Measuring Five-Nines Reliability: Sample-Efficient LLM Evaluation in Saturated Benchmarks

Sample-efficient LLM evaluation method for measuring five-nines (99.999%) reliability on saturated benchmarks, addressing rare failure probability estimation critical for reliability-sensitive deployment scenarios.