A failure mode where an LLM agent ignores retrieved authoritative tool output and instead answers from its own generated outputs.