A provider name or a plausible answer does not establish model identity. Record the model revision, tokenizer, precision, runtime, prompt, generated tokens and available logprobs before comparing a reference replay.
LMCV studies token logprobs and top-k distributions from a prefill replay. Calibrate normal variation across hardware before setting rejection thresholds. For MoE models, routing evidence may be needed; logprobs alone can be insufficient.
Sampled layerwise proofs check selected computation chunks after commitments. They do not prove all unsampled computation. Neither approach establishes factual correctness, privacy or settlement security. These are research methods, not evidence that every live request is verified.