OracleEvaluations & LLM Judges

12 min Intermediate 11 steps

An agent never answers the same way twice. How do you test it before it ships, and know when it got better?

Median correctness
?
Above threshold
5 by eye
Tool functions hit
?
AI Units / eval run
0
PayablesBOsholds, invoicesPolicyassistantAGENTPayables SOPPayables app
  • request
  • rows / data
  • model tokens
In this lessonStep 1 of 11