Docs
Reproducibility
The records needed to understand and, where practical, reproduce an analytical result later.
A reviewer may need to understand an old result after software, configurations, or models have changed. Reproducibility begins by recording the context that influenced interpretation.
Interpretation record
Where relevant, analytical output records:
- component, model, or method name;
- exact version;
- configuration and threshold version identifiers;
- schema and dataset versions;
- prompt identifiers and versions when language models are involved;
- processing timestamp;
- code revision;
- stochastic seed when practical.
Comparable does not mean identical
Scores produced by different versions are not assumed to be directly comparable. A changed model, baseline, configuration, or input population can change the meaning of a number even when its scale appears unchanged.
Bounded claim
Reproducibility records help explain how a result was produced. They do not validate the result’s real-world meaning or establish that a model performs adequately for operational use.