Skip to content
Ódinn ForgeÓdinn Forge Docs

Smart Routing (Darwin)

Darwin is an experimental model-routing scorecard. It records measured outcomes in the local ledger and computes a transparent weighted score.

Verified success and policy compliance carry more weight than speed or cost. A small-sample uncertainty penalty prevents one lucky run from dominating future choices.

Enable Darwin

Section titled “Enable Darwin”
Terminal window
odinn experimental enable darwin --confirm-impact

Record an observation

Section titled “Record an observation”
Terminal window
odinn experimental darwin observe \
  --run <run-id> \
  --provider <provider> \
  --model <model> \
  --verified true

Add a task class when comparing routes for a particular kind of work.

Inspect scores

Section titled “Inspect scores”
Terminal window
odinn experimental darwin stats --task-class bug-fix

The result explains recorded observations, uncertainty, reliability, performance, cost, and policy signals. Prompts and secrets are not stored in observation rows.

Choose a route

Section titled “Choose a route”
Terminal window
odinn experimental darwin choose --task-class bug-fix

The selection is a reasoned recommendation from recorded evidence, not a guarantee that the next remote model call will behave identically.

Use it well

Section titled “Use it well”
  • Compare routes only after collecting representative observations.
  • Treat verified outcomes as more meaningful than raw speed.
  • Keep task classes narrow enough to describe comparable work.
  • Investigate policy violations instead of allowing cost or speed to mask them.
  • Revisit a choice when providers, models, or pricing change.

Provider service behavior remains outside Ódinn’s local guarantees. See Models and providers for configuration and Public beta status for classification.