Research

Notes on method

Working notes on how we design tasks, capture expert reasoning and control quality. Method only: we write about our approach, not about results we have not produced.

Published on August 18, 2026 · 4 min read

How we capture professional reasoning

Why answer only datasets fall short for agentic training, and how we design tasks that make expert judgment legible: calibrated difficulty, externalised decisions and a trajectory schema built for learning.

Read the note

Published on September 1, 2026 · 4 min read

Quality control for expert data

Our approach to quality: credential verification, a practical entry exam, layered senior review, measured agreement between reviewers, audit sampling, and a standing preference for rejecting a batch over shipping an uncertain one.

Read the note