TOPLOC

A lightweight integrity proof technique designed for large language model inference. It computes compact commitment values over intermediate layer activations to verify whether an inference provider is running the claimed model, making it more resistant to deception than output comparison methods.