Technology
More than a model. An audit system.
arcOS is not a chat frontend on top of a language model, but a multi-layer audit system in which every layer has its own job. From reading documents cleanly to delivering an evidenced verdict, the layers work hand in hand.
The architecture
Five layers, one verdict.
Each layer solves exactly one problem. Together they produce an audit that is defensible — not just plausible-sounding.
Reading layer
Every document properly indexed
arcOS reads every format: PDFs, scans, emails, ZIP folders, office files. Even scanned legacy contracts, handwriting and mixed data rooms are indexed cleanly and page-accurately — the foundation every other layer builds on.
- OCR for scans and legacy contracts
- Page-accurate indexing of every document
- Mixed data rooms without pre-sorting
Knowledge layer
Curated German risk taxonomy
Land register, building encumbrances, zoning plans, heritage and milieu protection: a curated knowledge base feeds into every audit via RAG. It grows with every audited deal through curated rules — explicitly no training of models on customer documents.
- German risk taxonomy as the foundation
- Domain knowledge flows into every single audit
- Grows through curated rules, not through training
Application layer
Specialized audit agents
The audit is broken into specialized steps — extraction, findings radar, detail analysis, verdict — and each step is its own audit agent with its own mandate. Deterministic rules take over where rules are more reliable than a model, and human sign-offs sit at the decisive points.
- One audit agent per step, with a clear mandate
- Deterministic rules where they are more reliable
- Human sign-offs at the decisive points
Model routing
The strongest model per task
Each audit task runs on the strongest AI model for the job — across multiple providers, supplied with the knowledge layer's domain expertise and steerable per task in the app.
- Multiple AI providers, routed per task
- Supplied with the knowledge layer's expertise
- Steerable per audit task in the app
Evidence layer
Every statement evidenced
Every statement is bound to document and page; citations open the original with its highlight. The audit journal logs every step, and sign-off keeps decisions audit-proof.
- Citation binding down to document and page
- Audit journal logs every step
- Sign-off keeps decisions audit-proof
The loop
Every audit makes the next one better.
Insights from real audits flow back into the system as curated rules — not as training data.
An audit runs
The data room is read, audited, and every statement is backed by a citation.
Insights are curated
Which gap was critical, which wording carries risk — what matters gets curated.
Rules are sharpened
Audit rules and risk taxonomy are deliberately extended and refined.
The next audit benefits
The next deal is audited against the accumulated knowledge from the very start.
The difference
Why not just a chatbot with your PDFs?
The difference is not the model — it is the system around it.
Evidence instead of summary
A generic model summarizes. An audit system proves: citation binding instead of assertion.
Asking the right questions
A chatbot answers questions. arcOS asks the right ones: curated question catalogs and a radar for what is missing.
Growing without customer data
A model alone stays static. The curated knowledge layer grows with every deal — without customer data ever flowing into training.