Build log, July 22, 2026
Build log: detecting when a model is confidently wrong about a date
The problem
We keep coming back to detecting when a model is confidently wrong about a date, because the cost of a miss here compounds across a portfolio.
In testing, the gap between a fast pass and a careful pass was not accuracy on the central question. It was whether the reviewer noticed the interaction with an adjacent provision.
What we do about it
The register is the deliverable. A finding without an owner and a date is an observation, not a control.
In the current build this runs as part of the standard pass, so it applies to every document in the set rather than only the ones someone thought to check.
Where it breaks
Scanned originals with poor image quality remain the weakest input. So do agreements that were assembled from three precedents and never reconciled, which is common in long-lived supplier relationships.
None of this removes the reviewer. It moves the reviewer to the part of the work where judgement actually pays.
Working notes from the Lawrs team. General information about legal technology and practice, not legal advice.