TrueOpen

Original papers and experiment artifacts

TRUEOPEN / KNOWLEDGE

Can logprobs reveal model substitution?

Controlled Qwen experiments compare BF16, FP8, AWQ and smaller-model substitutes.

The bundled Worker-Verifier study compares a Qwen3-32B BF16 worker against reference configurations using vLLM 0.23.0, batch size 8 and 64 logprobs. The worker produced 36,882 rows. The table below is the 6000ws verifier subset, not an estimate for arbitrary commercial APIs.

Within this controlled dense-model experiment, the reported p99 deviation rises for FP8 and AWQ relative to BF16. Thresholds must account for legitimate runtime and hardware differences. A mismatch is evidence to investigate, not automatic proof of deliberate fraud.

For Qwen3.6-35B-A3B, the report finds sparse, position-sensitive logprob spikes. Layer-0 expert-routing comparisons add information. If a provider exposes neither suitable token traces nor routing, this method may not support a conclusion.

Qwen3-32B worker, 6000ws verifier: reported deviation statistics
Verifiermeanp99rank_delta_rate
Qwen3-32B BF160.01020.11670.0081
Qwen3-32B FP80.02680.29970.0205
Qwen3-32B AWQ0.06710.79620.0478

Original-source bibliography and citations

This page is an editorial overview, not the original paper. Cite the specific report or paper below for its findings. Bibliographic sources checked November 5, 2026.

Worker-Verifier Model Identity Verification with Logprobs and Routed Experts: An Experimental Study

TrueOpen.ai team

Publication date and report version not specified in the supplied report. ZIP export: 2026-09-17; export is not publication.

@misc{trueopen_worker_verifier_report,
  author = {{TrueOpen.ai team}},
  title = {Worker-Verifier Model Identity Verification with Logprobs and Routed Experts: An Experimental Study},
  howpublished = {\url{https://trueopen.ai/reports/TrueOpen-Worker-Verifier-Report.pdf}},
  note = {Publication date and version not specified; source: TrueOpen.ai team, https://www.trueopen.ai/research.html}
}

The related LMCV protocol manuscript credits TrueOpen Research / Decentralized Inference & Verification Group. It is a distinct work; that attribution must not be substituted for the author line of the bundled Worker-Verifier report. Exact report publication date and version await confirmation from its authors. LMCV manuscript ↗

Download scope: the ZIP files contain published report tables and report files, not the complete raw run logs. Read the included README before attempting reproduction.

Sources and artifacts

Cite this research overview

@misc{trueopen_model_identity_en,
  author = {{TrueOpen}},
  title = {Can logprobs reveal model substitution?},
  howpublished = {\url{https://www.trueopen.ai/en/research/model-identity/}},
  note = {Research overview; original experiment sources linked on page}
}

Metric definitions: mean and p99 summarize the absolute difference between worker and verifier selected-token logprobs. rank_delta_rate is the fraction of selected tokens whose rank changes; it is not a fraud-detection accuracy rate.

Related research

Evidence boundary: published experiments and protocol designs do not establish that every production request is verified. No claim of answer correctness, end-to-end privacy or settlement security is made by these experiments.