What's this about?
A cheap decision model handles ordinary comparisons, while low-confidence cases go to a more capable and expensive generative judge.
What is a confidence cascade?
It is a triage line: routine cases pass through the first checker, ambiguous cases receive expert review. Savings depend on confidence being well calibrated.
Why does this matter?
author-reported preprint; 21 HF upvotes.
- Claimed ordinary-task gap is under three points at 0.36% of comparator fee.
- Derivation checking and persuasive wrong answers remain weaknesses.
- The frozen cascade retains 99% of comparator accuracy on tested benchmarks.