Pricing Audits in Verifier-Guided Test-Time Scaling: Proxy Reliability and Budgeted Candidate Selection

In verifier-guided test-time scaling, a cheap proxy ranks candidates and an audit obtains the target score of the selected output. The best split of a budget between generation and audits depends on proxy reliability among still-competitive candidates. We study this reliability through decision-granularity correlation: the proxy–target association within a prompt's top-ranked shortlist. We connect classical range restriction to an audit-efficiency threshold for linear proxy-assisted estimation, then evaluate candidate selection under a shared generation-and-audit budget. PickScore–HPS correlation falls from 0.84 within full diffusion prompt pools to 0.32 in their top 10%, and an API judge agrees with the target on 53.4% of near-ties. On text, an approximate correlation diagnostic accepts GSM8K and abstains on ARC. Across five diffusion models, we separate ranking gains from allocation gains. On 80 new Stable Diffusion XL prompts, reallocating a random-audit policy improves the composite preference score by 0.01196, while screening retains only 0.00076 over tuned random auditing (95% interval includes zero); higher audit prices favor screening.