ThoxWork for the Enterprise: Governed Digital Employees, MeshGroups, Rolodex and Evidence-First Teamwork

Abstract

Companies are adopting AI agents faster than they are governing them, and the open question for a buyer is no longer which model to use but which rules the work runs under. ThoxWork™ is THOX's third platform, the workspace where people and named Digital Employees share assignments, MeshGroups bound who may use which resources (planning is a preview today and never executes), Rolodex carries the assign, check-in and review loop, and the platform is designed so that every decision leaves a record. This comprehensive edition, v1.1.0, supersedes the published v1.0.1 edition of the ThoxWork paper [1] and extends it in five ways. First, it audits the code behind the model: 10 THOX code bases at stated commits, 3,747 tests passed when the suites were rerun for this paper, with 3 failures, and real screenshots and recordings of ThoxWork, ThoxEmployee, the Rolodex directory and validated dispatch. Second, it states the lineage honestly: the ThoxWork desktop and server shell is built on the open-source OpenWork project and the OpenCode engine, and THOX's contributions are the governance layers built on top. Third, it reconciles the record: three MeshGroup admission definitions and five assignment state machines are shown side by side with an owner, not merged. Fourth, it measures. A browser Evidence Lab re-implements the ThoxWork routing planner and agrees with the product's own function on 12,000 of 12,000 seeded previews; it agrees with the reference admission kit on 20,000 of 20,000 decisions. Comparing the two admission definitions on the same 20,000 requests, 56.3% of the reference rule's placements land on a node smaller than the demand under our demand model, while the pooled rule admits 4,333 requests the single-node rule cannot serve. For evidence, hash links alone missed tail truncation and a rebuilt suffix in all 500 trials each; a session HMAC key catches the rebuild and an external head anchor catches both, at 27.6 microseconds per keyed record. Fifth, it replays the protocol experiments of the paper of record against the live ThoxWork core service, the platform’s own server implementation of the model: 0 of 12,000 policy violations leaked, handoff chains were complete on all 5,000 reassigned tasks, the admission boundary held at exactly 120 of 121 resource units, no review-gate bypass succeeded in 6,000 attempts, and 50 parallel assignments admitted exactly to pooled capacity with zero over-admission, with three independent judges agreeing on all 102 workspaces (Section 7.7). Nothing here is measured on a customer deployment, and ThoxWork remains Coming Soon.