# Run a first baseline

Canonical: https://savrn.com/cloud/docs/evaluations-first-baseline

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

A baseline gives later changes a stable point of comparison on the same task. This SAVRN Cloud walkthrough follows a frozen evaluation dataset into a sample run and explains which identities, settings and case outcomes must remain available when a candidate is reviewed.

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

## Run the local baseline

Open [Evaluations](https://savrn.com/cloud/console/#/evaluations) and choose **New evaluation**. Select the model, task environment and a dataset marked **Frozen evaluation set**, then start the simulation. Inspect deterministic sample scores and per-case results; **Export receipt** retains their label. Model detail **Evaluate this model** preserves model context. Candidate evaluation requires a dataset different from its training dataset.

A baseline is a reference point for the same task, not an arbitrary leaderboard score. The local result is synthetic and should retain that label in exports.

## Example report boundary

This illustrates the minimum provenance questions; it is not the application's import or API schema:

```json
{
  "simulation": true,
  "schema_status": "illustrative",
  "model": "Research 8B",
  "task_revision": "demo-1",
  "result_kind": "synthetic fixture",
  "measured_production_score": null
}
```

Inspect whether the local screen explains what ran and where to find the result. Do not fill a missing measured score with the fixture value.

## Before connected services launch

Freeze approved inputs, rubric, model offer and settings. Run a bounded sample, retain per-case outcomes and reconcile usage. Include failures, exclusions and uncertainty. Establish the minimum quality and operational thresholds before testing a candidate. A baseline should remain reproducible after later environment edits, model promotions and provider changes; preserve the original revision rather than overwriting its record.