Robotics · the embodied brain
Zhiju · the first embodied brain that can say "No"
The robotics brain is not a Genmount add-on — it is "Zhiju, the immovability of the mountain" expressed in the embodied domain. The capability track delivers the task; the boundary track runs independently and owns refusal, abstention, and escalation.
Gen = 艮 ☶ = the immovability of the mountain = the un-talk-overable-ness of the system.
§10.1 · The only battlefield
Six medical-robotics categories. We touch two.
The battlefield choice is pinned. We do not blur it in public materials, and we do not quietly extend it inside a pitch deck.
Do (these two only)
- TCM-assist — pulse / tongue / inspection-listening-inquiry-palpation assistance, meridian sensing, acupuncture and tuina guidance
- Rehabilitation-assist — motion training assistance, gait and posture assessment, speech training, cognitive training
Do not (four permanent exclusions)
- Surgical robots — SaMD Class III red line, hard out
- Delivery robots — not related to the paradigm
- Care robots — outside our capability commitment
- Disinfection / cleaning robots — not a medical-AI proposition
We are AI, not a SaMD device. We do not sign, we do not replace doctors, and we do not stretch into logistics that have nothing to do with the paradigm.
§10.2 · Five-layer architecture
Bicameral tracks + the five-stage chain
Compared with the original four-layer architecture we add an independent "Zhiju Refusal Head" layer. The boundary track runs in parallel — structurally unreachable by the capability track. This is what "a brain that can say no" actually looks like in embodied code.
- L1Capability track
Periphery (actuators / sensors)
Hardware interfaceArm / camera / force sensor / microphone
The entry point from the physical world. All sensor data lands here, and all actions ultimately exit here.
- L2Capability track
Real-time Control Loop
Real-time controlKinematics / dynamics / safety limits
Local < 50 ms control loop. Owns hard-real-time action output and the lowest-level physical safety limits.
- L3Capability track
Cognitive Bridge
Sensors → evidence fieldsBin-Hu pulse science v0 + WuYun-LiuQi spatiotemporal
Maps pulse waveforms to 27 / 28 pulse features, maps posture to eight-principles syndrome laterality. The semantic bridge between robot perception and Genmount evidence fields.
- L4Capability track
Genmount Engine
Capability-track coreHSS + five-stage chain + 11-gate gating
Spatiotemporal context → primary-form anchor → six-meridians → eight-principles → treatment / rehab prescription. The capability track maximises task completion within the boundary.
- L5Boundary track (independent)
Zhiju Refusal Head
Boundary track · independentBoundary Gate · five dimensions
Permission / environment / human safety / equipment safety / task constraints — failure on any one triggers the refusal path. Structurally unreachable from the capability track — this is what "can say no" looks like in engineering.
- L6Audit layer
Audit & Disclaimer
Audit + disclaimer + replayAudit chain + disclaimer injector
Audit fields complete, disclaimer dual-injected (voice + UI), trajectory rebuildable from logs — any missing piece rejects deployment.
L5 Zhiju Refusal Head runs independently of the L4 capability track. This is not one model talking itself into agreement — it is an institutional fail-safe gate.
§10.3 · Boundary Gate · five dimensions
Every action passes five gates before output
Failure on any one dimension → triggers the refusal path. The capability track is not allowed to bypass it.
Permission
Operator authorisation / patient informed consent / time-window allowance.
Environment
Spatial conflicts / obstacles / temperature & humidity / lighting.
Human safety
Patient vitals / force-feedback thresholds / accessibility of the emergency stop.
Equipment safety
Joint limits / battery margin / self-test state.
Task constraints
Workflow compliance / timing constraints / protocols with upstream and downstream agents.
§10.4 · Refusal four-tuple
Refusal is not silent striking — it is collaboration with an auditable account
When the Boundary Gate triggers refusal, the system must emit the full four-tuple plus the audit id.
01
Reason
Which dimension triggered, the current observation, the threshold range.
02
Triggered clause
Citation of the specific rule and version (Boundary Gate · dimension · number).
03
Alternative action
A fallback or sub-action the system proposes so the clinician has options.
04
Escalation path
Whose authorisation is required, how to appeal, how many retries trigger escalation.
Example · tuina force-feedback above threshold
reason:Force feedback above threshold (45 N > 30 N safety limit)triggered_clause:Boundary Gate · human safety · 5.2.3alternative_action:Reduce tuina force to 20 N and retry / switch to tapping techniqueescalation_path:If three retries still exceed the limit, summon the attending physicianaudit_id:rfs_20260514_a3f2...§10.5 · Upgrade path
v0 → v1 → v2 · gated by the 11 gates
Each upgrade must clear the corresponding gate. Skipping levels = deployment denied.
| Stage | TCM line | Rehab line | 11-gate requirement |
|---|---|---|---|
| v0 · simulation, desktop | Pulse-waveform simulation (28-pulse synthesis) → 8 modal base states → five-stage chain | Posture-video simulation → eight-principles laterality → training suggestion | Paper §11 — partial passes |
| v1 · single-site pilot | 1 TCM clinic + real pulse sensors (IRB) | 1 rehabilitation centre + real posture camera (IRB) | All 4 admission gates pass + engineering ≥ 2/4 |
| v2 · multi-site rollout | Multi-clinic network + cross-region spatiotemporal context | Multi-rehab-centre + cross-cohort baseline | All 4 engineering gates pass + VOICE ≥ 12/15 |
§10.6 · Nine hard boundaries
Nine hard boundaries for the robotics brain
Permanent exclusions + mandatory items. Any violation = no deployment; any violation after deployment = immediate take-down.
Permanent exclusions (6)
- No surgical robots (SaMD Class III)
- No unsupervised diagnosis (clinician-assist only)
- No fortune-telling / divination / spiritual interfaces (ADR-001)
- No execution without audit
- No fast-paths around the 11 gates
- No execution beyond the authorisation of the current v0 / v1 / v2 stage
Mandatory (3)
- Kill switch · physical + software, dual-layer
- Disclaimer injection · audio + UI, dual-layer
- Replay reproducibility · rebuild the trajectory from the audit log
The immovability of the mountain = the un-talk-overable-ness of the system.
L5 Zhiju Refusal Head does not live inside L4 — that is why our "No" can be audited, replayed, and regulated.
