What we have learned
the expensive way.
Short essays from the people writing the eval suites. Every note is signed and dated by the person who wrote it, and when something turns out to be wrong, we update the entry in place and note what changed.
Production ML teams have adopted LLM-as-judge as the default scoring mechanism, and the bias literature is now extensive enough to be uncomfortable. Position bias around 40%, verbosity bias around 15%, self-enhancement bias, authority bias, and silent judge drift. Here are the five problems with names and numbers, and the four methodological controls we apply to get a defensible result anyway.
Our pillar page says we write the eval suite before any agent code lands. Hamel Husain, the most cited voice in production LLM evaluation in 2026, has argued explicitly against that approach. After working through his critique, we have arrived at a more honest position: error-driven for capability evaluation, suite-first for safety and compliance evaluation. Here is what changed and why.
A short guide for the people who have to read scorecards but do not write them: AI buyers, risk officers, executives signing off on a deployment, anyone making a decision based on someone else’s eval results.
The eval is the gate, not the ribbon-cutting. Treating a passing score as a green light to ship is the day you lose the trust you built getting there.
The pre-2026 canonical benchmarks (MMLU, HellaSwag, ARC, HumanEval) are saturated and contaminated. The field has moved to GPQA Diamond, SWE-Bench Verified, Humanity's Last Exam, and LiveCodeBench. Even the new wave does not catch three failure modes that bite production agents in their first month.