Evals for AI Engineers

Chapter 4. Collaborative Evaluation Practices

A Note for Early Release Readers

With Early Release ebooks, you get books in their earliest form—the author’s raw and unedited content as they write—so you can take advantage of these technologies long before the official release of these titles.

This will be the 4th chapter of the final book. If you’d like to be actively involved in reviewing and commenting on this draft, please reach out to the editor at [e-mail].

Error analysis (Chapter 3) gives us a list of failure modes. The next question is whether different people on the team agree on what qualifies as a failure. This chapter covers how to align human judgment.

In the previous chapter (Chapter 3), we walked through error analysis: reading traces, identifying failure modes, and building a taxonomy of how your system goes wrong. That process depends on human judgment at every step. You decide what counts as a failure. You decide how to categorize errors. You decide whether a trace is acceptable or not.

Error analysis (Chapter 3) depends on human judgment at every step. You decide what qualifies as a failure, and how to categorize errors. But evaluation criteria are often subjective (what one person considers “helpful,” another might find verbose), and individual judgment is inconsistent (the same person might rate the same trace differently on different days). If the labels are unreliable, everything built on top of them (failure mode counts, automated evaluators, ...