GenmountGenmount OS

ACTION PLAYBOOK

How to Build an AI-Native Medical System

With a computable, auditable unified interpretive framework as the kernel, we build a hospital / wellness center, rehabilitation robots, and biopharma R&D support — three product forms sharing one reasoning spine and one end-to-end data-confidentiality base.

Honest status: the v0.2 paper is public; its §11 three gates (interpretability / reproducibility / falsifiability) have NOT yet passed. Every technical claim here is falsifiable and pending validation, verified gate by gate under the 11-gate governance framework.

THE ENGINE · FOUR PILLARS

Four Technology Pillars

The cosmology is the bedrock; the four pillars are the engine. All three product forms run on them.

Xirang

Hybrid State-Space Topography (HSS)

Represents the world through spatial and causal continuity rather than pure token-sequence stitching — answering the "language hallucination" of large models.

Chonglou

Tri-Tier Architecture of Permanence

Structural / episodic / working memory isolated across three tiers plus a policy layer — answering forgetfulness and loss of long-term consistency.

Zhiju

Bicameral Decision Engine

Dual-track decisioning with an independent safety gate; "no" and "I don't know" cannot be bypassed — the first system that dares to refuse.

Lianshan

Deep-Syntactic Discovery (Multi-hop)

Multi-hop reasoning, cross-chapter causality, auditable increments (evidence chain + version lineage) — a cross-domain closed loop.

Explore the Four Pillars →

REASONING SPINE

A Unified Reasoning Spine

Whether in the clinic, the rehab room, or R&D support, the system runs one structured medical reasoning chain; every conclusion traces back to a root anchor, and what cannot be traced back is rejected.

  1. 01

    Spatiotemporal environment sensing

    Including regional epidemiology priors, dietary geography, regionalized syndrome digitization

  2. 02

    Structured feature encoding

    Encodes observed features into a computable discrete representation

  3. 03

    Meridian-syndrome mapping

    Maps seasonal and channel-guiding agents against a meridian baseline

  4. 04

    Eight-principle synthesis

    Four-axis synthesis that converges user, environment, and task

  5. 05

    Treatment plan

    Three downstream branches: physician collaboration / self-cultivation / R&D support

Every external output is labeled: educational / pattern-recognition reference, not diagnosis / prescription / treatment advice.

PROJECTION DOMAINS

One Kernel, Projected Across More Domains of Life

The same unified interpretive kernel and reasoning spine serve human medicine first, and can project into further domains of life such as veterinary and plant medicine — domain differences live in the application layer, while the shared root anchor commensurates at the kernel. New domains must first pass the admission gates; until then each is marked a candidate, never claimed as supported.

Human medicine (TCM)

Mainline · first release

Currently the only domain systematically built and validated; still falsifiable, with the three §11 gates not yet passed — never claimed as clinically proven.

Veterinary medicine

Candidate · pending 4 admission gates

The same kernel projected onto animal-body syndromes — species differences sit in the application layer, the root-anchor commensuration stays in the kernel. A conceptual draft for now; admitted only after passing the 4 admission gates and validating the syndrome↔anchor mapping.

Plant medicine

Candidate · pending 4 admission gates

The same kernel projected onto plant signs (phenology / growth rhythm / disease), taking phenological input from the environment-sensing layer. A conceptual draft for now; admitted only after passing the 4 admission gates.

The 4 admission gates are part of the 11-gate governance framework; before passing, any domain is labeled a candidate, pending validation.

IMPLEMENTATION EXAMPLES

One Kernel, Six Worked Examples

The same unified interpretive kernel and reasoning spine project into each field as its own worked example. All of the below are architectural examples with an honest-status label — not validated clinical capability (the three §11 gates are not passed).

TCM

Mainline · first release · physician-collaboration

Chief complaint + spatiotemporal context (seasonal baseline) → structured syndrome encoding → six-meridian / eight-principle differentiation → treatment-plan candidate. Fully auditable and traceable back to the root anchor; output is a structured reference for physician collaboration, not a diagnosis or prescription. Still falsifiable — the three §11 gates are not passed.

