Datasets:

---

SlayerLab

/

polish-dynaword

Like 26

Follow

Fabryka AI 89

Datasheet, evidence file and a new per-record attribution sidecar for sejm_interpellations: Art. 4 wording, System Informacyjny Sejmu credit, acquisition route (v2#107).

What changes:

• data/sejm_interpellations/sejm_interpellations.md: the "Licensing" section opens with a "Legal status (our assessment)" paragraph. The interpellations, written questions and replies are treated as official documents excluded from copyright under Art. 4 pkt 2 of the Polish Copyright Act; Polish law only, status may differ elsewhere; public-domain (official documents) is a descriptive term, not a licence; no CC public-domain mark applied or implied; api.sejm.gov.pl states no licence or terms and no disclaimer was found; third-party fragments and personal data are outside the statement. The Open Data Act is cited as the reuse framework (article list now matches the other Sejm cards). The old line "Same basis as the shipped sejm_api shard" is gone.
• The SIS bullet carries the credit line Źródło: System Informacyjny Sejmu (Kancelaria Sejmu RP) (given as a precaution; the API states no terms and no source says its output is part of SIS). The commercial-offer clause is quoted and described as information only, not a licence term of this dataset. The terms page was read on 2026-10-19; the dated copy and its SHA-256 are in polish-dynaword-v2 (slayerlabs/polish-dynaword-v2#121).
• "Provenance and reproducibility": new Route (API directly, no intermediary) and Acquired bullets. The fetch date is not recorded; the raw file (SHA-256 cc061563...c14a09, 26,503 lines) has the file time 2026-09-05 19:42 CEST and the latest record is 2026-09-03, so the fetch ended between those days. The reproduction command is corrected: the card showed --terms 9,10 --workers 4, but the script has no --workers option and the registry records --terms 10.
• New "Attribution sidecar" section describing the file below and how it was recovered.
• data/sejm_interpellations/sejm_interpellations.license-evidence.json: new sis_terms.live_capture (2026-10-19 read, content hash, both clauses present) and the "live SIS page" entry of not_checked updated to point to it. Nothing else in the file changed.
• New data/sejm_interpellations/sejm_interpellations.attribution.jsonl (LFS, 25,950 lines, one per Parquet row in row order): id, source_url, term, num, kind, date, reply_key (replies), text_sha1, and identical_text_urls where other API documents have exactly the same text (129 rows, 553 documents, equal to the 553 exact duplicates the build dropped). Title and author are not shipped: the title is header chrome kept out of the text and can name private persons (NERGAL pass pending); the author is already a column.
• New src/build_sejm_interpellations_attribution.py and src/test_sejm_interpellations_attribution.py (22 tests). The shard does not record which duplicate survived the build's dedup, because the pool's chunk arrival order is not stored. The builder recovers that order from the shard (the order in which chunks of 2000 raw lines first appear among rows whose text occurs once), replays the dedup over the raw file in that order, and stops if the replay differs from the shard's text sequence, if a chunk cannot be placed, or if any survivor's author differs from the shard's. Ran on the real raw file with no mismatch.
• src/test_sejm_interpellations_contract.py: two tests pin the credit line, the information-only wording, the route sentence, the sidecar mention and the live-capture hash in the card.

Not changed: per-record license values and the registry label (v2#60), the Parquet, stats, src/sources.py, the builder build_dynaword.py. The Parquet added column is still 2026-07-19 (the build default) while the registry and card say 2026-09-05; that belongs to the rebuild.

Limits: the sidecar can be rebuilt only with the raw .jsonl.zst, which is not in the repository. It records which API document each text came from; it does not claim anything about the Kancelaria's rights in it.

Tests: cd src && python3 -m pytest -q -p no:cacheprovider -rfE gives 1 failed (the known test_studia_bas_contract.py::test_studia_bas_artifacts, missing data/studia_bas/studia_bas.parquet in this checkout, same as the baseline), 1236 passed, 7 skipped.

Refs slayerlabs/polish-dynaword-v2#107.

🤖 Generated with Claude Code