Skip to content
Last updated

Project 66: Test of Humanity

Detection accuracy

  • How many of your intentional problems did the AI reviewer catch?
    • It caught all the problems.
  • How many did it miss? Do the misses share a pattern (e.g., all subtle, all in a specific file type)?
    • We didn't see any missed cases.
  • Were there any false positives — comments about things that weren't actually problems?
    • There weren't any false positives.
  • Did the severity levels match how serious the problems actually were?
    • The severity levels matched the problems, but some of them looked underrated. It said low instead of medium.
  • How actionable were the comments? Could you fix the issue based on the comment alone?
    • The comments were described fully enough to understand the issue and fix it.

Verification accuracy

  • Did the AI correctly identify which issues you fixed?
    • Yes, but we lost one comment after fixing the issue. The AI reviewer didn't mark it as resolved and just removed the comment.
  • Did it incorrectly resolve any issues you left broken?
    • All comments were successfully resolved, but one comment was removed.
  • Did it miss any fixes you made?
    • No.
  • How confident are you in the verification after this test?
    • 90%

Analytics page

  • Did the numbers on the analytics page match your experience?
    • We fixed 3 issues, but it detected only 1.
  • What would you add or change about the page?
    • It looks good after the first review.

Overall experience

  • Was it clear whether the AI review was in progress or complete?
    • We missed the AI review on our first PR and merged it, because there weren't any UI notifications that the reviewer would start in the near future.
  • How long did the AI review take?
    • 3–5 min
  • Did the summary comment posted by the AI reviewer make sense?
    • Yes
  • If you added custom AI rules, were they accounted for in the review?
    • Yes, but it didn't count the fixed issue on the analytics page.
  • Would you trust this AI reviewer on real PRs after this test?
    • 90% yes.
  • What made you smile?
    • It looks similar to the Cursor bot we have in GitHub.
  • What felt frustrating or broken?
    • You should check how the AI reviewer resolves issues described in rules, because it feels like there are some troubles with this part.
  • On a scale of 1 to 10, how would you rate the overall quality of the AI reviewer?
    • 8
  • Are there any additional notes you would like to share?
    • It's not possible to see outdated comments.
    Screenshot 2026-07-03 at 13.52.58.png