Human Oversight Evidence in AI-assisted Decisions: The Procedural Doubt Layer as an Examinable Process Framework

Formal human presence in an AI-assisted workflow does not establish that human judgment was substantively exercised. Conventional operational logs often record that an action occurred without capturing the structure and substance of human engagement preceding it. This evidentiary gap creates a systematic problem in accountability analysis: when AI-assisted decisions cause harm, the review process frequently collapses to outcome, formal authority, and nominal human involvement, rather than a granular examination of how human judgment actually entered the process. This paper proposes the Procedural Doubt Layer-a methodological framework that shifts accountability analysis from formal presence to examinable process. Before resolving who is responsible, the Procedural Doubt Layer asks which specific element of the human decision process was present, absent, reliable, or unattributable. The framework is operationalized through an eight-link decomposition of the AI-assisted decision chain: authority, consequence record, AI reliance boundary, uncertainty posture, return duty, supervisor review, execution trace, and integrity of record. We examine the regulatory context in which this framework becomes relevant, with reference to EU AI Act Articles 14 and 26 and a 2026 Israeli Supreme Court ruling on AI-assisted municipal decision-making. We propose four design properties for Human Oversight Evidence records suited for legal examination, and illustrate the adversarial symmetry of the Procedural Doubt Layer-its availability to both claimants and decision-makers-as a structural feature rather than a limitation. The paper does not propose a defense against liability. It proposes a method for making liability claims more precise.