Western medicine

Comparison layer · not in the decision root-chain

The same presentation projected under the Western paradigm (ICD / lab reference ranges / evidence-based guidelines) is placed alongside the TCM projection on one neutral root anchor for cross-paradigm comparison. Reference only — not a diagnosis, not a replacement for either system.

Veterinary medicine

Candidate · conceptual draft

The same kernel projected onto animal-body syndromes — species differences sit in the application layer, the root-anchor commensuration stays in the kernel. A conceptual draft for now; admitted only after passing the 4 admission gates and validating the syndrome↔anchor mapping.

Plant medicine

Candidate · conceptual draft

The same kernel projected onto plant signs (phenology / growth rhythm / disease), taking phenological input from the environment-sensing layer. A conceptual draft for now; admitted only after passing the 4 admission gates.

Rehabilitation robot

Advisory output · no autonomous actions deployed

Structure the surrounding environment → 5 controlled action classes (avoid / prompt / assist / pause / hand off to a human); raw video / audio never leaves the device, only discrete features remain. Scope is strictly TCM rehab + modern rehab; output is advisory for now — action thresholds require clinical-pilot feedback plus a three-party sign-off before deployment.

Biopharma R&D support

R&D support · two hard stops

A structured medical-knowledge engine + auditable cross-evidence reasoning; anchored at the formula layer (reverse-traceable), with targets / mechanisms validated by statistical evidence; medical knowledge enters the hypothesis set only as phenotype-stratification / combination-mechanism priors. Two hard stops: no clinically usable companion diagnostic, and it does not present itself as regulatory filing material.

Every example output is labeled: educational / pattern-recognition reference, not diagnosis / prescription / treatment advice. Veterinary / plant medicine are candidates, admitted only after the admission gates.

IMPLEMENTATION PATH

Build Steps for Six Institution Types

All six share one build backbone; the differences are the domain and the gates. The path below is governed by the guardrails — candidate domains must pass the admission gates first; nothing is built, claimed validated, or substituted for a physician before the gates pass.

  1. Gate 0

    Admission

    Has this domain passed the 4 admission gates? Candidate domains must pass first, or all later steps are frozen.

  2. 1

    Data base

    Spatiotemporal sensing + local-first + session-scoped shred + reverse-lookup defense.

  3. 2

    Reasoning spine

    Five stages: environment sensing → structured encoding → syndrome mapping → synthesis → plan candidate.

  4. 3

    Capability layer

    CDSS three layers (visualization / dialogue orchestration / compliance & traceability), or the robot's 5 controlled actions.

  5. 4

    Human-in-the-loop

    Physician / expert sign-off gate + trace_id hash-chain audit replay.

  6. 5

    Pilot & launch

    Clinical / field pilot + three-party sign-off → limited launch; never claimed validated before the gates pass.

TCM–Western hospital

Gate open · §11 unpassed · physician-collaboration

The TCM five-stage chain is the decision root-chain; Western ICD / guidelines sit alongside as comparison reference (not in the root-chain). AI gives structured assistance, the physician signs off — not diagnosis / prescription.

Wellness center

Consumer companion · not CDSS

Uses the Stage-5 self-cultivation downstream (daily routine / diet / breathing / meditation / movement); builds daily-practice plans; supernatural claims pass the 3-criteria stripping test; a lifestyle companion that does not touch diagnosis.

Veterinary hospital

Candidate · must pass admission gates first

Gate 0 is a hard prerequisite — must pass the 4 admission gates + validate the syndrome↔anchor mapping; until then stages 1–5 are frozen and nothing is built. After passing, species differences live in the application layer.

Plant-medicine institute

Candidate · must pass admission gates first

Gate 0 likewise prerequisite; positioned as research observation (not a clinical institution), taking phenological input from the environment-sensing layer; frozen until the gates pass.

Rehabilitation robot

Advisory output · no autonomous actions

Stage 3 = 5 controlled actions (avoid / prompt / assist / pause / hand off); raw streams never leave the device; scope is TCM rehab + modern rehab; thresholds land only after clinical feedback + three-party sign-off.

