Is That API Really Serving the Model It Claims? A Reproducible Fingerprint Test

Short summary

A reproducible fingerprinting method tests whether LLM API endpoints actually serve the model they claim, using one-token preference distributions compared via Jensen-Shannon divergence. The open-source llm-fingerprint-detector CLI lets anyone audit OpenAI-compatible endpoints and publish verifiable match/mismatch/uncertain results. Mismatches don't prove fraud—updates, quantization, and load-balancing can shift fingerprints—but the protocol provides honest, repeatable evidence.

• •One-token repeated sampling creates a behavioral fingerprint for LLMs
• •Open-source TypeScript CLI audits OpenAI-compatible endpoints against trusted references
• •Public report schema ensures reproducibility with JSD metrics, timestamps, and limitations

Generated with AI, which can make mistakes.