Models and routing

Each kind of call — a project summary, a code analysis pass, an architecture pass, the synthesis, the Constitution, a Skill, a Skill review, a KEEL turn, a KEEL artefact — configures its own provider, model, temperature, output ceiling, timeout and retry count. That is what lets a cheap model do the cheap work and a reasoning model do the reasoning.

Structured calls send a schema, validate the reply against it, and retry once with the validation error appended before failing. Callers get typed data or an error — never raw prose to parse.

• The platform’s own models come out of a free allowance attached to your plan.
• Models reached by a key you connected are billed to you by that provider, and are selectable on any plan — your credential is not ours to ration.
• A model the server holds no credentials for is shown as unreachable rather than hidden. Not on your plan, not configured, and not reachable are three different reasons, and only some of them are yours to fix.

A model change applies to your next analysis. Nothing already produced changes, because every Skill version records the model that made it.

The reasoning steps return schema-valid output whose every field is UNKNOWN, with an explicit note that no provider was configured. Nothing is manufactured to fill the shape, and the eleven deterministic steps are unaffected.