Problem, findings and limitations
An answer that sounds plausible cannot tell a customer which model produced it. LMCV proposes comparing probability fingerprints from the requested reference model against the worker’s recorded output path. The aim is economical model-consistency evidence for open inference networks.
Once the input and generated tokens are fixed, a verifier can run teacher-forced prefill rather than generate another answer. The manuscript considers logprob differences, token ranks and top-k candidate distributions. Legitimate hardware, precision and runtime differences must be calibrated before a rejection threshold can be interpreted.
The protocol design adds deliver-first verifier selection, commit–reveal and a two-of-three majority rule. These are proposed coordination and anti-copying mechanisms, not experimental evidence that every production request follows the full protocol. Independence and the challenge schedule remain part of the threat model.
For MoE models, sparse logprob spikes can obscure the decision boundary, so the manuscript also considers layer-0 routed-expert fingerprints. Missing token or routing evidence can make a case inconclusive. The manuscript, model-identity experiments and latency benchmark are separate sources; their reported configurations should not be mixed into an unconditional security or cost claim.
Citation
@techreport{trueopen_lmcv,
title = {Logprob Majority-Consensus Verification (LMCV): A Low-Cost Execution-Verification Protocol for Open Inference Networks},
author = {{TrueOpen Research}},
institution = {TrueOpen},
url = {https://www.trueopen.ai/research/lmcv/},
note = {Publication date not stated in the source}
}The original source does not state a publication date. No date has been inferred from a download filename or repository timestamp. Bibliographic metadata remains incomplete until the authors confirm it.
Protocol design and controlled experiments are not proof of production deployment. This is not an independent security audit.