QUINCUNX: A Popperian Inference Engine — A prospectus and first registered demonstration: accelerating discovery of where the world's domains join, beginning where the data are sparsest—the world before print

QUINCUNX is a five-stage registered-inference engine — Mint, Select, Grade, Promote, Infer — for domains of found data: evidence accumulated for other purposes, unrepeatable, incomplete, survivorship-shaped and mutually dependent. Large language models supply the conjecture language and mass generator; the engine supplies the disciplines found data forces: demonstrated dataset independence, model memory treated as a contamination route, provenance tiers with an absolute evidence–inference firewall, adversarial rival generation by disjoint model families, and registration of every bet before its evidence is seen. Its inferences may direct the search for new evidence; they may not themselves become evidence.

This prospectus states the design, its Popperian rationale, and the registered conditions under which the thesis should be judged false. It is revised with the engine's first registered demonstrations on the epigraphic record of the Roman world: two Mint–Select runs, a Grade round, a clean-room rival battery over a 105,222-record eligible manifest of the Epigraphic Database Roma executed blind under frozen thresholds, and an interval-certification programme whose failures publish beside its successes. The paired technical record is published separately as First Registered Results from QUINCUNX (DOI 10.5281/zenodo.21839510).