Guardrails that prevent the AI agent from fabricating requirements, inventing APIs, or hallucinating behavior not present in the specification or codebase.
