This paper proposes the Observer-Dependent Pipeline Model (ODPM), a theory-building framework for evaluating generative AI systems with conversational front ends as full deployment pipelines rather than isolated models or interfaces. The framework separates concept formation, data/evidence quality, backend capability, orchestration, front-end fit, mission alignment, safety, observability, and post-release learning into analytically distinct yet dynamically coupled layers. ODPM introduces an upstream concept gate, a distinct data/evidence regime, backend and front-end phase indices, an integrated coherence index with a bridge penalty for backend-front-end mismatch, and a local emergence-mitigation (LEM) decomposition that distinguishes structural hazard from observed incident frequency. The paper is explicitly theory-building rather than fully calibrated prediction: it offers operational proxies, falsifiers, mandatory baselines, and two bounded deployment case studies to test whether the framework generates diagnostic value beyond model-only, chat-only, governance-only, or benchmark-centered framings. The central argument is that the relevant object of analysis is the deployed system as a coherent or incoherent lifecycle structure, and that accountability remains human and organizational even when conversational interfaces become highly salient.
Andrea Viliotti (Wed,) studied this question.