💭 if you asked an llm to generate 100 random numbers, itd obviously not be random and you could pattern analysis that to get a signature of what llm you're talking to. Other sorts of generations you could do the same, both 'list random' and any other patterns.

Could quantify somewhat how similar two llms are. With enough thoroughness and dimensionality could maybe be a robsut fingerprint Could quantify how much other context such as system prompt affects the llm Could be used to identify llms with masked identities such as previews