Concepts
Every Model is built from the same four parts, borrowed from how software teams already work.
| Part | What it is | Software analogy |
|---|---|---|
| Model | What Facet builds for one task, like support-triage | A repo |
| Change | A plain-language request, which Facet’s agent turns into a candidate Version | A pull request |
| Evals | Plain statements every candidate is scored on, each green or red | CI |
| Version | A built model you can call; the live one is Latest | A release tag |
Model
A Model does one narrow, typed task. It has what it takes, what it returns, and its Instructions (the prompt Facet writes and keeps current for you). Callers pass its name as model. Each Model has its own evals, examples, Versions and policy.
Evals
An eval is one plain statement of what good looks like (“picks the right team”), with examples, a way to check it (a code check, a score or a judge) and a bar to clear. Together they are the Model’s scorecard: green when a Version clears the bar, red when it doesn’t. Some examples are hidden: only the check runner sees them, so no Change can train on its own test.
Change
A Change starts when you say what should get better, in plain words. It moves through these states:
| State | What happens |
|---|---|
| Planning | Facet drafts the eval your words imply, measures Latest against it, and quotes the cheapest approach |
| Planned | It waits for a person to approve the quote. If Latest already passes, Facet says so and you can close it |
| Queued | Approved, and waiting for a free slot. It now has its number |
| Running | It works within its quote: better Instructions first, training only if needed, then every eval |
| Ready | It shows what got better and worse. By default it merges on its own; with Approve each merge on, it waits for a Model owner |
| Merged | Its Version is now Latest |
| Failed or Closed | It ended without merging. You pay only for what ran |
A Change has no Version until it merges, and can be stopped at any time.
Version
A Version is an immutable bundle: base model, Instructions and, if trained, an adapter. Latest is the Version callers get by default; name@vN pins one. Rolling back points Latest at an earlier Version, and can be undone.
Policy and spend
A workspace’s Owners run it: members, keys, billing and the spend limit. Each Model also has Model owners, who set its rules, such as whether each merge needs their approval and who may ask for changes. Each workspace has one monthly spend limit; at 100% Changes and calls stop until an Owner raises it or the month ends. You see every price before anything runs, and a Change never spends more than its quote. See Pricing.
Next steps
- Quickstart: the four parts in action
- Make a Model better: a Change from request to merge