Large language models (LLMs) are increasingly used for lexicographic support, neology detection, and semantic categorization, yet their behaviour on historical newspapers remains under-evaluated. This short paper describes an ongoing project that extends a DH2026-accepted two-phase methodology for extracting and tracking rumours in historical newspapers. From large-scale US and UK corpora (PleIAs/US-PD-Newspapers; biglam/hmd_newspapers), the DH workflow produces gold-standard sentence-level rumour instances with proposition-like “rumour content” spans (Rumour_Content/Cleaned_Content) and extraction-pattern metadata. Building on these historically grounded units, we propose an LLM-centered benchmark and analysis pipeline for assigning topical frames and evidential stance to rumour propositions, and for auditing “temporal projection” when models introduce anachronistic modern misinformation framings. For controlled cross-variety comparison we construct a strictly balanced benchmark of 800 instances over two well-attested bins (1840–1859, 1860–1879) and both national varieties (200 per country per bin). We outline prompt conditions (text-only vs time-aware vs historically calibrated) and self-consistency voting to quantify label stability and error modes. A small manually annotated subset supports evaluation, while the main contribution is the benchmark design, prompts, and reproducible protocol enabling community feedback before full-scale results are finalized.