What question did this study set out to answer?

This research aims to address the determinism-expressiveness gap in AI systems for regulated domains through the introduction of the MINERVA architecture.

synapse

⌘+K

synapse

⌘+K

May 5, 2026Open Access

MINERVA: A Production Architecture for Hybrid Agents with Deterministic Decisions in Regulated Domains

Key Points

This research aims to address the determinism-expressiveness gap in AI systems for regulated domains through the introduction of the MINERVA architecture.
Developed the HADD paradigm with five architectural invariants.
Implemented a polyglot holon structure to support language neutrality.
Validated the framework using real regulated-sector documents.
Proved compliance with EU AI Act, GDPR, and DORA regulations.
Demonstrated the effectiveness of the EMRE application, incorporating reinforcement learning from human feedback with mixture of experts.
Outlined governance consequences of cognitive/reactive holon distinctions.

Abstract

Presents MINERVA, a production architecture for AI agents in regulated domains that resolves the determinism-expressiveness gap in LLM-based systems through the HADD (Hybrid Agents with Deterministic Decisions) paradigm. Six contributions: (1) HADD paradigm codified as five architectural invariants; (2) cognitive/reactive holon distinction and governance consequences; (3) polyglot holon implementation with Language Neutrality property; (4) multi-tenant operational-cell formalism; (5) four reusable design patterns; (6) EMRE — first application of RLHF+MoE to the epistemic boundary of a deterministic BDI system. Validated in production with real regulated-sector documents. Compliant with EU AI Act, GDPR, and DORA by construction.

Read Full Paperexternally

Mark Helpful

Bookmark

Relay

View Full Paper