AI Architecture¶
How AI assists across the platform — always advisory, never authoritative.
This chapter describes the designed AI architecture so you can see where the platform is going. The single principle that governs all of it is firm and non-negotiable, and it is in force today.
The Principle: AI Is Advisory¶
AI may explain, summarise, diagnose, and draft. AI may never approve, publish, dispatch, or alter deterministic truth.
Your risk numbers are produced by deterministic, regulator-grade mathematics. AI never touches them. It helps your people understand and operate the platform — it does not make the platform's decisions. In force today.
One Advisory Plane, Three Roles¶
Rather than three disconnected AI tools, the platform is designed around a single governed AI Advisory Plane serving three roles:
flowchart TD
A[AI Advisory Plane] --> B[Config Doctor assistant]
A --> C[Run health analyst]
A --> D[Result analyst]
| Role | What it helps with | Status |
|---|---|---|
| Config Doctor assistant | Understanding readiness and repair before a run | Available today (on the deterministic Config Doctor engine) |
| Run health analyst | Interpreting run health and reliability | Available today as a health summary; deeper analysis on the roadmap |
| Result analyst | Explaining results and movement after a run | On the roadmap |
Designed Shared Components¶
The target plane is designed to share common, governed components so that behaviour is consistent and controllable across roles:
| Component | Purpose | Status |
|---|---|---|
| AI policy service | Central guardrails for what AI may do | On the roadmap |
| Context assembly | Builds tenant-scoped, governed context for each request | Available today in a focused form; broader assembly on the roadmap |
| Prompt registry | Managed, versioned prompts | On the roadmap |
| Model router | Selecting the right model per task | On the roadmap |
| Citation assembler | Tying every statement to its governed source | Available today in a focused form |
| Advisory artifact store | Retaining advisory outputs as evidence | Available today in a focused form |
| Evaluation framework | Measuring advisory quality | Available today for explainability checks; general framework on the roadmap |
Guardrails That Stay On¶
Whatever the role, the rules hold:
- deterministic risk facts remain authoritative
- AI reads governed facts and evidence references, not arbitrary files
- AI may summarise, explain, compare, and draft
- AI must not silently alter controlled configuration or result truth
- any action still flows through typed, authorised APIs and human approval
What This Means For You¶
- Your numbers are never at the mercy of a model. AI is a help, not an authority.
- Assistance is grounded, because it operates on governed facts and cites its sources.
- The plane grows coherently — new advisory capability arrives on the same governed foundation, not as a bolt-on.
Where The Platform Is Today¶
The advisory principle and the Config Doctor assistant are in force and available today, with a run-health summary available now. The unified plane's shared services and the full result-analyst experience are on the roadmap — see Capability Maturity And Status and the next chapter, Data, Explainability, And AI.