Problem, findings and limitations
The worker generates an answer token by token. A verifier receives the fixed input and output tokens and can evaluate that path in a prefill pass. This report compares those two timing paths, asking when replay can cost less than generating the answer again.
The report covers Qwen3-8B on one RTX 4090 and Qwen3-32B on one H100 NVL. Both use vLLM 0.29.0, BF16/bfloat16 and tensor parallel size one. The benchmark sweeps 48 input/output-length combinations; model, hardware and sequence lengths must accompany any quoted result.
Generation includes sequential decoding, whereas the fixed-token replay has a different workload shape. The measured ratio is therefore configuration-dependent. It should not be advertised as a constant verification discount for every model, batch or provider.
Single-request timing does not establish end-to-end service latency or total protocol cost. Queueing, networking, multiple verifiers and evidence handling require separate measurements. The report bundle contains published results; inspect the repository for the benchmark script and available run data before reproducing it.
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.
Citation
@techreport{trueopen_worker_verifier_latency,
title = {Worker vs. Verifier Latency Benchmark Report},
author = {{TrueOpen.ai team}},
institution = {TrueOpen},
url = {https://www.trueopen.ai/research/worker-verifier-latency/},
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.
Single-request timing is not production end-to-end latency. This is not an independent security audit.