When the Evaluator Became Part of the Generation System

In the first six articles of this series, we explored a recurring problem: AI-generated images of ancient Rome often looked convincing at first glance but contained subtle—and sometimes obvious—historical errors. Roman streets acquired Victorian lamps. Gladiators carried the wrong equipment. Architecture drifted across centuries. Prompting alone could improve results, but it could not reliably eliminate these failures. This episode examines what happened next. As the correction process matured, a surprising realization emerged: historically constrained image generation is not really a prompt problem. It is a systems problem. Different kinds of historical errors require different kinds of corrections. Architecture failures, social-status mistakes, equipment mismatches, environmental realism issues, and body-composition errors do not all repair the same way. That realization transformed the workflow from simple generation and re-prompting into a structured process involving evaluation, classification, targeted repair, and model specialization. The evaluator also changed roles. Instead of acting solely as a quality gate that accepted or rejected images, it became a production intelligence layer capable of identifying visible evidence, categorizing failures, and generating actionable repair targets. Once historical failures were tagged, localized, and classified, they could be addressed systematically rather than through trial and error. In this episode, we explore the evolution from single-model image generation to evaluator-driven orchestration, where specialized AI systems, structured historical rules, and recursive feedback loops work together to produce more coherent historical reconstructions. The result is not simply better Roman imagery—it is the emergence of a historical correction architecture designed to teach AI systems what history actually requires. Topics include AI image generation, historical reconstruction, prompt engineering, workflow design, model orchestration, evaluator systems, Ancient Rome, and the future of historically constrained visual production.