AI · 8 min read · April 26, 2026
Meta-predicates enforce evidence rules in clinical AI before deployment
A framework using domain-specific languages and epistemological type systems validates that clinical decision logic uses appropriate evidence sources, not just accurate predictions.
Source: arxiv/cs.AI · Michael Bouzinier, Sergey Trifonov, Michael Chumack, Eugenia Lvova, Dmitry Etin · open original ↗ ↗
Meta-predicates constrain what evidence types clinical AI rules may use, enabling pre-deployment validation of epistemological soundness.
- — Meta-predicates assert constraints on evidence types before rules execute, catching errors early.
- — Epistemological type system classifies evidence by purpose, domain, scale, and acquisition method.
- — Decision trees reformulated as unate cascades produce per-variant audit trails showing which rule fired.
- — Approach complements post-hoc explainability (LIME, SHAP) with preventive validation.
- — Demonstrated on 5.6M genetic variants; generalizes to any auditable decision domain.
- — Satisfies regulatory requirements for auditability under EU AI Act and FDA guidance.
- — Works with both human-written and AI-generated rules in AnFiSA platform.
Frequently asked
- A meta-predicate is a rule about rules. It asserts constraints on what types of evidence a clinical decision rule is allowed to use. For example, a meta-predicate might say 'diagnostic rules must use validated biomarkers, not anecdotal reports.' Meta-predicates catch epistemological errors before a system goes live, complementing post-hoc explanation tools.