What question did this study set out to answer?

The aim is to present OCHA, a neuro-symbolic architecture that ensures distinct functional roles in decision governance for high-criticality contexts.

March 21, 2026Open Access

OCHA: Objective-Constrained Hybrid Architecture — A Neuro-Symbolic Architecture for Auditable Decision Governance in High-Criticality Domains

Key Points

The aim is to present OCHA, a neuro-symbolic architecture that ensures distinct functional roles in decision governance for high-criticality contexts.
Introduced a two-layer architecture enforcing functional separation: Layer 1 for probabilistic perception and Layer 2 for deterministic decision-making.
Defined four architectural principles for governance and decision-making.
Added a governance function for regulatory alignment under evolving regulations.
Provided a proof-of-concept for the architecture using a specific financial scenario.
Established principles ensuring that perception does not equal authority, thus enhancing decision integrity.
Demonstrated a structured pathway for decisions that comply with stringent regulations.
Created an auditable outcome system where legitimate ignorance serves as a controlled decision state.

Abstract

This paper introduces OCHA (Objective-Constrained Hybrid Architecture), a neuro-symbolic architecture enforcing strict functional separation between probabilistic semantic perception (Layer 1 / LLM) and deterministic decision authority (Layer 2 / SMT solver) in high-criticality domains such as financial audit, legal reasoning, and regulatory compliance. OCHA is governed by four architectural principles:(1) Intelligence ≠ Authority — the capacity to perceive does not confer the right to decide;(2) Non-Optimization Under Authority Constraints — the system seeks the first feasible decision satisfying all formal constraints, not the globally optimal unconstrained outcome;(3) The Decision Singularity Point (DSP) — the precise instant at which probabilistic authority is irrevocably transferred to the deterministic subsystem, making OCHA justiciable;(4) Legitimate Ignorance — the inability to produce a formal decision given current evidence is a controlled, auditable outcome that preserves decision authority integrity. Version 5.3 (Definitive Final Edition) adds: Principle 4 (Legitimate Ignorance), a complete contributions table, and the Regulatory Monitoring Obligation (Regulatory Intelligence Function / RIF) — a governance role ensuring constitutional specifications remain aligned with evolving regulation (EU AI Act, MiCA, AMLA). The architecture includes: the Neuro-Symbolic Threshold (NST) with cryptographic API contract enforcement; the Integration Membrane routing Exploratory Kernel insights; the Sniper Protocol with game-theoretic grounding (Nash Equilibrium, sycophancy mitigations SM1–SM4); the Cognitive Audit Pipeline (5 stages); the Conditional Correctness Guarantee (CCG); and the Judicial Evidence Property (eIDAS Art.25 + FRE 901(b)(9)). A runnable Z3 proof-of-concept on the Wirecard scenario is provided in Appendix A.

Read Full Paperexternally

AI에게 질문

Bookmark

View Full Paper