Material Simulator is an open, interactive steel-metallurgy platform that serves the same physics through two synchronized surfaces: a visual, trilingual teaching application (Fe-Fe3C equilibrium, TTT/CCT kinetics with animated microstructure), and a machine-readable API in which every response carries a versioned reasoning contract — explicit assumptions, scope bounds, confidence tiers, refusals (claims the system declines to make), and verification needs. Alongside two educational physics tiers, the platform publishes a growing measured tier of curated transformation datasets (currently 100 datasets from 8 sources) with datum-level provenance: every value carries its exact source location, extraction method, extraction date, and the original publication's license and citation. The paper describes the architecture, the extraction and legal methodology (including the digitization of Davenport and Bain's founding 1930 isothermal-transformation dataset and the complete set of all 89 isothermal-transformation diagrams of the 1951 U.S. Steel Atlas of Isothermal Transformation Diagrams), and a pre-registered four-benchmark evaluation against a frontier large language model. Two of the authors' own initial hypotheses were refuted and are reported as primary findings: textbook equilibrium physics offers no correctness advantage over frontier model weights, and neither does closed-form kinetics computation once the model specification is disclosed — the tested frontier model reproduced a ~200-step Scheil additivity integration to four decimals by unaided arithmetic. What survived testing is a substitutability map across three layers: textbook knowledge and disclosed closed-form computation are both reproduced by the model unaided, whereas provenance-tagged measured values are not (33% accuracy from weights alone, rising only to 60% when full metadata was disclosed without the values, against 100% for an agent with live data access). Retrieval is an independent precondition: in an unprompted-discovery test no open machine-readable resource was located in 3 of 3 tasks. Datasets described in this paper are archived separately under CC BY 4.0 (concept DOI 10.5281/zenodo.21285189), as are the pre-registered benchmarks (10.5281/zenodo.21285520). Platform code is MIT-licensed. Preprint. Includes the PDF and the Markdown source of the manuscript.
Jan Mészáros (Sat,) studied this question.