Model Registry

A model registry is a centralized repository that stores trained machine-learning models along with their metadata and manages each model through its lifecycle. Think of it as version control for models: every trained artifact gets a unique version identifier, and models move through defined stages — development, staging, production, archived — with a clear record of who promoted what, when, and why. Where Git answers "which code is deployed?", the registry answers "which model is deployed?", a question that becomes surprisingly hard to answer the moment a team has more than one experiment in flight.

What the registry stores

Beyond the weights themselves, a registry captures the metadata that makes a model reproducible and auditable: training hyperparameters, evaluation metrics, the dataset version it was trained on, the code commit that produced it, and lineage back to the originating experiment. The large binary artifacts usually live in object storage; the registry tracks pointers, versions, stages, and approvals. Tools like MLflow and Weights & Biases implement the pattern for self-hosted and hosted setups alike, and the major cloud platforms ship their own equivalents. The registry entry is the difference between "a 14-gigabyte file named final_v2_REAL.safetensors" and an artifact you can actually trust.

Lifecycle and governance

The registry is the backbone of governance in MLOps. A model is registered when training completes, promoted to staging for evaluation, promoted to production when it passes, and archived when superseded — each transition recorded and, in disciplined teams, gated by an approval. When a production model misbehaves, the registry answers the questions that matter under pressure: which exact version is live, what data was it trained on, what changed since the last version, and how do we roll back to a known-good artifact in minutes instead of days. From the registry, an approved model is pushed to model serving and then watched with model monitoring; versioning is what makes safe rollout and rollback possible at all.

Why a sovereign operator should care

This is the same reproducibility discipline a sovereign operator wants over any critical system — a full, auditable history rather than an opaque artifact someone handed you. If you run local models, you already have an informal registry: a directory of GGUF files with version-bearing names and checksums. The formal concept scales that instinct. Anyone doing fine-tuning on their own hardware quickly accumulates dozens of checkpoints and adapter variants; without at least a lightweight registry — even a spreadsheet recording base model, dataset hash, training parameters, and eval scores per checkpoint — you will lose track of which variant produced which behavior within a week. Public model hubs act as de facto registries for open-weight models, with versioned repositories and revision hashes; pinning the exact revision you tested, rather than pulling "latest," is registry thinking applied to downloads.

The craftsman's rule

The registry embodies a rule every repair bench already follows: never modify the only copy, and always know which version you are holding. A model whose provenance you cannot state is a model you cannot debug, cannot reproduce, and should not trust in production — whether production is a bank's fraud system or the assistant answering questions on your own workshop server.

Getting started costs almost nothing. Self-hosters can run a full MLflow registry in a single container alongside their other services, but even without dedicated tooling the essentials are achievable with discipline: one directory per model version, an immutable naming scheme, a checksum file, and a short manifest recording base model, data, parameters, and evaluation results. The test of any registry, formal or improvised, is whether you can answer three questions six months later without guessing — what exactly is this artifact, how was it made, and would I get the same thing if I made it again. If the answers live only in your memory, you do not have a registry; you have a pile of weights.