The Pre-Registered ML Manifest: an open specification for committing the success criteria of an AI evaluation (metric, threshold, dataset, seed and five further fields) to canonical bytes and a SHA-256 digest before the run. Includes the canonicalisation rules, the manifest field semantics, the exit-code contract used by CI gates, and an amendment chain. Published under CC BY 4.0 and archived at DOI 10.5281/zenodo.20177839, with 21 conformance vectors and 14 reject vectors that any implementation can be checked against.
Fund this project