Virtual Model Variants — brainstormrouter/fast|cheap|quality
2026-03-23
What We Built
Three virtual model names that resolve to optimal concrete models using live performance data:
- brainstormrouter/fast → lowest latency (uses ModelPerformanceTracker latencyMean)
- brainstormrouter/cheap → lowest cost meeting quality floor (default 0.7)
- brainstormrouter/quality → highest quality score
Selection uses Thompson sampling arms from the 7-day performance window. Min 10 samples required. Filters out models with low validity (<0.5).
Lockstep Checklist
- [x] API: Virtual model resolution in router
- [x] Tests: 9 tests for identification, resolution per variant, edge cases
- [x] Docs: Ship log