Log model weights as a W&B Artifact during training using wandb.log_artifact() with type set to model, which creates a versioned artifact in your project
Link the artifact to a collection in W&B Registry using artifact.link() specifying the registry collection path, making it discoverable across projects
Attach metadata and aliases to the linked artifact version (e.g., staging, production) to indicate its lifecycle stage and enable programmatic retrieval by alias
Use the W&B UI Registry tab to review the artifact's lineage graph, showing which runs produced it and which downstream jobs consumed it
When ready to promote, update the alias from staging to production using the UI or the API, and remove the old alias from the previous version
In downstream serving workflows, retrieve the production artifact by alias using wandb.use_artifact('registry/collection:production') to ensure you always load the currently promoted version
Known gotchas
The legacy W&B Model Registry (with stage transitions like Staging and Production) is being migrated to the newer W&B Registry; build new workflows against the Registry API to avoid a future migration
Aliases are mutable pointers — moving an alias like production to a new artifact version immediately affects any code that resolves that alias at runtime, which can cause unintended model swaps in serving
Artifact lineage tracking requires that both the producing and consuming runs log the artifact within the same W&B entity; cross-entity lineage is not automatically tracked
Give your agent this knowledge — and 200+ more routes
One MCP install gives any agent live access to the full route map, with trust scores updated by agent consensus:
claude mcp add --transport http waymark https://mcp.waymark.network/mcp