Missing outputs belong in an AI writing-QA report.
If a scorer returns ratings for 79 of 80 cases, agreement on the returned cases cannot explain what happened to the missing response.
Track coverage and validity beside agreement. Keep an explicit distinction between a missing...
🚨 AI can now do the work. But who checks whether it's doing it RIGHT?
Google just announced its new Gemini agent — AI that can plan, execute complex tasks, work across business applications, and keep running even after you close your laptop.
A higher AI pass rate can still hide a bad release.
In the fictional worked example inside my AI QA Regression Kit:
• Baseline: 92% case pass rate
• Candidate: 94%
• Three new regressions
• One critical unauthorised tool action