livenerf
A long-running, append-only benchmark that checks whether a frontier language model quietly gets worse after it ships, running the same frozen prompts and graders daily against a fixed baseline set by launch day.
livenerf tracks one question: does a model's behavior drift after release. It runs once a day for 30 days against a panel of GPQA Diamond, MMLU-Pro, competition-math and AIME questions that were screened in advance for being neither always right nor always wrong, so a change in accuracy is more likely to reflect real drift than question difficulty.
Everything about the run is fixed in advance and logged: the prompts, the CLI version, the grading method and the statistical protocol, which follows a published method for adding error bars to evaluations. It is built on Inspect, the UK AI Security Institute's open-source evaluation framework, and the current run targets Claude Opus 5.5 through a Claude Max subscription via headless Claude Code.
The first ten days form a baseline, followed by two ten-day comparison windows, with the first possible verdict roughly a month after the model's launch.