LIDAR Method for Black-Box LLM Fingerprinting in Coding Agents
September 25, 2026
LIDAR identifies LLMs by analyzing agentic trajectories rather than text distributions, focusing on how models handle tool use and failures. The method fingerprints models through post-edit verification, recovery from transient failures, and specification-test conflict resolution.
HOW THIS AFFECTS YOU
●
builderBe aware that your agent's specific error-recovery patterns may uniquely identify the underlying model.
●
researcherYou can use active black-box fingerprinting to distinguish between models based on decision-making behavior.