the em-dash em-beds in Congress: frozen analysis package (code, derived per-release measurements, validation and run logs), 2021-2025

Frozen analysis package supporting the preprint "the em-dash em-beds in Congress: a population-level rise in em-dash frequency in U.S. congressional press releases at the dawn of the large-language-model era, 2021-2025" and the OSF preregistration DOI 10.17605/OSF.IO/U5NEY (frozen 2026-06-06, before any confirmatory number was computed). CONTENTS (26 files, 93 MB uncompressed) scripts/ - the frozen canonical pipeline (congress_press_run.py for corpus build, congress_press_aggregate.py as the canonical aggregator whose SHA-256 is recorded in the preregistration, congress_press_audit.py, congress_press_validate.py for dateline-cut validation) plus the post-registration analysis scripts, each labelled exploratory in the preprint: party and chamber symmetry, the lexical control, document-level and quarterly interrupted time series, positional sensitivity, and the W35 pass (three placebo cut-offs, the exact daily-cut-off H3, the exclusion decomposition, and the leave-out-top-5%-of-offices sensitivity). data/ - per-release measurements only, not article text. uspress_results_*.csv (146,239 releases), uspress_contexts_*.csv (201,800 dash contexts, each a 61-character window around one dash), and uspress_canonical_*.csv (the canonical analysis snapshot produced by the frozen aggregator, which is the input to every confirmatory number). Full release texts are deliberately not redistributed, in line with the OSF declaration; the source corpus is the public congress-press-releases scraper output. validation/ - the dateline-stripping validation sample (300 releases sampled, 123 manually adjudicated, 122 of 123 cuts correct, Wilson 95% CI 95.5-99.9%) and the adjudication report. logs/ - the run log of every analysis pass cited in the preprint. MANIFEST_SHA256.txt - SHA-256 of every file in the package. The three fingerprints recorded in the OSF preregistration (the canonical aggregator and the two registered CSVs) are reproduced and confirmed to match verbatim, which is what lets a reader check that the code and data behind the confirmatory numbers are the ones that were registered before the analysis. STATUS The preregistered residual-dateline-leak validation gate was formally breached: the full-corpus leak rate was 7.00 / 4.89 / 3.94 / 4.95 / 4.94 per cent per year across 2021-2025, above the registered 2 per cent threshold. The study's full preregistered decision rule was therefore not met, and the LLM-diffusion interpretation is reported as exploratory rather than confirmatory. The deviation is disclosed in Section 4.6 of the preprint. The scripts and logs deposited here are the material supporting that disclosure; nothing in this package was rerun or adjusted after the fact. ENVIRONMENT Python 3.13.6, numpy 2.4.4, pandas 2.3.3, statsmodels 0.14.6, Windows 11. Package built 2026-08-06.