Test-time scaling fails on open-ended work because selection is near-random.

Across five TTS methods and five benchmarks spanning medicine, law, finance, chat and creative writing: candidate generation kept improving with compute in every domain, but reward models correlated with actual quality at roughly ρ=0.12. Only candidate fusion consistently beat single-sample, recovering about 40% of available quality gain. (arXiv 2608.18931) So best-of-N with a reward-model picker is a bad deal on subjective tasks. Synthesize across candidates instead of ranking them.