TrueOpen

Original papers and experiment artifacts

TRUEOPEN / KNOWLEDGE

Community research: inference auditing, proofs and serving

Eight external papers, original-source bibliography and short editorial notes on their relevance and limits.

These are works by external researchers, not TrueOpen publications or endorsements. The short notes below are independently written reading aids, not translations of the full papers. Follow the original links for the authors’ complete claims and licensing terms.

Bibliography checked: November 5, 2026. Dates below identify the linked version; first submission dates are shown separately where applicable. No external PDFs or figures are redistributed here.

Are You Getting What You Pay For? Auditing Model Substitution in LLM APIs

Will Cai; Tianneng Shi; Xuandong Zhao; Dawn Song
arXiv:2504.04715v2 · 2025-09-29 (v1: 2025-04-07)

The authors test model-substitution audits and highlight how inference variability can undermine software-only checks. Their TEE approach is a useful counterpoint when evaluating logprob evidence.

IMMACULATE: A Practical LLM Auditing Framework via Verifiable Computation

Yanpei Guo; Wenjie Qu; Linyu Wu; Shengfang Zhai; Lionel Z. Wang; Ming Xu; Yue Liu; Binhang Yuan; Dawn Song; Jiaheng Zhang
arXiv:2602.22700v1 · 2026-02-26

Selective cryptographic audits target substitution, quantization abuse and incorrect token billing. Compare the audited fraction and threat assumptions before comparing costs with a replay-based verifier.

TOPLOC: A Locality Sensitive Hashing Scheme for Trustless Verifiable Inference

Jack Min Ong; Matthew Di Ferrante; Aaron Pazdera; Ryan Garner; Sami Jaghouar; Manveer Basra; Max Ryabinin; Johannes Hagemann
arXiv:2501.16007v2 · 2025-05-30 (v1: 2025-01-27)

TOPLOC fingerprints intermediate activations with compact hashes. Its experiments motivate comparing activation evidence with output logprobs; reported detection results remain specific to the tested conditions.

Gaussian Variant of Freivalds' Algorithm for Efficient and Reliable Matrix Product Verification

Michael V. Mascagni; Hao Ji; Yaohang Li
NIST publication record · 2017-05-30; revision number not supplied

Gaussian projections check matrix products without recomputing the full multiplication. The numerical-error analysis is background for matrix verification, not a security proof for a complete inference service.

Slalom: Fast, Verifiable and Private Execution of Neural Networks in Trusted Hardware

Florian Tramèr; Dan Boneh
arXiv:1806.03287v2 · 2019-02-27 (v1: 2018-06-08); ICLR 2019

Slalom delegates linear computation from trusted hardware to an untrusted accelerator and checks the result. Any comparison must preserve its TEE trust assumption.

zkLLM: Zero Knowledge Proofs for Large Language Models

Haochen Sun; Jason Li; Hongyang Zhang
arXiv:2404.16109v1 · 2024-04-24; accepted to ACM CCS 2024

Lookup and attention arguments address expensive non-arithmetic steps in LLM proofs. This is related proof-system research, not a measurement of TrueOpen. The source forbids redistribution of this author version; we link to it without hosting its PDF or figures.

Efficient Memory Management for Large Language Model Serving with PagedAttention

Woosuk Kwon; Zhuohan Li; Siyuan Zhuang; Ying Sheng; Lianmin Zheng; Cody Hao Yu; Joseph E. Gonzalez; Hao Zhang; Ion Stoica
arXiv:2309.06180v1 · 2023-09-12; SOSP 2023

PagedAttention reduces KV-cache waste to improve serving efficiency. It informs inference-runtime design; runtime batching is a different problem from protocol-level batch verification.

Narwhal and Tusk: A DAG-based Mempool and Efficient BFT Consensus

George Danezis; Eleftherios Kokoris Kogias; Alberto Sonnino; Alexander Spiegelman
arXiv:2105.11827v4 · 2022-03-16 (v1: 2021-05-25)

Separating data dissemination from transaction ordering offers a useful architectural comparison for evidence distribution. Inclusion here does not mean TrueOpen adopted this consensus or achieved its throughput.

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.