Problem, findings and limitations
The September 17, 2026 report tests a seal-then-sample pipeline on GPT-2, TinyLlama-1.1B and Llama-2-70B. Commitments are fixed before the challenge; selected chunks are proved and checked. This page summarizes that report version, not every later paper revision.
For Llama-2-70B on a CPU server with 2 TB memory, the report records 163 sealed boundaries, five proved chunks, 1,259 seconds (about 21 minutes) for sealing plus proving, a 4.34 MiB proof and 46.3 seconds of verification. This excludes model registration and the separately reported 48,706-second integer-inference stage; it is not end-to-end latency or a production promise.
The report proves selected chunks of a 12-bit quantized canonical model, not the original floating-point serving model. Its 70B and TinyLlama proving experiments use context lengths of 8–16; production-context proving costs were not measured. Most reported figures are single runs without repeated-run variance (except E09 and E12). These limits matter when comparing the prototype with an inference service.
For a forgery confined to one eligible chunk fixed before the challenge, detection by a single audit is limited by sampling coverage. The report gives about 1.9% single-audit detection probability (3 randomly sampled chunks out of 161 eligible chunks, alongside 2 mandatory anchors) for one fixed invalid chunk in the 70B configuration and demonstrates challenge grinding with manifest-only sampling. External ordering and explicit threat assumptions are needed.
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
Read the later September 23 arXiv preprint separately
Citation
@techreport{trueopen_slp_poc,
title = {SLP Verifiable Inference: Proof-of-Concept Experiment Report},
author = {{TrueOpen.ai team}},
institution = {TrueOpen},
year = {2026},
url = {https://www.trueopen.ai/research/slp-poc/},
note = {Version dated 2026-09-17}
}Sampled checks do not prove unsampled computation. This is not an independent security audit.