The history of Latin in Polish lands spans from the threshold of the second millennium to at least the end of the 18th century (Axer 2004). Despite significant scholarship, a synthetic approach to the historically variable functions and forms of Latin in Polish territories remains absent. The project DiaLexPoL aims to create a lexical database of Neo-Latin (ca. 1550–1800) building upon the existing Dictionary of Medieval Latin in Poland (SŁŚ) as a diachronic baseline.

We present two experiments evaluating LLM-assisted sense assignment as a core method for compiling this resource (cf. Tiberius et al. 2024; Lugli 2025), focusing on high-frequency, polysemous sociopolitical terms such as CIVITAS, CONSILIUM, and IMPERIUM, whose referents changed substantially across the medieval–early modern transition (McGillivray & Nowak 2025). In Experiment 1, the LLM reassigns dictionary quotations to existing sense definitions; in Experiment 2, it processes novel concordance lines from a Neo-Latin corpus. We vary models (GPT-4o, Claude, Llama, Qwen, Mistral), prompting strategies, and definition languages, measuring accuracy against expert gold standards and inter-annotator agreement. Cases where no existing sense fits are analyzed for genuine semantic innovation in the Neo-Latin period.

We discuss the results within the post-editing lexicography framework (Lew 2024; Jakubíček & Rundell 2023) and reflect on the well-documented risks of LLM adoption, including irreproducibility and loss of evidence-based workflow. Importantly, enthusiastic adoption of LLMs as a way of coping with funding cuts may prove counterproductive if the efficiency gains we demonstrate are used to justify eliminating the very expert labour they currently augment.