From Explanation to Evidence: A Method-Agnostic Pipeline for Regulatory-Grade XAI Artefacts Under the EU AI Act

High-risk AI systems deployed in the European Union must satisfy transparency and human-oversight requirements under the AI Act (Regulation 2024/1689) by 2 August 2026, yet compliance-artefact generation today remains predominantly manual and dominated by expert time. This white paper proposes a four-step, implementation-conditional pipeline that transforms the output of any attribution method whose deployed implementation satisfies the Shapley Efficiency axiom in the chosen output space into a machine-verifiable, multi-regime regulatory evidence bundle in sub-second wall time. The pipeline operates through (1A) attribution via SHAP as primary method and LIME as baseline comparator, (1B) algorithmic property verification including an Efficiency residual ε with an empirically justified interpretation threshold, (2) compilation into a signed artefact via Computable Governance Notation (CGN), and (3) automated mapping to specific articles of the EU AI Act, ISO/IEC 42001:2023, and ANMAT Disposición 64/2025. The paper is a formalisation contribution rather than an empirical discovery: it turns the informal intuition that SHAP's conservation property should matter for audit into a mathematically precise statement with explicit preconditions. The closed-form identity of Proposition 1 characterises the coverage-score gap ΔCOV = COV(B_SHAP) − COV(B_LIME) as a structural consequence of XAI method selection under a given regulatory mapping. The pipeline is evaluated on four canonical datasets (COMPAS two-year recidivism, Adult Income, German Credit Statlog, and Cleveland Heart Disease) and on a unified probability-space cross-model study over 45 configurations (three tree-ensemble families × three datasets × five random seeds). SHAP satisfies the Efficiency precondition in 44/45 configurations; LIME in 0/45. Under the published regulatory mapping and when the preconditions hold, the identity yields ΔCOV_Art.13 = +1/3. The scope of this work is deliberately narrow: the paper investigates one computable property of attribution artefacts (algebraic self-verifiability in the Shapley-Efficiency sense) and its structural relationship to a regulatory mapping. It does not claim that this property suffices for substantive compliance, that SHAP is universally superior to LIME, or that the pipeline is a production-ready compliance engine. The coverage scores reported are research metrics and not legal attestations. All four datasets used are publicly available through their original sources (ProPublica and the UCI Machine Learning Repository); the full regulatory mapping database is published inside the paper itself (Appendix A).