Biopharma plant

R&D support · two hard stops

Stage-5 downstream: a structured knowledge engine + auditable reasoning, anchored at the formula layer (reverse-traceable), targets validated by statistical evidence; two hard stops — no clinically usable companion diagnostic, and not presented as regulatory filing material.

Throughout: never claimed validated before the gates · AI is not a medical device and does not replace physicians · candidate domains are not built before the admission gates · raw data local-first + shredded · neutral external wording.

THREE FORMS

Three Product Forms

One kernel, projected in three directions into three AI-native medical products.

①

AI-Native Hospital / Wellness Center

  • Three layers: visualization / dialogue orchestration / safety-compliance traceability
  • Chinese-first ASR/TTS, warm human presence, refusals with a human touch
  • Attached voice / video runs as a dumb channel; no digital humans or virtual-physician personas

We are AI, not a medical device — the product form is "physician collaboration"; the physician is the perception input, and AI provides structured assistive reference.

②

Rehabilitation Robots

  • Exclusive battlefield: TCM rehab + modern rehab only — no surgery / delivery / housekeeping
  • Environment features structured into 5 gated actions (avoid / prompt / assist / pause / hand off to human)
  • Raw video / audio streams never leave the device; only discrete features are kept

Currently advisory output only, no autonomous actuation; action thresholds are deployed only after clinical-pilot (>=1 rehab cycle) feedback and three-party sign-off.

Robotics →
③

Biopharma R&D Support

  • Structured medical knowledge engine + cross-evidence auditable reasoning
  • Anchored at the formula layer (reversible), targets validated by statistical evidence
  • TCM knowledge enters the hypothesis set only as phenotype-stratification / combination-mechanism priors

Two hard stops: produces no clinically usable companion diagnostic; never claims to be a regulatory filing itself. The system says "I don't know" and stops loss earlier.

END-TO-END · DATA CONFIDENTIALITY

End-to-End Data Confidentiality

All three forms share one end-to-end confidentiality base, layered four deep.

01

Local-first

Rehab-robot raw streams never leave the device; only discrete features remain.

02

Session-key burn-after

Perception data is encrypted with a session-level key and destroyed immediately at session end, with no grace period; never bound long-term to a user identity.

03

Anti-join guard

Engineering-level ban on "location + medical input -> outcome" aggregation queries, enforced by integration tests (offending requests return 400).

04

Network isolation

Services run on an isolated intranet; database and monitoring ports are never exposed publicly; the master key is readable only inside the running container.

Reversible value traces (90 days) flow through a separate retention channel, physically isolated from raw perception data.

GOVERNANCE

Gates Before Deployment

The capability roadmap is surrounded by governance guardrails and ships only after gate-by-gate verification.

  • 11-gate governance: 3 paper gates (interpretability / reproducibility / falsifiability — all currently unpassed) + 4 engineering gates + 4 admission gates, verified gate by gate and externally traceable.
  • Falsifiable commitment: before the three gates pass, we never claim "already verified / already clinically proven / already peer-reviewed".
  • Open-source strategy: platform layer open-sourced (AGPLv3), method and artifact layers tiered; the base model is replaceable.
See the 11-Gate Map →

SYSTEM MAP

The AI-Native Medical OS Map

One picture: from the unified interpretive kernel to three product outlets, surrounded by governance and floored by data confidentiality.

Governance · 11 gates / "AI not a medical device" / falsifiable
Data ConfidentialityPerception LayerReasoning SpineFour PillarsInterpretive KernelThree Product OutletsHospital / WellnessCDSS 3 layersRehab Robots5 gated actionsBiopharma R&DStructured engine

Four Pillars Engine

Xirang · HSSChonglou · Tri-TierZhiju · "No"Lianshan · Multi-hop

Unified Reasoning Spine

01Environment02Encoding03Meridian mapping04Synthesis05Treatment plan

Perception Layer

Physician collaborationRegional priorsRobot environment sensing

End-to-End Data Confidentiality Base

Local-firstSession-key burnAnti-join guardNetwork isolation
We don't chase scale. We chase physical honesty about the real world.

Genmount OS — The Bedrock of World Intelligence.

ACTION PLAYBOOK — Genmount OS