Urgent.News

What's breaking now, across thousands of outlets.

More in AI

When a Safety Reviewer Rejected Everything and Still Passed Its Test

This is a submission for DEV's Summer Bug Smash: Smash Stories powered by Sentry . ARGUS is a data-catalog governance swarm.

  • Safety reviewer rejected 86 of 112 proposed repairs in ARGUS test
  • Arbiter failed to distinguish between interpreting evidence and making new claims
  • Revised prompt introduced explicit categories to balance test outcomes

More from Sunday 23 August →