Safety Under Scaffolding: How Evaluation Conditions Shape Measured Safety

ArXi:2603.10044v1 Announce Type: cross Safety benchmarks evaluate language models in isolation, typically using multiple-choice format; production deployments wrap these models in agentic scaffolds that restructure inputs through reasoning traces, critic agents, and delegation pipelines. We report one of the largest controlled studies of scaffold effects on safety (N = 62,808; six frontier models, four deployment configurations), combining pre-registration, assessor blinding, equivalence testing, and specification curve analysis.