HowLLMDecision Agents Fail in the Wild: AReproducible Failure-Cell Framework

Frontier language models are increasingly used in consequential deployment settings, but how they fail in those settings is poorly understood. The term “failure” is often used loosely; the actual deployment variables that trigger one are rarely isolated; the reasoning trace surrounding a fail ure is rarely inspected as a diagnostic in its own right; and the fixes proposed against failures are rarely costed in the cells where they would ac tually be deployed. We introduce Operational Alignment, a framework that treats deployment failure as a reproducible experimental object. The framework provides four pieces designed to be adopted directly: a precise operational definition of when a model has failed (a violation of a stated operative rule under a controlled decision config uration), a method for constructing the smallest deployment-realistic configuration that triggers the failure the same way every time, trace-level diagnostics for what kind of failure has occurred, and an intervention-transport test that exposes what each candidate fix costs in cells beyond the one it was tuned on. The framework is built so that decision-level failures can be reported, re produced, and composed into the long-horizon failures that arise in agentic deployments. We instantiate it on three regulated decision testbeds and release the corpus, the manipulation library, the diagnostic audit, and a reporting template. The framework is built so failure cells can be composed into the longer-running, multi-agent deployments where these failures actually occur in production.