Determinism and reproducibility#
Level: Beginner
You will: understand how configuration, seeds, named streams, and
fingerprints make a run repeatable.
Before you start: the Quickstart.
Services: None.
Reproducibility is part of FraudTwin’s data model, not only a property of the test suite. A run records enough identity and configuration information to explain how its outputs were produced and to compare them with another run.
Configuration and seed identity#
A validated configuration and seed determine the logical world: entities, payment identities, event timestamps, fraud scenarios, labels, and output artifacts. The run manifest records the resolved configuration, schema versions, counts, and output fingerprints.
Changing a relevant configuration value should produce a new run identity. Keeping the same configuration and seed should reproduce the same logical identities and values.
Named random streams#
FraudTwin derives named random streams for independent parts of generation instead of relying on one mutable global random state. This limits unrelated changes: adding or changing one controlled behavior should not silently replace every entity ID or payment identity.
Parallel workers may finish in a different order. That can change execution order, but it must not change logical IDs, event meaning, or recorded fingerprints.
Fingerprints are evidence#
Use the manifest and stage fingerprints to compare runs. A matching fingerprint supports the claim that the relevant logical output is the same; it does not mean that every incidental file timestamp or physical partition layout is identical.
Record the configuration, seed, package version, schema versions, and selected view or cutoff with any benchmark or evaluation result. Without those boundaries, a result may be repeatable but not explainable.