We coded 184,000 rejected redlines where the reviewer gave a reason. A third were correct and not worth the fight. Fewer than one in five reflected a misreading of the playbook. The rest are more interesting.
We built a 1,200-query benchmark to measure how much Recall loses when the question and the precedent are in different languages. German loses little. Mandarin loses some. Arabic loses more, and we think we know why.
Across 112 accounts onboarded since Review 4.0, median accept rate on proposed redlines rose from 71% in week one to 94% in week four. Here is how we count, what drives the curve and what it does not tell you.
A simulated tribunal that rules without explaining is a coin toss with a gavel. We describe how Simulation records each ruling, how we checked 600 of them against practitioners, and where it still disagrees.
What should a voice agent do when the caller starts crying? We spent six months with three clinicians and 140 mock intakes finding out. The answers were mostly about pacing and consent.
Across 184 anonymised acquisition agreements with European biotech targets, what milestone types, efforts standards and dispute mechanisms look like, and what Review does with that.
A dismissed flag costs a reviewer forty-one seconds. That is the small number. The larger one is what happens to the next three flags.
Zero retention means we cannot read the contracts our model reviews. So how do we know it is improving? A method built on synthetic drafts, in-tenant scoring and a panel of lawyers who disagree.
Our in-house model: what it was trained on, what it was not, how it was evaluated against the models we previously routed to, and where it still falls short.
Simulation models venues, not named arbitrators. This note sets out what a corpus of published awards actually supports, where selection bias defeats it, and why we stopped short.
