Skip to main content
A success criterion states what the transcript must demonstrate. For example, a booking assistant should confirm the requested date before requesting a reservation.
Criteria authoring and stored-result reads are available. Run conversation evaluation currently returns the existing record; it does not calculate a fresh score.
With agent edit access, review platform_settings.evaluation.criteria. Each criterion has a stable id, a name, and a conversation_goal_prompt of up to 2,000 characters.
1

Write an observable criterion

Describe the required evidence and how to handle missing information. Keep unrelated outcomes in separate criteria so one success cannot hide another failure.
2

Preserve the criteria list

Save through the Analysis editor or agent draft. Review the complete criteria list and identifiers before saving a replacement.
3

Review stored results

Read analysis.evaluation_criteria_results in conversation details. Inspect each result and rationale, plus score and max_score when present. Compare them with the relevant transcript turns.
Keep unknown separate from success and failure. It does not provide enough evidence to claim the criterion passed.
No. Saving a definition does not recalculate existing conversations.