Governance & AI
Daritas is the evidence-and-governance layer for AI in production: a bounded context AI can safely rely on, a ledger of how it was used, and a scan of where AI already touches your OT.
AI in and around the installation
AI copilots are increasingly used to explain, query, and even act on control systems. The risk is not the AI itself — it is AI asserting things that were never verified, or acting on the installation without a record. Daritas addresses this with three things: a bounded context AI can safely rely on, a ledger of how AI was used, and a scan of where AI already touches your OT.
AI-ready control context
The governed context is a machine-readable view of a system’s verified model, built for an AI copilot to consume. It deliberately separates what can be relied on from what cannot:
| Part | What it carries |
|---|---|
| Verified behaviours | The signed-off facts an AI may rely on. |
| Known unknowns | Flagged / unverified items, marked "do not assert". |
| Usage policy | The read-only / advisory / no-control boundary the model is served under. |
| Coverage & sources | How much is verified, and the approved sources behind it. |
The AI runtime ledger
The ledger records AI events in and around your systems — what was asked, what was answered, which model, and the grounded context it was given — with an influence level from advisory (L0/L1) up. The Advisor logs to it automatically; your own AI systems can post events to it via the API. It is the audit trail for “how has AI been used here?”.
The governance evidence pack
The governance pack aggregates, per system: verification coverage (the baseline of what is actually confirmed), the AI-use inventory from the ledger with its influence levels and any approval gaps, and the resulting open risks. It downloads as a white-labellable report — evidence you can put in front of an auditor or a safety committee.
AI-in-OT exposure scan
The exposure scan runs on the floor over your source and configuration and flags where AI already touches the control system — LLM calls, APIs, agent frameworks, SDKs — with an estimated influence level and the gaps that matter:
- a control-write path (AI that can act, not just read);
- no logging of what the AI did;
- no human approval step;
- sensitive data in reach.
Findings — not your source — are shipped to the app for review, and a control-write finding becomes an open risk in the governance pack. It answers a question most operations cannot: where is AI already wired into our OT, and what can it do?