Latest themes
Notes for teams engineering behavior, evidence, and expert judgment.
Our writing connects practical development decisions to the deeper problem of keeping human intent intact as AI systems become more open-ended.
Generic judges are only a starting point
Why enterprise AI needs correctness specifications grounded in internal expertise rather than generic model preferences.
Research directionCoverage is a design problem
A deliberate approach to finding the behavioral dimensions, difficult combinations and edge cases that matter.
Explore coverageExperts should decide, not annotate
Designing agent-assisted workflows that reserve scarce specialist time for judgment that materially changes the benchmark.
Explore elicitation