About

Built to make
AI trustworthy.

Universal Kernel is deterministic AI compliance infrastructure — built from first principles for the era of regulated artificial intelligence. We did not retrofit compliance onto an existing system. We built compliance as the system.

Core principles

01
Determinism before probability
Every claim is verified deterministically before any probabilistic AI system is consulted. Classical approaches are exhausted first. LLMs are called last — and only when the deterministic engine declares a genuine knowledge frontier.
02
INDETERMINATE is a valid answer
Most AI systems fabricate answers when uncertain. Universal Kernel does not. When 24 classical variations are exhausted without resolution, the motor returns INDETERMINATE — a documented, auditable declaration of a knowledge boundary.
03
Human oversight is inviolable
No result enters the Knowledge Base without documented human approval. This is architecturally enforced — not a configuration option, not a policy guideline. The Human Gate cannot be bypassed regardless of how the system is used.
04
Every decision is auditable
SHA-256 hashes chain every event from session creation to human approval. The audit trail is immutable and export-ready for regulatory submission under EU AI Act Arts. 12-15. Nothing is logged selectively — everything is logged always.

Timeline

2024
Research phase
Deep investigation of the failure modes of probabilistic AI systems in regulated domains — oncology, legal, finance. The core insight: compliance cannot be added to an AI system. It must be the architecture.
2025
Constitutional Compiler v7.0
First-order logic engine with formal verification support. The DIPEngine exhaustion cycle is defined: 6 classical approaches, 4 variations each, 24 total before declaring a knowledge frontier. Validated against benchmark domains.
2026 Q1
Production deployment
Universal Kernel v1.0 goes live. PostgreSQL stateless architecture, Redis async queue, EU AI Act Arts. 12-15 compliance confirmed. VAAR paper submitted to NeurIPS 2026.
2026 Q2
API access opens
Developers can now build AI Act compliant products on Universal Kernel infrastructure. One instance per client. Motor stays on our servers. Code never leaves.

EU AI Act compliance

Art. 12
Record keeping
SHA-256 hash chain from session creation to human decision. Every event logged immutably. Export-ready for regulatory submission.
Art. 13
Transparency
VAAR protocol classifies every claim. No result passes as VERIFIED without traceable source. INDETERMINATE is documented, not hidden.
Art. 14
Human oversight
Human Gate is architecturally inviolable. No KB update without documented human approval. Delegation protocol configurable per domain.
Art. 15
Accuracy and robustness
INDETERMINATE is a first-class output. Classical exhaustion before probabilistic inference. No fabrication by design.
Thiago de Sousa Barbosa Monteiro
Founder — Sovereign Systems, Chamusca, Portugal
Researcher and architect of the Genesis Engine v3.0 and Universal Kernel v1.0. Author of the VAAR paper (NeurIPS 2026 submission). Work focused on the intersection of deterministic verification, hybrid AI systems, and regulatory compliance under EU AI Act 2024/1689. ORCID: 0009-0009-5896-1502.