Transparent process · Human-verified output

How Daritas works

From your control-system code to an evidence-backed model — graded machine-generated or engineer-reviewed, coverage and unknowns disclosed — a clear, documented process at every step.

📝
01

Submit your system

Upload your control program — a PLCopen XML export or the project source — plus any schematics, manuals, or notes. One controller or a whole multi-unit system: a system assessment composes several units into one model, each at the depth it needs. A guided form walks you through scope and what you most need to get out of it.

🔍
02

Daritas surveys & condenses

We survey the whole codebase, map its structure, and condense it to its logic essence — stripping the tool-noise (graphics, layout, regenerated wrappers) while keeping the behaviour and the exact source locations.

🔧
03

Grounded extraction

Daritas extracts every behaviour it can identify and ties each one to the evidence that supports it — source-derived claims down to their source locations. A claim whose quote does not occur at its cited location is rejected automatically — a provenance check the machine can make; whether the evidence truly supports each claim is what expert review adds. What the code establishes and what is design-level structure — interfaces, intent — are reported as separate tiers, never blended, and coverage is disclosed: what was read, what was out of reach, what stayed unknown. Whatever the sources cannot settle becomes an open question put to you and your team, never a guess.

🎓
04

Expert review & sign-off

A qualified automation engineer verifies the model against the evidence, resolves ambiguities, and signs off within a stated scope. The reviewer is named and accountable — an attestation of specific claims, not a blanket guarantee.

📦
05

Delivery & revisions

You receive an interactive model to explore and a report structured the way your installation is — system, unit, source file, function — with per-unit dependency diagrams and every finding linked to its evidence. Multi-unit systems get an Interface Catalog (ICD) covering the connections inside the system and out. Plus document exports (Markdown, Word, PDF) and a structured export you own. Revisions are included — send your comments and we turn them around fast.

What every assessment includes

📎
Evidence-grounded
Every claim links to the evidence that supports it — source-derived claims to their source locations — and coverage is disclosed: what was analysed, and what was not.
🔒
Confidential
Your code and data are never shared. NDAs on request.
🎓
Expert-signed
In the expert lane, a named automation engineer reviews the model and attests to its claims within a stated scope.
Fast turnaround
An Orientation in days; deeper scopes by agreement.

What you receive, concretely

Four structures every assessment ships — shown from the AHU-01 sample, a synthetic demo air-handling unit run through the real engine. Abbreviated, not mocked: the hashes, counts, and the claim are the sample's actual output.

1 · A claim card

“FZ-01 has a safety-relevant alarm: freeze stat trip — heating coil frost protection.”

state sourceSupported · confidence 0.8 · safety-relevant
grade machine-generated · engineer review pending
evidence exports/alarm_list.csv · line 2
“ALM_AHU01_FREEZE, critical, safety, Freeze stat trip - heating coil frost protection”
source sha256 7f47d158…
2 · The coverage manifest
sources ingested 10 · content-addressed (sha256)
src/AHU01_Project.xml · 995719ae…
exports/alarm_list.csv · 7f47d158…
model 78 nodes · 57 edges · 6 claims · 93 evidence items
confidence mean 0.77 · min 0.40 · max 0.90
open clarifications 0

What was not read is listed, never assumed.

3 · The signed attestation

Ships with reviewed grades. It binds, at minimum:

  • reviewer — name & qualifications
  • scope — claims covered, exclusions, unknowns
  • review depth — exhaustive / risk-based / sampled
  • artifact hashes the review was performed against
  • date & supersession conditions
  • obfuscation status

No attestation is shown for AHU-01 — the sample is machine-grade, and we do not fake signatures.

4 · The change impact

What a re-ingest reports when FB_FreezeProtection.st changes:

claims re-checked 6
invalidated 1 — the freeze-stat claim above
unaffected, retained 5
re-review scope the invalidated claim only

Illustrative flow — the version diff ships with every re-ingest.

A model you can use, not just read

The report is the first thing you get, not the last. The model behind it stays yours, stays current, and answers questions.

Ask it questions, and see where the answer came from

Ask about an alarm, an interlock, a mode you don’t recognise. Answers are drawn from your assessed system and cite the model behind them — so you can check the reasoning instead of trusting a chat window. Scoped to one system: the same alarm code means different things on different machines.

For whoever is standing in front of the machine.

It tells you what changed

Each assessment is kept as a baseline. Re-ingest after a change and you get a diff: which findings still hold, which were invalidated, and the review needed — scoped to what actually moved, not the whole system again.

For the person who has to sign off after a modification.

Your own tools can read it

The verified model is available over an authenticated read API, and to AI copilots over MCP — read-only, scoped to your organisation by your own key, so your assistant reasons over evidence-backed facts instead of guessing. Every assessment also exports in portable form: it is your model, usable in your systems, and it leaves with you.

For your engineering and IT teams — and whatever AI they already use.

A named engineer stands behind it — or the machine says it didn’t

In the expert lane an automation engineer reviews the claims and attests to them within a stated scope, and the attestation records who and when. In the machine lane nothing is signed and the deliverable says so. You always know which you are holding.

For whoever carries the risk of acting on it.

It asks you and your team — and their knowledge stays

Some things aren't in the code: why a setpoint is what it is, which alarm everyone ignores, the workaround that never got written down. We ground what the sources can prove and ask your people the rest — the controls engineer, the operator, the integrator who built it, the vendor's support desk. Each question is written to be forwarded: what is asked, the line that prompted it, and what we currently believe, so whoever receives it can answer or pass it to the person who can. Open questions stay visible in the portal against the behaviour that raised them, and every answer is recorded against the claim it explains, with who said it and when — kept as testimony, a separate tier from what the code proves, so the knowledge outlives whoever happens to know it today.

Bring the people who know the system, including the ones you'd forget to ask.

Three ways to run it

Security claims are only honest per mode — this is the table.

ModeWhat can leave your siteHuman review path
Fully air-gappedNothing — a local model, or the deterministic engine alone with the reduced depth honestly markedYour engineers, or a reviewer working inside your network
Local engine + your AI APIThe configured payload — source excerpts and prompts to the provider you chose, under your provider contract, including its training termsLocal, or controlled remote review
Hosted collaborationThe evidence model and the artifacts you approve for uploadDaritas reviewers or your own, in the portal

Obfuscation reduces what any mode exposes — proprietary names replaced before review, restored on delivery, with your approval of the diff before anything goes. Structure and logic remain your IP in every mode: you decide what leaves.

🔄 A model that compounds, instead of leaking away

The intelligence layer

Most engineering knowledge walks out the door when a person does. Daritas is built the other way — every assessment, and every reviewer correction, makes the next one faster and more accurate.

🤖→👤

The AI proposes, the engineer disposes

When the model is proposed, the reviewer sees every claim beside the exact evidence. If something is off, they see precisely why — file, line, and a suggested correction. The work gets sharper as it is done.

👤→🤖

Reviewers sharpen the model

Every correction a reviewer makes is captured. The next time a similar pattern appears, those corrections inform it. Assessment #50 benefits from everything learned in #1–49.

📖

Every assessment starts with context

Before anyone reviews a behaviour, the system builds an overview — the systems, their structure, the risks. Human and AI start from the same shared context, not a blank page.

🏛️

Knowledge outlasts any individual

Customer-confirmed history accumulates. When a great engineer moves on, their confirmed judgements stay. The platform accumulates expertise; it does not leak it.

Assessment #100 will be faster, cheaper, and more accurate than #1 — not because we hire more people, but because the system retains everything it learns, customer-confirmed.