Skip to content

THOX.ai Research · Technical report v1.1.1 · 2026-11-10

Preprint, not peer reviewed

THOX Community Mesh: Unlinkable Subscriber Tokens, Digest-Audited Stage Execution and a Witness-Cosigned Settlement Ledger for Member-Contributed Inference

How MeshStack™ subscribers can pool their own hardware, and reach cloud models through it, without the pool learning who is asking: a design, a working virtual lab and the measurements behind both.

Craig Ross

Chief Executive Officer

ORCID 0009-0005-9872-3673

Tommy Xaypanya

CTO and Co-Founder (corresponding author)

ORCID 0009-0000-4425-9565

THOX.ai LLC

DOI 10.5281/zenodo.24120004 · Concept DOI 10.5281/zenodo.24115595

Abstract

Most people who want to run a capable language model can afford neither the hardware nor a metered cloud account, while their neighbours’ computers sit idle. This paper specifies and tests a community mesh for THOX MeshStack™: members add their own devices to a shared pool, and any subscriber can run distributed inference on that pool, or reach a hosted cloud model through it, without the pool learning who is asking. The design combines blind-signed subscriber tokens (RFC 9474) and sealed envelopes through a relay (RFC 9180) for requester unlinkability; signed per-hop receipts with sampled replay audits for integrity; contribution credits that raise a member’s token budget and tier, with no currency; and a witness-cosigned Merkle ledger that settles credits and penalties after the request, off the token path. We state the limit plainly: a contributor that executes a stage can read its activations, so the pool gives identity privacy, not content privacy, and requests that need the latter are refused rather than routed. On a real 135M-parameter model split into three stages, a token-exact re-execution audit, the natural choice, exposed 69 of 102 cheated requests and missed an fp16 weight substitution on all 17 prompts; a SHA-256 digest of the stage boundary exposed 102 of 102. Minting an unlinkable token costs 7.9 ms at RSA-2048 and verifying one 0.12 ms; the ledger costs 0.92 microseconds per receipt. A virtual lab of 15 processes with real cryptography ran 12 scenarios, including a contributor killed mid-request, an owner reclaiming a device, a cheating contributor caught and penalised, and a forged ledger head refused by every witness; all passed and every lab request was token-identical to the whole model. A simulation shows the cost of unlinkability for fairness: fixed issuance budgets isolate heavy users but waste capacity, and indexing issuance to pool load recovers the waste. This edition is consolidated with a companion THOX paper and lab on the same subject: we adopt its plan-wide admission rule, execution contexts and replica-first default, reproduce its 62 tests and 150-trial campaign on a second host, and measure the replica-first claim in our own lab (6.6% lower per-remote-call round trip for one contributor holding both remote ranges than for two contributors in sequence on loopback, 30 matched pairs; corrected in v1.1.1, see Section 6.7). Nothing here is a THOX device, wide-area or throughput result.

In one line

MeshStack subscribers pool their own idle hardware into one private community mesh: the pool never learns who is asking, stage work is digest-audited, and contribution settles in credits after the request, off the token path.

Erratum v1.1.1

This edition corrects five points in v1.1.0 found on independent review: the route comparison now uses matched prompts (118.0 ms against 126.4 ms per remote call on a quiet host, Section 6.7); the witness proposition states its prefix condition (Section 1); replica placement keeps the first layers on the requester’s own device in both routes (Section 3); fairness waits are reported with never-started requests counted (Section 6.4); and the lab control services are stated not to be a security boundary (Section 7).

118.0 ms

One contributor holding both remote ranges, per remote call, against 126.4 ms for two contributors in sequence; matched prompts on a quiet host, 30 pairs, all token-exact (erratum v1.1.1)

Virtual lab, measured

102/102

Cheated requests exposed by the SHA-256 stage-boundary digest audit on a real SmolLM2-135M model split into three stages

Virtual lab, measured

62 of 62

Companion Community Compute lab tests passed on a second host; 150 of 150 trials numerically equal, 164 receipts verified

Independent reproduction

1.91M

Median served tokens at 0.6 load with load-indexed token budgets, against 1.38M with fixed budgets, across 30 seeds

Simulated

Download the paper (PDF)Zenodo record Read the paper (HTML) Community Mesh Virtual Lab

Contents

  1. Executive summary
  2. 1. Introduction
  3. 2. Background and related work
  4. 3. System and threat model
  5. 4. Design
  6. 5. Implementation
  7. 6. Evaluation
  8. 7. Discussion and limitations
  9. 8. Conclusion
  10. 9. Roadmap
  11. Appendix A. Reproducibility
  12. Appendix B. Notation
  13. Appendix C. Authorship and disclosure
  14. References

Where the paper lives

  • Zenodo record (DOI)

    Version of record: the v1.1.1 erratum edition, published 2026-11-10, CC BY 4.0. Concept DOI 10.5281/zenodo.24115595; corrects v1.1.0 (DOI 10.5281/zenodo.24118129), which superseded v1.0.0 (DOI 10.5281/zenodo.24115596)

  • GitHub paper repository

    The thox-community-meshstack-paper repository: the v1.1.1 release carries the paper PDF, the HTML edition and the Virtual Lab demo; the full source pack is on main under v1.1.1/

  • Read the paper (HTML)

    The v1.1.1 erratum edition as a single-file HTML page, hosted on thox.ai

  • Community Mesh Virtual Lab

    The paper’s interactive virtual lab: plan-wide admission, pipeline and replica placement, grant lifecycle, privacy drills and the settlement ledger, running client-side in the browser

  • Hugging Face whitepapers dataset

    The v1.1.1 erratum edition (PDF, HTML and demo) in the community-meshstack folder of Thox-ai/thox-whitepapers

  • THOX Model Factory catalog

    The research_publications entry for this paper in the thox-model-factory catalog (catalog revision 2026-11-10.5)

Cite this paper

@misc{ross_craig_2026_24120004,
  author       = {Ross, Craig and
                  Xaypanya, Tommy},
  title        = {{THOX Community Mesh: Unlinkable Subscriber Tokens,
                   Digest-Audited Stage Execution and a Witness-Cosigned
                   Settlement Ledger for Member-Contributed Inference}},
  month        = nov,
  year         = 2026,
  publisher    = {Zenodo},
  version      = {v1.1.1},
  doi          = {10.5281/zenodo.24120004},
  url          = {https://doi.org/10.5281/zenodo.24120004}
}

Licensed CC BY 4.0, matching the Zenodo record.

Your AI. Your Data. Your Rules.™