smevals keeps harness failures out of model scores

smevals is an MIT-licensed CLI framework that records immutable runs for task and model-or-harness pairs, then grades them with ordered checks and reusable checkers. Why it matters: A network drop or crashed runner is a harness failure, not a model result. Keeping those records without grading them makes a routing comparison easier to audit.

Try this: Build a ten-task eval around a recurring review job, add a deterministic citation or schema checker, and compare two model or harness configs before changing the default.