Em-dash and function-word shift in Polish parliamentary interpellations (2015–2025) — frozen analysis package

Abstract

Frozen analysis package for a preregistered study of the em-dash (U+2014) as a putative marker of LLM-assisted writing in 60,275 Polish parliamentary interpellations, 2015–2025. The registered decision rule was not met: the decisive cell gave a paired conditional p = 0.00214 but a Newcombe interval of [−4.44; +8.07] pp for the difference, and three registered disconfirmation conditions fired. The registered result is negative. Follow-up tests point to the cause: models writing in Polish, and machine translators, substitute the en-dash (U+2013) — the marker is anglocentric. An exploratory analysis on an external, pre-specified instrument (Śmigiel corpus, parliamentary domain, frozen with its reading rule before measurement) finds a shift of function-word frequencies toward the model profile: rho = +0.368 (95% CI +0.228–0.461) against −0.070 at a term-change placebo. The shift is small in any single office (median rho +0.125) but points the same way in 61 of 84 offices measurable separately (72.6%, one-sided p = 2·10⁻⁵) against 53.8% at the placebo. The package contains every analysis script, every run log, the registration artifacts with SHA-256 manifests, the register of withdrawn claims, the provenance supplement tracing each number to a file and line, and the report of an independent reimplementation carried out without access to the author's code. Source documents are not redistributed: they are public in the Sejm API.