Model Router

Status: FUTURE (Phase 5 execute, Phase 7 learn). No router service exists.

Contract

Input: task type, project id, requirements blob, budget (max candidates, max cents, max ms).

Router queries Pastes aggregates for that project and task class.

Output: model_version_id, prompt_version_id, temperature, tools, evaluation_policy_id, candidate_count.

Then: generate -> evaluate -> store -> update rolling stats.

Example matrix (illustrative, not measured)

These are not live rankings. They show the shape of the artifact the router will read:

• UI generation -> Model B
• PHP debugging -> Model C
• SQL generation -> Model A
• Brand design -> Model D
• Data analysis -> Model C
• Vision inspection -> EE VIZN model
• Documentation -> Model A

Until Pastes has scored runs, publishing a matrix would be marketing. Do not.

Cold start

Phase 5 uses staff-declared priors in Command. Phase 7 replaces priors with empirical win rates once n is large enough per cell (task x app). Below threshold, explore (extra candidate) instead of exploit.

Failure policy

If evaluations disagree with humans more than a project threshold, freeze auto-routing and require Command approval.

Diagram

See diagrams/model-router.mmd.