Model signing is meant to authenticate the weights of a released large language model, but the version users actually run is usually a quantized artifact. A new arXiv paper identifies a verifiability gap: the signature covers the full-precision weights W, not the quantized weights Q that get deployed.

According to the paper, the transformation from W to Q is not uniquely determined under current disclosure practices. The same signed full-precision model could plausibly map to multiple quantized artifacts, so a user cannot cryptographically confirm that the Q they run is faithful to the original W.

That gap matters because quantization is a standard step for deploying LLMs on memory-constrained hardware. The finding implies that model signing alone does not extend to the artifacts users actually run, leaving a blind spot in the LLM supply chain.