Google DeepMind announced a double-blind evaluation pilot on August 27 to test models without sharing confidential questions with the model developer or giving the evaluator proprietary model weights.
The pilot runs a Gemini Flash Lite model in Google Cloud Confidential Space against prompts supplied by outside evaluators. The Singapore AI Safety Institute, OpenMined, AVERI and MLCommons are listed as partners. Google says participants can verify how a test runs while keeping each party's protected inputs from the other.
Keeping test questions private could reduce benchmark contamination, where a model appears stronger because its developers have already seen the questions. The pilot addresses that exposure route; it does not demonstrate that every source of contamination has been eliminated.
This is an evaluation infrastructure pilot. It does not release a new Gemini model or a consumer feature.