From Proposal Error to Wrong Final State: A Controlled Simulation Framework for Failure Propagation Across Agentic Architectures

From Proposal Error to Wrong Final State examines a systems-level question in AI reliability: when a model or agent makes an upstream error, what determines whether that error is contained or propagates into an incorrect real-world outcome? The paper introduces a controlled simulation framework that follows faults beyond model output, tracing their progression through proposal, admission, commitment, external effects, reconciliation, and final authoritative state. The study comprises 144 synthetic enterprise scenarios across six fault classes and three architecture bundles, producing 5,184 protocol-conformant simulated trajectories with event-level traces. The framework separates proposal correctness from downstream system behavior, enabling analysis of how the allocation of authority across models, software controls, and execution infrastructure shapes failure propagation and containment. The broader objective is to provide a reproducible methodology for evaluating AI systems at the level of operational outcomes rather than model outputs alone. This release includes the research manuscript, generated datasets and event traces, analysis and reproducibility materials, and a Stage 1–9 provenance archive.