How an inference provider can prove they're not serving a quantized model