Pradra StudioDocs
Sign in

Models & evaluation

A model is a trained, versioned artifact that backs exactly one capability at a time. This page covers how a candidate earns trust — evaluation — and how a trusted model reaches, and leaves, production.

The model lifecycle

StageMeaning
Trained / DraftFreshly produced by an experiment.
EvaluatedScored on held-out data; report available.
PublishedPassed the gate; eligible for production.
ArchivedSuperseded or retired.

Models are versioned automatically: any change to configuration or data produces a new version, while metadata-only edits do not. Version numbers only ever go up — nothing is overwritten.

Evaluation

Before a model can be trusted, it is evaluated on data it never saw during training, producing a metric report: accuracy, precision, recall, F1 and friends for classification (with a confusion matrix you can read at a glance), error measures for regression.

That report is then checked against the capability's acceptance criteria. The gate is strict on purpose: if any rule fails, the model cannot be promoted — even if its headline metric improved. A model that is more accurate overall but worse on the class you care about does not get a pass because one number went up.

Promotion and the registry

Promotion is governed by the model registry, which moves a published model through Registered → Staging → Production → Archived. Two invariants hold at all times:

  • Only models that passed acceptance can enter or advance.
  • Exactly one version is in production for a capability at a time.

Rolling back is a pointer switch, never a retrain. If the new model disappoints in the real world, you point production back at the previous version and are done in seconds — the artifact never went anywhere.

Working with models in the Studio

  1. Go to Models in the sidebar for every model in your workspace, with status and version.
  2. Open a model's detail page: its evaluation report (metric tiles, charts, confusion matrix for classifiers), the experiment that produced it, and its lifecycle actions.
  3. Publish a model that passed evaluation, then promote it to production for its capability. Both actions re-check the gate — the answer is always current, never cached.
  4. A model that fails shows exactly which rule blocked it, so the next run knows what to fix.

Next