OpenAI on evaluating agent behaviour
Guidance for evaluating model behaviour on a specific task and different kinds of input.
- Publisher
- OpenAI
- Published
- Unknown · maintained documentation
- Reviewed
- Independent Skufosaur reproduction
- Pending
What the source supports
Supports task-specific criteria and coverage of ordinary, edge and difficult cases.
Where the conclusion stops
The examples and thresholds are context-specific. The guide does not certify these recipes, and its initial publication date is unknown.
What an independent check would need
Comparable tasks, documented environment versions, a quality threshold, baseline and candidate runs, complete available costs and reproducible artefacts.