Section 020 · Chapter 4, Statistical Tests for AI Quality

Comparing Versions with t-tests

A t-test can help Confidence Engineers decide whether a difference in average scores is likely to be real or just sampling noise.

t-testp-valuestatistical significancepaired datacompare prompt or model versions

What to do

  1. Use judgment and risk analysis to decide whether the improvement is safe and worth shipping.
  2. Start with the data-generating process, not a favorite test.
  3. Ask what the experimental unit is, whether the same units saw both versions, what kind of outcome was measured, and which observations can influence one another.
  4. Compare versions within prompt, summarize or model the repeated runs, and calculate uncertainty at the prompt or cluster level.

Evidence to preserve

  • Preserve the inputs, versions, configurations, raw outcomes, and results for t-test, p-value, statistical significance, paired data needed to reproduce work on Comparing Versions with t-tests.
  • Report results for t-test, p-value, statistical significance, paired data by relevant slice, separate blocker failures from averages, state uncertainty and blind spots, and connect the result to a release decision.

Expert note

When the system matters, use paired tests when the same cases run through both versions. Pairing reduces noise from case difficulty. Also inspect assumptions: outliers, non-normal differences, multiple comparisons, and category-specific regressions can all make a tidy p-value misleading.

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. "Comparing Versions with t-tests." Testing AI Knowledge Edition, section 20.

https://jarbon.ai/testing-ai/knowledge/ch020-compare-versions-t-tests.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