Reading artifact manifests

The normative manifest schema and measurement schema live in the core contract. This page does not duplicate their field tables.

When comparing runs, check the producer and payload format, configuration digest, source revision, input content identities, and task semantics before comparing measurements. A configuration hash alone does not identify the dataset or weights. A copied path is provenance, not proof of compatibility.

null means not measured or not applicable; zero is an observed value. Sampled live process-tree memory and scheduler peak memory measure different things. Coverage counts are diagnostics; compare actual prediction keys with the requested index.

Core's runner uses measured_artifact for publication and measurements. Pass repo=algorithm_root explicitly to identify the algorithm checkout. Overwrite preserves a backup until publication succeeds; a failed replacement restores it, or reports its recovery path if restoration also fails. New outputs use one rename; overwrite uses two and can briefly hide the destination. Prefer immutable artifact paths for concurrent readers. See the core package for implementation ownership and limits.