Daritas Docs

Security & Data

Industrial control code is sensitive. Daritas is built so you never have to upload source to a cloud you don't control, and so the knowledge you gain stays yours.

💡
Preparing a formal review? The full Security Overview for IT / OT reviewers covers data handling, deliverable properties, connectivity design, and our known limitations — implemented vs pending, honestly marked.

Your source stays on your floor

The engine runs inside your own network and ingests your source where it lives. Everything you receive works air-gapped forever. You decide what, if anything, leaves the premises.

For the exact statement — what leaves in each of the three ways of running it, including the fact that a synced model carries the verbatim excerpt each claim cites — see Local mode.

Bring your own AI

Where the deep reading uses an AI, it can run under your account rather than ours — but which of these you get depends on where the engine runs, so here is the current state plainly rather than the intended one.

How the reading happensWhose account, and where the key lives
You run it — the CLI on your own machine (local mode)Your own API key, read from your own environment. It is never sent to Daritas and Daritas never stores it — there is nowhere in the product that holds a customer’s provider credentials, by design.
We run it — the assessment you order in this portalOur account and our contract with the model provider. There is currently no way to supply your own key here; if you need the reading to happen under your own contract, run it yourself.
No AI at all — the deterministic engine aloneNothing leaves. Shallower, and the deliverable says so rather than quietly reading less.
ℹ️
A local model is planned, not built. Running the reading against a model on your own hardware (llama.cpp, Ollama) is the design intent and the option people ask for most, and it does not exist yet. Until it does, the private options are your own API key on your own machine, or no AI at all.

Evidence, not opinion

Every statement in a model is bound to a verbatim source quote, and a proposed statement whose quote is not found in the source is rejected before it is stored. The deliverable is grounded and checkable — not an opaque summary you have to take on faith.

You own the model

Your model exports to an open, documented format that is dependency-free and readable without us. The knowledge is a durable asset you keep, not a rental that disappears if you stop subscribing. See The Model & Evidence.

IP protection

When work does involve our reviewers, options exist to reduce how much of your source is exposed — so a sensitive system can be understood without handing over more than necessary. For the most critical systems, you can remove external exposure entirely by verifying with your own engineers (below).

Your reviewers, no external eyes

For critical systems you can staff verification with your own engineers — employees or trusted contractors — on Daritas’s rails, so the model is reviewed with no external eyes at all. See Verification & review.

Isolation in the cloud

When you do use the cloud app, your data is isolated:

  • Each organization’s model is locked to that organization; access is by invitation.
  • The model tables are protected at the database level, reachable only by the server after an authorization check — never directly by other tenants.
  • API keys are stored hashed and scoped to a single organization.
ℹ️
The short version: the sensitive reading happens where your code already lives, the deliverable is evidence you can check, the knowledge is yours to keep, and for critical systems no outside person needs to see the code at all.