Haruna, a Reasoning Governance Scaffold for Context-, Time-, and Harm-Aware Artificial Intelligence

Current large language models (LLMs) increasingly produce outputs that are coherent, fluent, and technically correct. However, a growing class of failures does not arise from incorrect reasoning, but from correct reasoning applied within an insufficient, implicit, or ethically unexamined frame. These failures often lead to human, social, or long-term harm without triggering conventional safety mechanisms. This paper introduces Haruna, a lightweight, language-based reasoning governance scaffold designed to constrain how an AI system reasons rather than what it knows. Haruna explicitly structures reasoning along dimensions of context sufficiency, assumptions, trade-offs, time, irreversibility, and human impact. It does not modify model architecture, training data, or internal representations, but operates as a transparent overlay on the reasoning procedure itself. Haruna is proposed as a pre-incident innovation: a framework intended to address classes of AI failure that are otherwise only recognized after harm has occurred. Its relevance increases as AI systems become more capable, autonomous, and socially embedded.