Code and data accompanying "Beyond cognacy" (corrected study)

The input data are drawn from established, curated, publicly available sources (Lexibank, Grambank, Glottolog). All derived results were produced by the included pipeline in a single documented analysis campaign (August 2026) with fixed random seeds; the complete run logs are part of the deposit. The final evaluation tables are identical to the numbers reported in the corrected version of the accompanying paper (arXiv:2507.03005).

Complete: the deposit contains the full working directory of the analysis, including the primary data as downloaded on 2026-08-03, all intermediate results (wordlists, distances, cognate clusterings, alignments, character matrices), the tree inference outputs for the full dataset, the 15 families and the 100 subsamples, the final evaluation tables, the figures, the code as run, the conda and Julia environment specifications, and the run logs. The parameters of the pair-Hidden Markov Model, which were trained on a corpus shared with a related project, are included as code/phmm_parameters.json. Excluded is only an abandoned earlier run that contributed to no reported result.

The deposit follows the FAIR data principles. All input data come from publicly available, versioned sources with documented licenses. File formats are open and non-proprietary (CSV, PHYLIP, Newick, JSON, plain text). Software dependencies are specified in machine-readable environment files, and the workflow is documented in the code repository.

All results stem from a single analysis campaign with fixed random seeds, not from a mixture of runs. The final evaluation tables can be re-derived from the included character matrices and trees, which in turn derive from the included primary data via the included pipeline. The archive passed an integrity test, and the MD5 checksum of the uploaded file was verified against the local copy.

The study was carried out at the University of Tübingen. The underlying paper was peer-reviewed and published in the Proceedings of SIGTYP 2025. The post-publication corrections are transparently documented in the corrected paper, in the code repository, and in this deposit; the development history is public at https://codeberg.org/profgerhard/sigtyp2025_code/.

All files are machine-readable with standard open-source tools. The analysis stack (Julia, Python, R, RAxML-NG, GNU parallel) is free and open source; exact dependency versions are pinned in the included environment files (conda yml, Julia Project.toml and Manifest.toml).

The deposit supports research in computational historical linguistics and linguistic typology: it allows full replication of the reported comparison between expert cognacy, automatic cognate clustering, and multiple sequence alignment as bases for phylogenetic inference, and the derived character matrices, alignments and trees for 974 languages can be reused in follow-up studies.

The deposit reflects the corrected study of August 2026, which supersedes the originally published 2025 analysis. The dataset of the original version remains available on Zenodo (doi:10.5281/zenodo.15678468). The Lexibank, Grambank and Glottolog data are the versions current at the download date (2026-08-03), which is recorded in the deposit.

DEPOSIT_README.md at the root of the archive provides a detailed inventory of the contents. The analysis workflow is documented step by step in the code repository (workflow.md and README). The methods are described in the accompanying paper; the corrections relative to the originally published version are documented in the paper's title footnote. The lexical data follow the CLDF conventions of Lexibank.

The conda environments are specified in sigtyp2025_code.yml and sigtyp2025_r.yml and can be recreated with:

mamba env create -f sigtyp2025_code.yml mamba env create -f sigtyp2025_r.yml

The Julia environment is set up with:

cd code julia --project=. -e "using Pkg; Pkg.instantiate()"

The pipeline steps and the sbatch scripts for the tree searches are documented in code/ and in workflow.md of the code repository; the driver scripts of the actual run (run_pipeline*.sh) are included at the top level of the archive.