Building great evals part 8: The discriminatory property of evals

realmadhuguru · x · 2026-08-25

Part 8 of the series on building great evals discusses the discriminatory property. A good eval should separate AI systems that are meaningfully different. If scores are too close (like giving a 5th-grade math test to PhDs), the eval has low power. The goal is not arbitrary difficulty but a balance: realistic + difficult + sensitive to capability differences. As models improve, good evals saturate, requiring strategy shifts.

Original post →

More from coding & agent

coding & agent channel →