Root Cause Analysis

Root cause analysis (RCA) asks “why did this happen?” rather than stopping at “what happened?” When a model produces a harmful or incorrect output, the root cause might be missing training examples, an ambiguous instruction in the system prompt, a flaw in the behavioral specification, or an evaluation gap that failed to catch the issue before deployment. Treating only the symptom — adding a single example or a narrow fix — often leads to whack-a-mole problem solving. For behavior architects, RCA is a discipline of investigation: tracing a failure back through the system to its origin, so that fixes address the actual problem and the same category of failure doesn’t recur in a slightly different form.