Vals independently tests frontier AI models on real-world tasks, building private, domain-specific benchmarks that resist leakage, cherry-picking, and academic proxy metrics.