About PRISM

PRISM (Prompt Reliability Through Intelligent Semantic Multiplexing) is a benchmark and local desktop application for evaluating how small, locally runnable language models behave when the same question is presented through different, semantically equivalent prompt formulations.

It was created to make prompt-dependent model behavior directly observable and measurable rather than leaving it as an assumption.

PRISM sends the same question through several prompt formulations, evaluates each response independently, and reports where a model's behavior stayed stable and where it did not.

The desktop application performs benchmark execution locally, while the public website provides documentation and access to published benchmark results.