Reproducible AI Benchmarks Require Frozen Dataset, Oracle, and Metrics Bundle

A coding-agent benchmark score is meaningless unless the dataset, oracle, metrics, and environment are locked together as a single versioned bundle before any evaluation runs. Publishing a success rate without these frozen components is comparable to reporting a race time without confirming the distance or the position of the finish line. The proposed framework organizes tasks into a structured directory, each with a stable identifier, a frozen prompt, fixture files, and an oracle definition. A manifest file hashes every component that could affect scoring, and this manifest must be generated before any agent process begins to prevent post-hoc oracle changes. The harness presented is a reproducibility scaffold for future benchmarks, not a record of a completed evaluation.