EviLedger: governing clinical AI with a verifiable evidence ledger

Abstract Clinical evidence evolves, yet most AI systems cannot reconstruct the evidence state supporting a past decision. We introduce EviLedger, an evidence ledger that converts guidelines, drug labels, and EHR events into immutable assertions linked to hashes and rollback. On 2,000 guideline-change events and 1,200 cross-source contradiction cases (Cohen’s κ = 0.87), EviLedger achieves drift F1 94.2% and contradiction F1 94.0%. A blinded semantic audit finds that 93.7% of extracted assertions are semantically supported by their evidence spans (κ = 0.84), and external validation on ESC/JCS guidelines yields F1 >90%. As an auditable memory layer for LLM-based retrieval, EviLedger reduces stale citations from 14.7% to 1.2% and unverifiable citations from 28.4% to 2.3%, while supporting p95 rollback in 5.13 s at 78M assertions. In a 6-month hospital pilot, EviLedger detects 5.5× more actionable guideline changes with 104× faster triage than manual surveillance.