Skip to content
Docs

Evidence model

DialogAssert normalises each agent run into canonical evidence:

  • Customer turns
  • Assistant responses
  • Tool calls
  • Tool results
  • Final output
  • Acceptance criterion results
  • Findings and severities
  • Judge metadata when LLM judging is used

Normalising evidence is what makes framework-agnostic evaluation possible: the same acceptance criterion can be judged against an OpenAI Agents SDK run, a LangChain run or an HTTP agent, because all three arrive at the evaluator in the same shape.

It is also what lets DialogAssert distinguish the card the customer selected from the card the agent acted on — a distinction the final message often hides.