Dependency-Aware Integrity Controls for AI-Assisted Financial Decision Support: Evidence Semantics, Runtime Dependencies, and a Falsifiable Evaluation Protocol

This research protocol presents a falsifiable framework for evaluating integrity controls in AI-assisted financial decision support systems. It introduces and operationalizes the AICOS research architecture together with the Financial Decision Integrity Benchmark (FDIB), with particular emphasis on financial-semantic integrity, evidence qualification, runtime evidence consumption, dependency-aware reassessment, policy binding, human authorization, and reproducibility. The study distinguishes evidence availability from evidence admissibility and actual runtime consumption; model recommendations from institutional decision authority; and decision integrity from predictive correctness. The evaluation protocol compares an ungoverned AI baseline, a provenance/traceability baseline, a deterministic financial-validation baseline, and the full research architecture under controlled financial-semantic, temporal, dimensional, evidence, execution, policy, and authorization mutations. The primary evaluation endpoint is the Unsupported Critical Claim Rate (UCCR), complemented by measures of supported decision coverage, appropriate abstention, dependency-impact precision and recall, reconstruction completeness, and governance overhead. This version is a research protocol and architecture preprint. It does not claim validated credit-prediction accuracy, regulatory certification, production deployment, proven commercial performance, or completed confirmatory experimental results. Empirical performance claims are intentionally deferred until preregistered experiments and statistical validation have been completed. The associated research program is designed around falsifiability, version-bound validation, evidence provenance, reproducibility, explicit uncertainty, and strict separation between implementation, testing, empirical validation, and deployment. Author: Dr. Yasin KalafatoğluORCID: 0009-0009-5063-999XResearch architecture: AICOS®Benchmark: Financial Decision Integrity Benchmark (FDIB)Document type: Research Protocol / Architecture PreprintVersion: 1.0Date: 29 August 2026