A model registry is the system of record for an organization's machine learning models. It stores each model version alongside its lineage, the code, data, and parameters that produced it, its evaluation results, its approval status, and its lifecycle stage, from staging through production to archived. When something answers a customer or scores a transaction, the registry is how the team knows exactly which model version it is, where it came from, and who approved it.
The registry is the control point that makes disciplined deployment possible. Promotion from staging to production becomes a governed action with review gates rather than a file copied to a server. Rollback becomes trivial because every previous version remains registered and retrievable. Audits become answerable: a regulator asking what model made a decision on a given date gets a precise answer, which is why registries anchor compliance under frameworks like the EU AI Act and ISO 42001. Common implementations include MLflow's registry, SageMaker Model Registry, Vertex AI, and Weights and Biases. In LLM-based systems the same discipline extends beyond weights to prompts, retrieval configurations, and evaluation sets, since in practice the prompt version changes system behavior as much as the model version does.
At arosplatforms a registry, or a lightweight versioned equivalent for smaller teams, is part of every production deployment we build. We register models with their evaluation evidence and model cards attached, wire promotion into canary deployment pipelines, and version prompts with the same rigor, so every behavior change in production is traceable to a reviewed, reversible release.