The method, in the open.

SagaBench's measurement method — the counterfactual scoring against doing nothing, the replication design, and the group-level findings summarized on this site — is written up in a full technical paper. It will be released together with a public replay verifier, so every number in it can be re-run by anyone.

That hasn't happened yet. The paper has not been published and has not been submitted for peer review. When it is, the DOI will appear here — and nothing on this site depends on it.

What the numbers stand on in the meantime is their own reproducibility: 2,415 internal validation runs, zero deviations. If you want to check that rather than take it from us, we'll take your technical team through the method in detail and give you a run to replay yourself. Talk to us →

What we don't claim: published or peer-reviewed status, model rankings, or safety guarantees. What we do claim, you can re-run.