Reproducibility in Production

Reproducibility in a production context means being able to recreate the exact conditions that produced a specific model version or output. This goes beyond the research-oriented reproducibility discussed in data engineering - it's about operational and regulatory needs. When a customer disputes a decision made by an AI system, you need to demonstrate what model was running, what data it received, and how it arrived at that output. When an audit requires you to justify your model's behaviour at a specific point in time, you need to recreate that model and its environment. Achieving this requires disciplined versioning of everything: training data, preprocessing code, model artefacts, serving configuration, feature pipelines, and infrastructure specifications. Container images and infrastructure-as-code help capture the serving environment. Comprehensive logging of inputs, outputs, and intermediate states provides the raw material for post-hoc analysis. In practice, perfect reproducibility is an ideal that's approached asymptotically. Floating-point non-determinism, external API dependencies, and the sheer complexity of modern AI stacks mean that exact reproduction is sometimes impossible. The practical goal is to get close enough to answer the questions that matter - what decisions were made, based on what information, using what model. Organisations in regulated industries often have specific requirements around reproducibility, and meeting them requires planning and investment from the outset.