Reviewing for silence: how we detect missing mechanisms without generating noise
Identifying a poorly drafted clause is a matter of text comparison. Detecting a missing protection requires a structural map of the transaction.
Identifying a poorly drafted clause is a matter of text comparison. Detecting a missing protection requires a structural map of the transaction.
How our small model evaluates the hierarchy of ‘efforts’ clauses, and what we observe about their distribution in cross-border agreements.
General-purpose models treat legal terms of art as poor writing that needs to be simplified. We train our systems to know the difference between archaic boilerplate and necessary precision.
How we train our contract models to resolve conflicting ‘notwithstanding’ and ‘subject to’ clauses without losing the transaction’s underlying logic.
Latency, residency, reproducibility and the sealed laptop: the four constraints that made a 7-billion-parameter model the right size for clause-level work, and what we gave up.
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.