# Make a Model better

When a Model gets something wrong, you say what should change, in plain words. Facet works out the fix, tells you what it costs, and puts it live only when its evals pass and nothing the live version gets right breaks. You never edit prompts or training data yourself.

## 1. Ask for a change

Tell your coding agent what should be different:

> Tickets about refunds are billing, not account.

Add the inputs it got wrong if you have them. They become examples for the new eval, and some are hidden to check the fix.

Your agent calls `request_change`, which makes a Change at once. It shows on the Model's **Changes** tab in the web app too.

## 2. Read the plan

Facet answers with a plan:

- **The eval** your words imply, such as "labels refund requests billing", with its examples.
- **How often today's version misses it,** measured on those examples.
- **The quote:** the most the fix can cost. Facet picks the approach: rewriting the Model's Instructions costs cents and takes minutes; building a new version costs more and takes longer. It tries the cheaper one first.
- **Setups,** when the fix tries more than one: another base or window size, each with its reason, what 1,000 calls would cost on it, and the longest input it takes. Only one goes live: the best on the Model's validation rows, or the cheaper per call when two score within the noise.

If today's version already passes, Facet says so and suggests closing the Change.

Until you approve, the Change waits as **Planned** on the **Changes** tab, with its price and both buttons on its page.

## 3. Approve

Approve the quote and the fix starts, and the Change gets its number. It never spends more than the quote. If it runs out first, it stops and shows how far it got, with **Raise quote and continue** if a bit more would finish it. Approving from an agent asks you to confirm in your agent's window, since it spends money.

## 4. Watch it run

The fix runs on its own; you can close the tab. Its page shows each step as it happens. **Stop** ends it at any point, and you pay only for what ran.

## 5. Review and merge

When the fix is ready, its page shows every eval side by side with the live version: what got better, what got worse, and what stayed the same.

- By default, a fix merges on its own when every required eval is green and no eval the live version passes turns red. See [When a version goes live](/docs/evals#when-a-version-goes-live).
- With **Approve each merge** on, under the Model's **Settings, Merging**, it waits for a Model owner to choose **Approve and merge**.

Once merged, the fix is the new live version (also called Latest), and every caller using the Model's name gets it on their next call.

## Roll back

If a merged change turns out wrong, open **History** and choose **Roll back** on the version before it. Callers get the earlier version at once. Rolling back can itself be undone.

## Keep it current

A Model's owners can turn on **Suggest retraining** under the Model's **Settings, Retraining**. On a schedule, weekly by default, Facet scores the live version on dataset rows added since it merged. When an eval drops, Facet drafts a Change with a quote and emails the Model's owners. Nothing trains until a person approves it, and the checks themselves stay under a monthly cap you set.

## Compare with a frontier model

Once a version is live, you can score it next to the frontier model you use today, on the same hidden examples. Choose **Compare against…** on the version's page, or ask your agent, which calls `compare_version`. You see correct answers on each side, and what 1,000 correct answers cost on each. It is quoted first and runs only with your yes. A workspace Owner chooses which providers hidden rows may go to under **Settings, Frontier comparisons**.

## Next steps

- [Write good evals](/docs/evals): a sharper eval makes a cheaper fix
- [Pricing](/docs/pricing): what fixes cost
- [Concepts](/docs/concepts): Changes and Versions in one picture