When No Benchmark Exists: Validating Comparative LLM Safety Scoring Without Ground-Truth Labels

The contribution here isn't a new benchmark but a meta-level argument: a formal account of when an audit can stand in as deployment evidence even before labeled ground truth exists. That distinction matters because most safety tooling assumes benchmarks precede deployment, when in practice the order is often reversed.

This sits in direct conversation with two recent pieces. ML-Bench&Guard (early May) built multilingual safety benchmarks grounded in regional regulations, but still assumed labeled data was available to validate against. FinSafetyBench similarly produced a reusable red-teaming methodology for financial deployments, yet both works operate in the regime where benchmark construction is possible. The current paper addresses what happens before that regime exists, which is arguably the more common situation for teams deploying into niche languages or sectors. The leaderboard critique covered in 'Why Global LLM Leaderboards Are Misleading' adds relevant pressure here: if global rankings are already unreliable, the case for rigorous pre-benchmark audit contracts becomes stronger, not weaker.

Watch whether SimpleAudit's validity framework gets adopted or cited by domain-specific benchmark projects like FinSafetyBench or ML-Bench in their next iterations. If it does, that suggests the field is converging on a shared evidentiary standard for pre-benchmark safety claims.