I Put Jev Behind a TLA+ Spec and Ran 1,680 Chaos-Tested Pharmacy Decisions. Zero Wrong Verdicts.

A developer has demonstrated a rigorous approach to validating AI-driven decision-making in high-stakes environments like pharmacy systems. By utilizing Jev, a language model that outputs probabilities rather than prose, the author created a system where consensus can be model-checked. Using a TLA+ specification to define safety invariants and quorum policies, the author built a Rust-based kernel that treats the AI as a noisy oracle. The system was subjected to 1,680 simulated pharmacy decisions under intense chaos-testing conditions, including adversarial inputs and system failures. The results showed zero wrong verdicts, with the system correctly opting to escalate ambiguous cases to human pharmacists. This project highlights the importance of measuring the 'noise floor' of AI models and using formal methods to ensure reliability, proving that while AI can be unpredictable, the protocols surrounding it can be engineered for safety.