Research
Programme
Foundational R&D powering REALM's privacy engineering
and cognitive governance capabilities.
From research to product proof
Realm is not a product wrapper. Each research direction terminates in an artifact a customer's reviewers can actually open - this is how technical depth becomes proof.
Privacy Engineering Studio
Rigorous privacy engineering — differential privacy, synthetic data, and attack validation for enterprise data workflows.
Governed Runtime
Memory governance, RAG privacy, and policy verification for regulated AI work with full auditability.
The architecture, in plain English
The moat that separates Realm from governance dashboards is architectural. Here is what it means without the jargon.
CVPL
A formal way to measure real re-identification risk - showing that data can be "compliant" on paper yet still leak. It powers PAMOLA's attack testing.
Read the CVPL paper →MARM
A memory scaffold for AI agents with consistent structure and metadata, so what an agent remembers, reuses, or forgets stays inspectable and policy-bound.
ATE
A deterministic, Bayesian engine that decides whether an action is allowed, with zero LLM calls in the verdict path. The same case and policy produce the same verdict - byte for byte, every time.
Model-is-a-Plugin
Inference is swappable. Bring any model - the governance layer stays the same, so changing models never weakens your controls.
Verdict recomputation
A reviewer independently recomputes a verdict from the same inputs and policy version and gets the same result - not a log to be trusted, but evidence that can be re-derived.
Why governance must be testable
A policy you cannot test is a promise. Realm makes governance executable, so it can be attacked, measured, and proven - not just declared.
Research behind the runtime
Realm's positioning rests on real technical depth, not a product wrapper. The CVPL framework - the basis of PAMOLA's privacy attack testing - is published research by Realm CTO Valery Khvatov and co-author Neyman.
The same foundations carry into AYITA: governed memory (MARM), the deterministic trust engine (ATE), and verdict recomputation are what make Realm's governance executable and its evidence reproducible.
Lead author of the CVPL framework on measuring real re-identification risk (with co-author Neyman).
Read the CVPL paperInterested in our research?
We're building foundational technology for privacy-preserving AI. If you're working on related problems — let's talk.