How UK AISI and EvalEval Are Making Benchmark Results Reproducible

The UK AI Security Institute is using EvalEval's shared infrastructure to publicly release evaluation results for AI models in a standardized format, making those results more reproducible and verifiable. Evaluation results are currently reported across many different formats and platforms without sufficient information to reproduce them, making it difficult for researchers to verify findings or compare results across studies. The two organizations are collaborating to adopt a shared schema called Every Eval Ever and an open platform called Evaluation Cards that combine benchmark data, evaluation-run information, and model metadata into a common structure. This effort aims to improve the quality and reliability of AI evaluation science by creating transparent reference points that help researchers understand how different evaluation setups influence reported model performance.