DOORM Genmount WorldModel-OS

A Governance-First Architecture for Auditable Agentic Reasoning Hybrid State-Space · Structurally-Primed Discrete Anchors · Anti-Scale Training Discipline

This repository hosts the defensive preprint, governance documents, threat model, and disclosure boundary for DOORM Genmount WorldModel-OS — a hybrid local/cloud operating system for auditable agentic world models, paired with the model-agnostic Finetune Workbench.

---

TL;DR

• What this is. A position + system paper proposing seven cognitive increments for auditable agentic reasoning, grounded in a discrete state-space (8 base / 64 paired / 384 ordinal anchors), a four-level rollback contract embedded in every inference trace, and an anti-scale training discipline. Applications: Traditional Chinese Medicine (TCM) syndrome differentiation and rehabilitation robotics.
• What this isn't. An empirical benchmark paper. No metrics are claimed in v0.2. The reproducibility envelope, three release gates, non-coverage list, and falsification criteria are specified in §11.
• Open vs closed. L1 (platform skeleton) + L2 (training methodology) are AGPL-3.0-or-later / CC BY 4.0. L3 (production weights, clinical adapters, partner data, commercial terms, device auth keys) is held under separate commercial / clinical governance.

---

Repository contents

File	Description
PAPER-v0.2-en.md / .pdf	English paper
PAPER-v0.2-zh.md / .pdf	Chinese paper (中文版)
LICENSE	CC BY 4.0
CITATION.cff	Citation metadata
SECURITY.md	Security contact

---

Companion code repositories

The platform and workbench code (separately governed):

• Genmount WorldModel-OS — L1 platform skeleton (AGPL-3.0-or-later)
• Genmount Finetune-Workbench — L2 training workbench (AGPL-3.0-or-later); base-model agnostic

---

How to cite

Use the CITATION.cff entry (GitHub auto-generates citation buttons), or:

@misc{doorm2026worldmodel,
  title  = {DOORM Genmount WorldModel-OS: A Governance-First Architecture for Auditable Agentic Reasoning},
  author = {Tong, Li and Tong, Tong and Wu, Lihua},
  year   = {2026},
  doi    = {10.5281/zenodo.20053545},
  url    = {https://github.com/doorm-ai/Genmount-WorldModel-OS-Governance}
}

---

Authors

Author	Role	Affiliation
Tong Li¹*	Conceptualization; Methodology; Software (system architecture); Writing	DOORM AI PTE. LTD.
Tong Tong¹	Software (implementation, deployment); Data curation; Validation; Project administration; Funding	DOORM AI PTE. LTD.
Wu Lihua²	Conceptualization; Investigation (literature review); Writing – review	Independent Researcher

¹ DOORM AI PTE. LTD. (UEN 202441729W), Singapore ² Independent Researcher * Corresponding author: service@doorm.ai

Author order reflects relative contribution, not academic seniority.

---

License

• Paper + governance documents: CC BY 4.0
• Linked code repositories: AGPL-3.0-or-later (see those repos individually)

---

Disclosure boundary

Tier	Status	Scope
L1 Platform skeleton	✅ Open (AGPL-3.0)	Service architecture, gateway, audit, state space, redactor
L2 Training methodology	✅ Open (AGPL-3.0 + CC BY 4.0)	5-class data pipeline, 3 quality gates, bucketed evaluator
L3 Production artifacts	❌ Closed	Production LoRA weights, clinical adapters, partner medical data, commercial license terms, device authentication keys

L3 is held under separate commercial / clinical governance by DOORM AI PTE. LTD.

---

Contact

• Corresponding author: service@doorm.ai
• Issues: GitHub Issues
• Security disclosure: see SECURITY.md

---

Three release gates (per §11 of the paper)

No artifact is considered officially released by DOORM AI until all three gates are passed:

1. Reproducibility gate — A third party can run run_dataset_v1.py → train → eval on a single consumer GPU (RTX 3090 / 4090, 24 GB VRAM) within 24 hours, producing a v1 adapter matching the reference adapter_version.
2. Benchmark gate — At least one bucketed metric in the domain evaluation (TCM + rehabilitation robotics) with a pre-registered rollback threshold.
3. Real-adoption gate — At least one MOU or paid deployment with a medical institution. (Partner identities not disclosed in this draft.)

Until all three are passed, every claim in this paper is, by our own framing, falsifiable conjecture — not validated science.