L1 platform skeleton — the operating system layer for auditable agentic reasoning.
Coming soon. Not released yet — by design.
This is the L1 platform skeleton of the DOORM Genmount WorldModel-OS architecture: a hybrid local/cloud operating system for agentic world models that puts auditable rollback, state-space discipline, and disclosure boundaries ahead of scale.
Modern agentic LLMs ship a fundamental contradiction: they are tasked with high-stakes domains (medical, legal, financial) but lack the structural primitives to be audited. Hallucination is treated as a tax, not a fault. Rollback is a nice-to-have, not a contract.
WorldModel-OS treats audit as a first-class architectural concern, not retrofit:
- Discrete anchor state-space — every inference produces a coordinate in an 8 / 64 / 384 anchor space, not just free-form text
- Four-level rollback contract embedded in every inference trace
- Redactor + append-only audit trail built into the gateway, not bolted on
- L1 / L2 / L3 disclosure boundary — what's open, what's closed, and why
When released, this repository will contain:
| Module | Purpose |
|---|---|
gateway/ |
Request entry, authentication, rate limiting, redactor |
state_space/ |
8 / 64 / 384 anchor encoding + decoding |
audit/ |
Append-only trace store + four-level rollback hooks |
redactor/ |
PII and clinical-data egress filter |
runtime/ |
L2 adapter handshake + adapter_version pinning |
Base-model agnostic — orchestrates any compliant L2 adapter.
| Gate (§11 of paper) | Status |
|---|---|
| Reproducibility (single-GPU 24h reference run) | ⏳ Pending |
| Benchmark (TCM syndrome differentiation + rehabilitation) | ⏳ Pending |
| Real-adoption (MOU or paid deployment) | ⏳ Pending |
All three must pass before this repository goes public. No code release for demo's sake.
See §11 of the companion paper for full gate definitions and falsification criteria.
| Tier | Status | Scope | This repo? |
|---|---|---|---|
| L3 Production artifacts | 🔒 Closed | Production weights, clinical adapters, partner data, device auth keys | No |
| L2 Finetune Workbench | ⏳ Pending | Training pipeline + three hard quality gates | Sibling repo |
| L1 Platform skeleton | ⏳ Pending | Gateway · State space · Audit · Redactor · Runtime | ← You are here |
L2 produces adapters. L1 consumes them. They communicate via a stable adapter_version handshake — neither is locked to the other's release schedule.
- ⭐ Star this repo to track release
- 👁 Watch → Custom → Releases for the first public drop notification
- 💬 Discussions for early design feedback
- 📬 Issues for post-release bug reports
- 📄 Paper + governance docs — DOI 10.5281/zenodo.20053545
- 🛠 Genmount Finetune-Workbench — L2 training workbench (sibling repo)
For commercial deployment outside AGPL-3.0-or-later terms, partner integrations, or clinical L3 access:
DOORM AI PTE. LTD. (UEN 202441729W) · Singapore