A Deterministic Architecture for Pre-Execution Governance

Le résumé fourni par la source

AI systems increasingly operate in environments where governance cannot be retrofitted after execution. Policy documents, constitutional alignment, post-hoc audits, runtime guardrails and reactive shutdown mechanisms share a structural defect: they act too late, on the wrong layer, or with signals that are themselves probabilistic. This paper describes PREEXEC®, a deterministic pre-execution governance architecture that evaluates every instruction before any AI action is taken. The instruction is scored on three structural signals, equally weighted by default — syntactic structure, semantic determinacy and affective stability — yielding a ClarityScore that drives a review decision; permissibility is enforced separately by a red/amber policy layer with an engine-fixed legality floor. The resulting verdict is ternary (EXECUTE / HOLD / BLOCK): unclear instructions are held for structured clarification, impermissible instructions are blocked regardless of clarity. Every decision is committed to a tamper-evident audit chain with signed Merkle-root commitments. Version 3.0 supersedes the two-axis (Clarity / Awareness) architecture described in version 2.0: the Awareness axis was empirically falsified in production-scale evaluation and removed, policy conformance was restructured from a score component into a separate enforcement gate, and all three clarity signals were rebuilt against sealed specifications after their prior operationalisations were falsified. The paper documents the architecture of the signed release build as validated — including the negative results that motivated the revision — and updates the regulatory framework to Regulation (EU) 2026/1744 (Digital Omnibus on AI). An empirical companion study, measuring the shipped clarity screen on 60 agentic instructions with all raw data published, appears as DOI 10.5281/zenodo.21939298. PREEXEC® is a registered EU trade mark (No. 019301368) of Clarity IP Holdings Limited. Note: this publication does not disclose proprietary algorithms or the operationalisation of the described signals.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.