A Guardrails Library That Publishes Its Misses

The new jamjet-guardrails library offers a deterministic approach to LLM input and output validation, prioritizing transparency over performance benchmarks. Unlike many existing tools that provide only binary risk scores, this library publishes precision and recall metrics for every check against a committed corpus. It explicitly documents its limitations, including specific case IDs for known failures, ensuring developers understand exactly where the tool might fall short. The library includes nine built-in checks—covering areas like PII, secrets, and structural injection—with zero runtime dependencies. By requiring that all performance data be gated in CI, the project forces developers to confront the accuracy of their guardrails directly. It also provides a clear migration path for users of the now-archived llm-guard, while maintaining compatibility with broader ecosystems like guardrails-ai, offering a robust, audit-friendly alternative for securing LLM interactions without relying on opaque black-box models.