From Causal Attribution to Legal Addressability: An Evidence-Constrained Framework for Human-AI Governance

This preprint presents an evidence-constrained framework for analysing legal and normative attribution in human–AI systems. The framework introduces a decision-relative System Attribution Profile, SAP = (C, I, A, P, D, H, R), covering causal contribution, interaction structure, operational autonomy, epistemic provenance, delegated authority, effective human control, and risk/harm/reversibility. Evidential confidence is treated separately as a meta-level annotation rather than a system property. The paper develops two inferential constraints: a revised Non-Exceedance Principle, which limits the factual strength of attribution claims to the evidence supporting their premises, and Inferential Containment, which requires cross-domain normative conclusions to identify an explicit legal or normative bridge premise. Its core architecture is Q → [B_Q ↔ E_Q(SAP_Q + Δ_Q)] → N_Q → F, in which a legal or normative question, candidate bridge premises, and question-relevant evidence form an iterative interpretive-evidential process rather than a one-way derivation from technical properties to legal consequences. Legal addressability is distinguished from procedural addressability, which is confined to the final feasibility stage. The framework’s use is illustrated through Moffatt v. Air Canada, Robodebt, SCHUFA, State v. Loomis, SyRI, Dun & Bradstreet Austria, and a bounded high-delegation hypothetical. These are illustrative, single-coder applications; the paper does not claim empirical validation. The contribution is methodological: an audit discipline for making causal, institutional, epistemic, interpretive, and procedural warrants explicit in AI governance, without treating automation as normatively irrelevant or as presumptively person-like.