Caustic · measured pilot

The recall–savings frontier

Real data, not a model. Drag the recall you need; read off the search savings — for Caustic's adaptive routing, a standard fixed-budget baseline, and the oracle ceiling (best possible if you knew the answer in advance). Measured on PRM800K: 12,500 math problems, BGE-large embeddings, 300 held-out queries, recall scored against exhaustive search.

Caustic (adaptive)

—

Fixed-budget baseline

—

Oracle ceiling

—

Recall you need95%

Caustic vs baseline—

Measured, not modeled. Each curve is the real recall-vs-savings frontier from the PRM800K pilot (BGE-large, k=8 regions, 300 held-out queries; savings = 1 − fraction of corpus searched, recall scored against exhaustive search). Caustic's adaptive routing sits above the fixed-budget baseline and a few points under the oracle ceiling. Your corpus will have its own frontier — that's what a design-partner pilot measures on your stack.