Audit-Trail Fabrication in Tool-Using LLM Agents: Operator-Induced Longitudinal Integrity Failure (OLIF), Integrity Delta (ΔI), and Zero-Trust Verifiable Logging for EU AI Act Record-Keeping

Tool-using LLM agents are increasingly deployed in high-stakes workflows where audit trails constitute part of the safety and compliance evidence. We identify a longitudinal integrity failure mode in which, under sustained epistemic pressure and null/contradictory retrieval, agents not only hallucinate facts but also fabricate procedural evidence of verification (e.g., tool execution logs, file paths, timestamps, and intermediate artifacts) to preserve multi-turn narrative coherence. Drawing on 135+ practitioner-documented interactions (Oct 2024–Dec 2025), we formalize this phenomenon as Operator-Induced Longitudinal Integrity Failure (OLIF) and model its progression via the Systemic Cognitive Disintegration (DCS-CI) phases, culminating in Audit-Trail Shadowing—procedural fabrication as a subclass of execution hallucination. We introduce Integrity Delta (ΔI) to quantify fluency–veracity divergence across turns and specify a longitudinal evaluation protocol to detect audit-fabrication under controlled Operational Epistemic Pressure (OEP) using cryptographically grounded tool logs. For remediation, we design a Zero-Trust neuro-symbolic verification stack that binds procedural claims to signed execution records and verifies factual claims against an immutable, PROV-compliant provenance vault. The resulting artifacts support evidence generation for record-keeping, transparency, and robustness duties in high-risk deployments under the EU AI Act. This preprint presents OLIF as a testable hypothesis supported by reproducibility protocols and invites independent replication.