Announcing 'Decision Stack', a New Architecture Redefining Decision-Making in the AI Era

SHIRO & Co., Inc. (Headquarters: Tokyo, Representative: Kosuke Shiroko) announces the 'Decision Stack' decision-making architecture as an answer to the structural challenges of decision-making in the AI era.

This architecture redefines decision-making not as an AI directly outputting an 'answer,' but as a 'controlled process.' It places 'accountability, audit, and human final judgment'—which are insufficient in discussions of speed and accuracy—as 'design prerequisites' rather than afterthoughts. Generative AI is positioned as a lower-level inference/generation engine, with a stack of decision-making layers above it that separate meaning, interpretation, trust control, and execution.

Positioned as an Upper Layer of AI The Decision Stack is not an 'extension' of a single AI model or prompt design, but rather 'an upper architecture that bundles multiple AIs, rules, and human judgments.' Generation and inference are treated as lower-level engines, and prior to their output, the layers above sequentially control meaning candidate selection, context-dependent interpretation selection, 'HOLD' based on risk, and execution into business operations. The Decision Stack is the layer that determines 'when, with which interpretation, to execute/stop,' before 'what the AI says.'

Innovation and Paradigm The innovation of this concept lies not in 'fine-tuning generation' such as improving accuracy or prompts. By separating meaning, interpretation, trust control, and execution, and by treating 'not executing' as a normal outcome rather than a failure—going this far replaces AI from an 'answer device' to a 'decision-making foundation that can stop, branch, and explain.' This is not a functional addition, but a 'paradigm inversion.'

Relationship with Patents, Academia, and Precedent Cases The Decision Stack is presented not as an 'improvement' on existing generative AI governance, explainability (XAI), RAG, or human-in-the-loop collaboration, but as an 'independent design proposition centered on layered responsibilities and HOLD.' The contrasts and differences with pending applications, peer-reviewed papers, and industry precedents will be clarified sequentially as disclosures, publications, and implementation details are released.

What's Different (3 Points) ・Structure: Inference is not completed by a single unit. Responsibilities and records are separated for meaning, interpretation, trust control, and execution.

・Output: Value is not placed solely on 'action.' Hold, stop, and resume are treated as designed outcomes.

・Operation: Explanations are not deferred to post-hoc. Branches, thresholds, and policies are embedded beforehand.

Comparison with Conventional Models (Input–Output Model vs. Decision Stack)

Conventional (Input → Output) Decision Stack Core Concept AI as an 'answer machine' that directly outputs results. AI as a 'decision-making process controller' that manages execution. Focus Speed and accuracy of output. Controllability, explainability, and reliability of the decision process. Handling of Uncertainty/Risk Attempting to minimize or ignore it to achieve a single answer. Explicitly managing and controlling it through 'HOLD' and branching. Role of Human Judgment Often an afterthought or for error correction. Integrated as a core component for interpretation, trust control, and final decision. Output Value Primarily the final output (e.g., text, code). The entire decision process, including the decision to hold, stop, or proceed. Explainability Often post-hoc analysis or model-specific. Built-in through layered responsibilities and pre-defined policies.