Model ids
Three names exist for one model, and only two of them are the registry's.
| Name | Example | Whose |
|---|---|---|
| Sku id | anthropic/claude-sonnet-5 | Vitna's |
| Slug | claude-sonnet-5 | The market's |
| Call name | whatever the API expects in a request body | The provider's |
For the first-party models those three currently coincide. That is a fact about how these providers happen to name things, not a rule anybody promised: OpenRouter already spells Opus 4.5 as claude-opus-4.5, and a dated snapshot id or a -fast variant would break the coincidence tomorrow.
So the call name is written down by hand, the same way the catalog's listing id is, and a test keeps it honest.
A registered sku can still be unnameable
Three reasons, each a property of the sku's market rather than a gap in a file:
| Reason | Why |
|---|---|
hosted | The same model is served under a deployment name the customer chose. There is no id Vitna could know, and guessing one produces a config that 404s. |
open_weight_serving | Each host spells somebody else's weights its own way in its own catalogue, which Witness does not read. |
first_party | A back-catalogue model no workload names. Adding it is one line, and it belongs to the change that points a workload at it. |
Unnameable is a refusal, not an omission
A sku that cannot be named produces a named refusal from `aperture:config`, carrying the reason above.
The alternative would be a config quietly missing a line, which leaves the consumer running whatever it already had while the record believes it emitted a decision.
