Section 033 · Chapter 5, Judges, Humans, and Disagreement

Using Raters Well

Human raters are not a checkbox. They are an evaluation instrument that needs selection, calibration, workflow design, and quality control.

exact assertionsLLM judgehuman raterraters

What to do

  1. Use overlap deliberately.
  2. Keep raters blind when possible.
  3. Track disagreement as data.
  4. Track inter-rater agreement, rater-specific bias, calibration drift, fatigue effects, adjudication outcomes, and whether the rater population matches the user population.

Evidence to preserve

  • Track disagreement as data.
  • Track inter-rater agreement, rater-specific bias, calibration drift, fatigue effects, adjudication outcomes, and whether the rater population matches the user population.
  • Preserve the inputs, versions, configurations, raw outcomes, and results for exact assertions, LLM judge, human rater, raters needed to reproduce work on Using Raters Well.
  • Report results for exact assertions, LLM judge, human rater, raters by relevant slice, separate blocker failures from averages, state uncertainty and blind spots, and connect the result to a release decision.

Expert note

At scale, treat raters as measurement instruments. Track inter-rater agreement, rater-specific bias, calibration drift, fatigue effects, adjudication outcomes, and whether the rater population matches the user population. If raters and users disagree systematically, the eval is measuring the wrong audience.

Continue the conversation

Apply this to your context.

Save your product context once, then open a focused conversation that combines it with this concept.

Cite this page

Jason Arbon. "Using Raters Well." Testing AI Knowledge Edition, section 33.

https://jarbon.ai/testing-ai/knowledge/ch033-raters.html

Shared across the Knowledge Edition

Adapt every concept to your world.

Describe your product, role, users, risks, constraints, or current quality problem. This stays in this browser until you choose to send it to ChatGPT.

Saved only in this browser.0 / 2400