Double Blind Evals: Resolving the Dual Confidentiality Dilemma in AI Safety Auditing
AI-generated summary Authors present the Double-Blind Evaluation framework using secure hardware enclaves to provide cryptographic mutual secrecy for external AI safety auditing. The preprint reports evaluating Google DeepMind's Gemini 2.5 Flash Lite against a private benchmark corpus within an NVIDIA H100 secure enclave.