NMAI Core, NHS Clinical CUI, UK Model Handoff

This page summarises the technical implementation route for the NHS Sovereign AI Replacement Architecture.

The purpose is not to build another chatbot. The purpose is to prove a controlled route:

The architecture is designed so the NHS can pilot, test, audit, stop, replace, and scale the system without surrendering the workflow to a model-first dependency chain.

The technical stack separates the workflow into distinct layers:

The implementation route uses a lightweight local or hosted API layer. Core endpoints:

NMAI Core creates a decision state before model handoff:

The system is designed to stop safely. A failure must not fall back to an external model.

The prototype and pilot stack uses open-source or inspectable components:

The preferred pilot path is CUI-first. LibreOffice integration can operate as an office-layer replacement route, but the NHS Clinical CUI is the cleaner first pilot path because it avoids being blocked by locked desktop estates.

nhs-sovereign-ai/
|-- nmai/
|   |-- engine.py          (NMAI Core deterministic conditioning)
|   |-- markers.py         (Marker detection module)
|   |-- rules.py           (Ruleset loader and validator)
|   |-- audit.py           (Audit trail writer)
|   +-- config.py          (Configuration loader)
|-- rules/
|   +-- nhs-rules.json     (NHS marker rules)
|-- api/
|   +-- main.py            (FastAPI service)
|-- audit/
|   +-- logs/              (JSONL append-only audit logs)
|-- model-gateway/
|   +-- handoff.py         (UK model endpoint handoff)
|-- vllm-server/
|   +-- Dockerfile
|-- nginx/
|   +-- nginx.conf         (Reverse proxy configuration)
|-- libreoffice/
|   +-- plugin/            (LibreOffice extension files)
|-- docker-compose.yml
|-- requirements.txt
|-- Dockerfile
+-- README.md

The rules file defines the deterministic boundary for all NHS AI workflow:

The NHS Clinical CUI is a controlled workflow interface — not a gimmick chatbot.

Its functions:

• Input capture: Receives NHS workflow input in structured form
• Task selection: Presents approved task categories (rewrite, summarise, restructure, format, draft)
• Marker display: Shows detected markers to the user
• Warning display: Surfaces caution markers with explanation
• Blocked state: Clearly prevents blocked tasks from proceeding
• Human review state: Holds output for clinical review before release
• Escalation route: Routes specialist tasks to appropriate teams
• Output state: Displays conditioned output with audit reference
• Audit reference display: Shows unique audit ID for every transaction

The CUI preserves context throughout the workflow. It does not allow open-ended chat that could drift off-domain.

The LibreOffice plugin provides office-layer integration:

• Document check before export: Validates content against NHS rules
• Document check before AI assistance: Routes content through NMAI Core conditioning
• Python-UNO route: Uses LibreOffice's native Python bridge
• Local API call: Communicates with NMAI Core service on local network
• Blocked-state display: Clear warnings when content is blocked
• Warning-state display: Surfaces caution markers with required actions
• Output handling: Returns conditioned output back into the document
• Local-only configuration: No external network required

Plugin workflow: User selects text or whole document → clicks plugin action → document text sent to NMAI → NMAI checks input → returns blocked, warning, review, or approved state → output returned to LibreOffice with audit reference.

vLLM serves as the post-conditioning execution module — not the primary system.

Permitted model tasks:

Rewriting approved text, summarising permitted administrative content, restructuring meeting notes, formatting internal policy text, converting structured notes into draft documents, producing plain-English wording from approved source material.

Blocked/restricted tasks:

Direct clinical diagnosis, patient-specific medical advice, medication recommendations, unsupervised treatment planning, any task bypassing NMAI conditioning.

The full NHS stack deploys as a Docker Compose multi-container system:

Key commands:

# Build and start all containers
docker-compose up --build -d

# Check running services
docker-compose ps

# View logs
docker-compose logs -f nmai

# Test NMAI conditioning endpoint
curl -X POST http://localhost:8000/condition \\
  -H "Content-Type: application/json" \\
  -d '{"text": "Summarise this meeting agenda"}'

# Stop stack
docker-compose down

The model is not selected to replace ChatGPT. The model is selected only to execute bounded permitted tasks.

Model selection checklist: Licence review → NHS use approval → No external API calls → Version lock → Capability boundary → Safety limits → NMAI Core compatibility → Hardware envelope check → Replacement and exit route confirmed.

The pilot measures the full route, not model speed alone.