THOX.ai Research · Paper v1.1.0 · 2026-11-04
Preprint, not peer reviewedThoxWork: One Workplace, One Connected Team. A Policy-Bounded DigitalHuman Work Platform for Enterprise Use
THOX.ai LLC · Cedar Park, Texas · November 4, 2026
Abstract
Enterprise adoption of AI agents is outrunning enterprise governance of them. Vendor surveys report organizations experimenting with agents at scale while only a minority operate policies to secure them, workplace benchmarks show the best agents completing a minority of consequential tasks unaided, and orchestration research shows that at team scale the deciding factors are not model capability but discovery scope, handoff fidelity, and the structure of human review. THOX’s answer is a work platform, not another agent framework. In the ThoxWork model, work lives in shared workspaces; DigitalHumans are named, roleholding colleagues who can be assigned tasks and held to account for them; Rolodex is the team interface through which a person assigns a task, checks in, and reviews results; a MeshGroup binds a named team to the exact set of resources it is approved to use; and MeshStack connects those approved resources. This edition reports evidence in three labelled tiers. Simulated: the model’s five invariants (unapproved resources never run; oversized work is denied rather than spilled; reassignment preserves a complete handoff chain; no task reaches Done without eligible human review; queues complete predictably as teams and groups scale) are evaluated in a deterministic, seeded virtual lab, where every pre-registered bar is met. Measured, code and lab: an evidence lab re-implements the product’s routing planner and the reference admission kit and agrees with both on every seeded case (12,000 of 12,000 previews; 20,000 of 20,000 kit decisions); a divergence study re-run under the paper’s own admission semantics shows pooled admission admitting 3,882 of 20,000 probed requests that no single node could serve, and quantifies how a looser port of the same rule over-admits; and a re-implemented detection matrix shows what a keyed, externally anchored audit chain detects that hash links alone cannot. Measured, live implementation: the same evaluation generators that produced the simulated results were replayed against thoxwork-core, the live Rust implementation of the model, and every bar was met again on the running system, including the exact 120 admitted / 121 UNPLACEABLE capacity boundary and a three-judge agreement check over 102 workspaces. The edition also reconciles the record it found: three admission definitions and five assignment state machines are shown side by side, the one material disagreement is dispositioned, and the platform’s upstream lineage (OpenWork, OpenCode, Hermes Agent) is stated with its license posture. ThoxWork as a purchasable enterprise platform remains in development; THOX’s own product copy says so, and this paper claims nothing further.
0 / 12,000
Unapproved-resource attempts leaked across direct, cross-group and escalation attacks (E1)
Simulated
1.000
Handoff-chain completeness over 5,000 reassigned tasks and 15,000 chain entries; 0 of 4,000 negative probes leaked (E2)
Simulated
100%
Of 4,000 oversized tasks denied UNPLACEABLE; 0 silently placed, 0 spilled to another group (E3)
Simulated
0 / 6,000
Review-gate bypasses succeeded across six attack classes; 4,000 of 4,000 legitimate tasks completed through review (E4)
Simulated
Contents
- 1. Introduction: the third platform
- 2. Background, related work and positioning
- 3. System and threat model
- 4. The ThoxWork model, reconciled
- 5. What is built: an evidence-bounded account
- 6. Method: three instruments
- 7. Evaluation
- 8. Governance mapping
- 9. Reading the results
- 10. Product implications
- 11. Security considerations
- 12. Limitations
- 13. Decisions and roadmap
- 14. Provenance, consent and document control
- Appendix A. Reproducibility
- Appendix B. Notation
- Appendix C. Errata register
- Appendix D. Disclosure boundary
- References
Where the paper lives
- Read the paper (HTML)
The v1.1.0 expanded edition (HTML), hosted on this site
- Download the paper (PDF)
The v1.1.0 expanded edition (PDF), 45 pages, hosted on this site
- Browser demo
Interactive demo implementing the paper’s rules: the Rolodex board, MeshGroup admission control and the review gate
- GitHub paper repository
Public paper repository: https://github.com/ttracx/thox-research-paper/tree/main/papers/thoxwork/v1.1.0
- DOI record
Zenodo DOI 10.5281/zenodo.23965699
Cite this paper
@misc{ross_craig_2026_thoxwork,
author = {Ross, Craig and
Xaypanya, Tommy},
title = {{ThoxWork: One Workplace, One Connected Team.
A Policy-Bounded DigitalHuman Work Platform
for Enterprise Use}},
month = nov,
year = {2026},
publisher = {THOX.ai Research},
version = {v1.1.0},
url = {https://thox.ai/research/papers/thoxwork}, doi = {10.5281/zenodo.23965699},
note = {Preprint, not peer reviewed.
}
}Preprint, not peer reviewed;
Your AI. Your Data. Your Rules.™