Inspect is an open-source framework for large language model evaluations, created by the UK AI Security Institute together with Meridian Labs. It gives researchers and developers a standard way to assess model performance across many dimensions, including coding, reasoning, knowledge, behavior, and multimodal understanding, and to do so in a reproducible, sharable form rather than through ad hoc scripts.
The framework is built from composable building blocks. Datasets supply labeled samples with input prompts and target answers. Solvers generate model responses, ranging from a single generation call to sophisticated multi-turn agents. Scorers evaluate the outputs using text comparison, model grading, or custom schemes. Tools let agents interact with the world through bash, Python, web search, and sandboxed execution. Inspect ships with more than 200 pre-built evaluations, supports over 20 model providers, and includes a web-based viewer and a VS Code extension for monitoring and debugging runs.
Because a government safety body built it and released it openly, Inspect has become common infrastructure for the kind of agentic and dangerous-capability evaluations that frontier AI governance increasingly relies on.
For a business reader, Inspect matters because shared, credible evaluation tooling is what turns vague safety promises into measurable, comparable results across different models and labs.