BBEH routing: cost and answer quality

Explore the final performance of policies chosen on calibration. Each symbol is a distinct saved policy; there is no new selection or frontier fitted to final answers.

Cost and accuracy

Cost basis Answer model cost · all methods Answer + known router fees

Filter methods or policy IDs

Black circles: always Luna / always Sol. Dashed line: analytic expected random routing between those arms, with zero router inference fee.

Select a symbol or a table row for its uncertainty. Axis bounds may zoom to the data. Error bars are pointwise 95% intervals for the selected policy, not a confidence band for the curve.

Selected policy

Intervals resample whole families within each task, retaining the observed task weights. They are conditional on these tasks and do not provide simultaneous coverage across policies or a guarantee on future requests.

All calibration-frontier policies

Final measurements alongside the calibration values used to include each policy. Always-arm baselines are shown once. Calibration cost includes encoder fees for embedding methods and excludes unpriced hosting for native/CPU methods.
Choose	Method	Final $ / 10,000	Final accuracy	Sol share	Calibration accuracy	Calibration $ / 10,000

Interpretation and provenance

True task metadata is an assisted control: it uses the source task label, which an ordinary question-only router does not receive. Predicted task from text is a separate method.

The random line is an expected mixture, not an observed random run. A policy’s calibration-budget random comparator keeps its calibration mixing probability; its final spend need not match the policy.

Exact input file hashes

This file is self-contained and makes no network requests. It only formats the supplied reports; it performs no grading, training, or policy selection.