Notes from the lab
Working notes, dated. How we build, why a person stays in the loop, and what we measure.
Section 1 — notes
No figures appear in a note until a pilot report supplies them, with its basis.
The measurements themselves are separate: four public benchmark suites, with the data and the scorer published.
All notes
26 Aug 2026
How we measure extraction accuracy
An accuracy figure is only as good as its definition. This note says which definitions we use, and why a single headline number is never enough.
5 Aug 2026
A pilot ends with a written report the partner can hand to their client. This note lists its sections and what each one is for.
14 Jul 2026
Why a person stays in the loop
Every system we ship has a human review queue. Not as a fallback for a model that is not good enough yet, but as part of the design.
Section 2 — all notes
Newest first.