# Promote, deploy and roll back

Canonical: https://savrn.com/cloud/docs/lab-promote-deploy-rollback

SAVRN Cloud · Coming soon. Public documentation and a browser simulation are available; connected services are not yet available.

Evaluation, approval and activation establish different facts about a model candidate. This SAVRN Cloud preview guide follows those decisions through a simulated promotion and rollback, showing which artifacts and serving evidence a connected release would need to retain.

**SAVRN Cloud · Coming soon · Public preview · Browser-local simulation · No live compute or payments**

## Review a candidate

Open [Candidates](https://savrn.com/cloud/console/#/candidates) after inspecting a training run. Compare the candidate's model, dataset context and associated evaluation evidence. Read its current state before using any promotion action. A simulated candidate is a local record, not a downloadable trained model.

Choose **Evaluate** with a **Frozen evaluation set** whose ID differs from the training dataset. A passing rubric permits **Approve**; record an independent demo reviewer and note, then **Approve local candidate**. **Promote**, followed by **Create demo deployment**, creates a named synthetic deployment in [Deployments](https://savrn.com/cloud/console/#/deployments). Confirm its candidate relationship; its readiness label still represents simulation.

## Preserve the review decision

Evaluation completion establishes that a result record exists. Approval establishes that a reviewer accepted a proposed action. Activation establishes that the intended serving change occurred. These are separate facts.

**Roll back promotion** marks its demo deployment **deleted** and returns the candidate to **approved**. Inspect both records. This rehearses the decision flow; it does not restore actual weights or reroute live traffic.

## Before connected services launch

Bind promotion to an immutable artifact, qualified serving configuration, acceptance thresholds and an authorized decision. Canary traffic and verify observed model identity. Track errors, quality, latency and cost during the acceptance window. Maintain a tested rollback target and preserve the original offer history. A failed activation must remain visible; never mark the release complete merely because an approval or deployment request succeeded.