Model routing
Horme uses deterministic rules to choose among accepted models available on your Mac.
How selection works
Horme considers task class, complexity, risk, provider status, runtime model availability, explicit model pins, user defaults, dependencies, continuity, and review needs. Routing follows product rules and can change when a provider or model is unavailable.
Deterministic, not self-learning
The current Alpha does not learn routing policy from your code or send routing decisions to a network learning service. A failed implementation can use one bounded escalation when an accepted rule covers it.
What is not exposed
The product reports which model owned each step and why that assignment fits at a useful level. It does not publish private planner prompts, scoring weights, or hidden evaluation instructions.