Proposals · sp-0003
Adopt a common format for evaluation records
draftDiscussion and state illustrated
- Problem
- Evaluation records written in different formats cannot be compared or aggregated.
- Change
- Store evaluation records in an established common format for AI evaluation results, including failed and aborted runs.
- Consequences
- Evaluators export records in that format; the site reads them directly.
- Source
- Standard Explorer §5, Evaluation protocol.
Discussion
The steering group decides after the discussion and publishes the decision with its reasons here.
Which existing building blocks would fail conformance after this change? A list would help maintainers plan new versions.