"Inspect AI: Writing Reproducible Evals and Safety Tests for LLM Systems"
Large language models rarely fail in obvious ways, and that is exactly why evaluating them demands more than dashboards, ad hoc prompts, or borrowed benchmarks. *Inspect AI* is written for experienced practitioners building, shipping, or governing LLM systems who need rigorous, repeatable evidence about capability, reliability, and safety. It treats evaluation as an engineering discipline: one grounded in specifications, threat models, and operational decision rules rather than intuition or one-off red-team exercises.
The book shows how to turn product goals and policy boundaries into measurable behaviors, design high-signal eval datasets, build trustworthy graders and rubrics, and define thresholds that support real release decisions. It also goes deep on reproducibility: versioning models, prompts, datasets, and graders; capturing run provenance; and separating true regressions from benchmark drift. From end-to-end system evals for retrieval, tools, and memory to adversarial safety testing, prompt injection robustness, and CI/CD-integrated EvalOps, readers will learn how to construct evaluation programs that remain credible as systems and attacks evolve.
Designed as a technically dense, implementation-aware guide, the book assumes familiarity with modern LLM application architecture and software delivery practices. Its distinguishing strength is its focus on the full lifecycle of evaluation: not just writing tests, but maintaining them as durable operational assets for production AI systems.
