Beyond the Traffic-Light Model: Legal Epistemic Traceability, Human- Accountable Reasoning and the Attribution of AI-Assisted Legal Acts A Contextual Framework for European Union and Spanish Law after the 2026 Digital Omnibus

The rapid normalisation of generative and agentic artificial intelligence in legal practice exposes a structural weakness in prevailing regulatory discourse. Practical guidance commonly asks where AI may or may not be used and answers through traffic-light classifications, professional cautions, or system-level risk categories. This article argues that those approaches are insufficient because legal permissibility is not a stable property of an AI tool. It is a contextual property of a concrete AI-assisted legal act. The article formulates an integrated framework for assessing such acts under European Union and Spanish law after Regulation (EU) 2026/1744, with particular attention to the AI Act, the GDPR, professional secrecy, the right of defence, judicial independence, procedural fairness and evidentiary reliability. It advances three original analytical constructs. First, Legal Epistemic Traceability (LET) is defined as the capacity to reconstruct, independently of the AI system itself, the normative and evidential chain through which an AI-assisted proposition becomes attributable to a legally responsible human actor. Secondly, an Epistemic Adoption Test (EAT) specifies when machine output may legitimately be treated as a professional proposition: the responsible actor must know the role played by the AI, understand the material proposition, independently verify it to a degree proportionate to consequence, retain genuine capacity to reject it, and consciously adopt it as part of the actor's own legal reasoning. Thirdly, the article replaces weak human-in-the-loop formulations with human-accountable reasoning and, for agentic AI, with bounded automation based on authority, reversibility and legal consequence. The framework is applied to five legal scenarios, including public-source drafting, confidential RAG, autonomous legal research and filing, judicial drafting, and AI-mediated evidentiary analysis. The article concludes that the next phase of legal AI governance should move from tool classification to professional epistemic governance, where attribution, contestability and reconstructibility become central rule-of- law safeguards.