UK AISI Adopts EvalEval for Reproducible Benchmark Results

The United Kingdom AI Security Institute and the EvalEval Coalition have formalized a collaboration to standardize and publicly release artificial intelligence benchmark evaluations through EvalEval's open infrastructure. Building on research initially presented at a joint NeurIPS 2025 workshop, the partnership transitions from theoretical framework development into active deployment of the Every Eval Ever schema and Evaluation Cards platform. This initiative directly addresses a critical bottleneck in AI safety research: the widespread inconsistency and lack of reproducibility in published model evaluations. The institute is now publishing verified benchmark results, methodological configurations, and contextual metadata via Evaluation Cards. The initial release covers five primary benchmarks and two complementary cybersecurity evaluations, examining six frontier language models: Claude Opus 4, Claude Opus 4.5, Claude Opus 4.6, GPT-5, GPT-5.2, and GPT-5.4. The shared data accompanies the institute's recent publication analyzing how inference-time compute and evaluation protocols influence model performance. By detailing setup parameters and token usage limits, the release provides verifiable reference points that allow researchers to isolate how specific testing conditions affect reported scores. EvalEval's infrastructure resolves fragmentation in the evaluation ecosystem by consolidating benchmark metadata, execution data, and model specifications into standardized records. This approach complements existing UK AI Security Institute methodologies such as OptStop for efficiency and HiBayES for statistical rigor. The combined framework enables the scientific community to conduct reliable meta-research, identify reporting gaps, and compare capabilities across varying experimental setups without incurring prohibitive re-execution costs. The publication of structured, transparent evaluation data supports both academic inquiry and policy development. As AI deployment scales, standardized reporting mechanisms become essential for verifying system capabilities and assessing risks. The collaboration establishes a scalable model for open evaluation science, with the EvalEval Coalition actively encouraging broader participation from AI safety organizations and research groups. Future iterations of the platform will continue to refine schema definitions and expand coverage across emerging benchmark categories, reinforcing transparent, reproducible standards for frontier AI assessment.