The UK AI Security Institute is publishing benchmark methods and setup details through EvalEval’s Evaluation Cards.
The UK AI Security Institute is using infrastructure from the EvalEval Coalition to share AI evaluation results in a more consistent format. The announcement, published by Hugging Face, says the release includes verified results, context, and configuration details for five benchmarks.
Those benchmarks are HealthBench, FrontierMath, Humanity’s Last Exam, SWE-Bench Pro, and Terminal-Bench 2.0. The data accompany AISI research on how inference-time computing and evaluation rules affect frontier language-model performance.
The infrastructure has two main parts. Every Eval Ever is a shared reporting schema: a standard set of fields for describing an evaluation. Evaluation Cards organize those details into records covering benchmark information, model information, and the specific evaluation run.
That matters because a benchmark score is not always a simple measure of model ability. Results can change with the evaluation protocol—the rules used to run and score the test—and with how much computing a model receives while answering. The announcement gives Humanity’s Last Exam as an example: models solved more tasks as they used more tokens, particularly when they received correctness feedback after each attempt.
With setup details recorded, researchers can ask whether two scores are genuinely comparable or came from different conditions. They may also be able to reproduce a run without guessing which model settings or testing rules produced the result.
This is a reporting standard and shared publishing system, not proof that every evaluation becomes reproducible. The coalition’s release does not establish how widely other organizations will adopt the schema or whether independent researchers can reproduce all the listed results.
For engineers, the useful next step is to inspect the Evaluation Cards before comparing benchmark scores—and to report future runs with the same level of detail.