small, confident distortions — a number from the wrong year, a clause from a superseded draft. Those are exactly the errors verification is designed to catch.

A model that says “I can’t find support for this” is more useful than one that is right most of the time and never tells you when it isn’t.

We now publish our evaluation method alongside every release, including the areas where Orchid still falls short.

Create a free website with Framer, the website builder loved by startups, designers and agencies.