TRACER ships formal LLM routing library

TRACER stands out because it turns LLM routing into a controlled systems problem instead of a vague “cheap model when it feels safe” heuristic. The guarantee is only about surrogate-teacher agreement on handled traffic, but that is still a practical bar for teams trying to cut inference spend without flying blind.

• –The surrogate-plus-gate design is the right shape for this problem: cheap local inference for easy cases, explicit fallback for edge cases.
• –Calibrating for maximum coverage subject to a teacher-agreement floor is more operationally useful than chasing a single aggregate score.
• –Auto-selecting L2D via the Pareto frontier removes method-selection guesswork, which matters when teams want a default policy rather than a research menu.
• –The qualitative audit outputs matter a lot here; routing models need to explain boundary cases, not just report averages.
• –The Banking77 result is promising, but the real test will be noisier, drifting production data where trace quality and label skew are less forgiving.