Datasets:

---

SlayerLab

/

polish-dynaword

Like 24

Follow

Fabryka AI 87

Tooling only. No data, README or CHANGELOG changes. Tracks slayerlabs/polish-dynaword-v2#61.

src/verify_paper_claims.py parsed the old paper (file.tex), which no longer exists. It now checks polish-dynaword-v2/main.tex:

• tab:releases: documents, tokens and source counts per release, against src/plot_version_growth.py and the release field in src/sources.py. The counts are cross-checked against CHANGELOG.md.
• tab:additions: each release's addition rows, and its "Total added" row, must sum to the change in that release's totals.
• tab:sources: for each source, documents, tokens, first release, and licence (against src/sources.py and the datasheet). Also the Total row, and whether any source is missing from the table or listed without being released.
• Prose: corpus-wide document, token and source counts. A sentence that names one release is checked against that release. A sentence that gives a percentage is skipped, because it describes a subset.

All repository files are read at one ref. The default is the commit that the paper's \snapshot pins (7ae4ac7). --ref worktree reads the files on disk instead. The tolerance is half a unit in the last digit the paper prints. --tex is now required.

src/test_verify_paper_claims.py is now a pytest suite with 44 tests. It previously had no test_ functions, so pytest collected nothing.

• One test checks that a consistent two-source fixture passes.
• 24 parametrised tests each introduce one contradiction and check that the script reports it.
• The rest cover a dual licence, which must WARN rather than FAIL, an unreleased source, a git-pin test against a drifted worktree, the tolerance rule, licence comparison, and parsing the pin.

Every one of 10 hand-made mutations of the script makes at least one test fail.

AGENTS.md "Cut the release" gets one new step: append the release to plot_version_growth.py, then run the paper check. The source-rights-review skill step now gives the full command.

Output on the real paper

At the pin (python3 src/verify_paper_claims.py --tex ../polish-dynaword-v2/main.tex), there are 8 FAILs, all from one cause. At 7ae4ac7, src/plot_version_growth.py still holds the 18-source draft figures for v0.2.5 (4,319,200 docs / 9,639,062,573 tokens). The stats files, the CHANGELOG and the paper all give 4,366,651 / 9,676,513,980. A separate PR fixes the literal on main. The tag itself is immutable, so the check will keep failing there. Every per-source row, the Total row, the prose and the licences pass. There are 2 WARNs: parliamentary is public domain but redistributed as CC-BY-4.0, and the sejm_api datasheet has no **License:** line.

With --ref worktree on current main there are 42 FAILs. These are the expected v0.2.6 changes against the v0.2.5 paper: dropped sources, rebuilt sources, and relabelled licences for eurlex and 1000_novels. The paper is edited after the dataset freeze, so nothing in it changes here.

Checks

• python -m pytest from the repo root: 660 passed.
• render_sample.py: not applicable, because no Parquet file changes.

🤖 Generated with Claude Code