System Architecture Archaeology in Practice Case Study II: AI-Assisted Analysis of Roman Dodecahedrons

System Architecture Archaeology in Practice: Case Study II applies the System Architecture Archaeology (SAA) framework to the Roman dodecahedron—one of the most persistent “mystery objects” of the Roman world—demonstrating how structured metadata and AI-assisted modeling can be used to generate falsifiable, system-level hypotheses about anomalous artifacts. Rather than attempting to “solve” the object, this study treats the Roman dodecahedron as institutional hardware: a surviving physical interface whose operational meaning was lost when the administrative and technical systems that produced it collapsed. Using a curated dataset of published museum measurements, material analyses, and geographic distribution data, the artifact is evaluated through the three diagnostic stages of SAA—Institutional Intent, Structural Mirroring, and Operational Semiotics. These data are encoded in a standardized metadata framework (MIMETA: Minimal Institutional Metadata for Artifact Analysis) and provided to an AI model operating under SAA constraints. The model generates multiple competing, system-level theories—each with defined institutional context, operational workflows, and testable predictions—rather than a single speculative interpretation. In this case study, three distinct hypotheses are produced: a military rangefinding device, an industrial calibration gauge, and a textile-standardization tool. Crucially, this paper does not select among these theories. Instead, it demonstrates a scalable research workflow in which AI is used to rapidly generate a falsifiable hypothesis space, while the human researcher remains responsible for data curation, methodological constraint, and empirical testing. This division of labor transforms AI from a narrative generator into a scientific force multiplier for archaeology. Together with the companion SAA framework paper and the Phaistos Disc case study, this work establishes a reproducible computational methodology for investigating undeciphered or functionally orphaned artifacts and provides a template for future large-scale analysis of archaeological “mysteries” grounded in evidence rather than speculation.