TrueOpen
TRUEOPEN / Experimental report

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

TrueOpen.ai team · Publication date not stated in the source

What can probability traces tell us about the model behind an inference response?

Problem, findings and limitations

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.

Record generation → Replay fixed tokens → Compare metrics → Apply calibrated limits1Record generation→2Replay fixed tokens→3Compare metrics→4Apply calibrated limitsStatistical consistency evidence, not a cryptographic proof of complete execution.
TrueOpen explanatory diagram. This illustrates the method; it is not an additional measured result.

Published evidence

Download report and results bundle

The ZIP contains report files and published tables, not a complete raw-log dataset. See the repository for available run artifacts and reproduction instructions.

Additional experimental context

Citation

@techreport{trueopen_model_identity,
  title = {Worker-Verifier Model Identity Verification with Logprobs and Routed Experts: An Experimental Study},
  author = {{TrueOpen.ai team}},
  institution = {TrueOpen},
  url = {https://www.trueopen.ai/research/model-identity/},
  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.

Statistical consistency evidence, not a cryptographic proof of complete execution. This is not an independent security audit.