wikipedia: licence labels by contribution date (v2#110), seven pages withheld

artifacts/source_findings.md

@@ -2281,8 +2281,8 @@ faults in `drop_residue`, fixed for every source:
   they can be a sentence about the option; they are now removed only next to a pipe;
 - the trailing-options rule took the `| class=...` of a table start `{| class=...` for a file
   option and left a lone `{`; it now skips a pipe after `{`. Four wikivoyage rows released in Hub
-  PR 125 have this fault (`wikivoyage_2055`, `2564`, `9129`, `10490`); a follow-up PR rebuilds
-  them.
+  PR 125 have this fault (`wikivoyage_2055`, `2564`, `9129`, `10490`). A rebuild would only
+  turn the lone `{` back into the table markup `{| class="wikitable"`, so they are not rebuilt.
 
 Known limits, as matched in the cleaned text (documents; a match is not always residue):
 `...thumb|` glued to the word before it 10, a line holding only `mały` 18, `Kategoria:` inside a
@@ -2340,6 +2340,91 @@ both. `wikibooks_3083`, whose page states CC BY-SA 3.0, was 4.0 by its last edit
 
 README, CHANGELOG and totals are left to v2#80.
 
+## wikipedia — licence labels by contribution date, seven pages withheld (update to the sections above)
+
+**Investigated** — 2026-10-18 (slayerlabs/polish-dynaword-v2#110 option B, as for wikivoyage and
+wikibooks above; the deletion logs of released pages missing from the 2026-10-01 stub dump).
+
+**Verdict** — `kept`, seven more pages withheld, relabelled: 1,468,647 documents, 3,140,823,417
+characters, 1,226,906,115 tokens (was 1,468,654 / 3,140,838,409 / 1,226,912,191). Labels per
+record: `CC-BY-SA-3.0` 719,119 documents (364,335,667 tokens, 29.7%), `CC-BY-SA-3.0 AND
+CC-BY-SA-4.0` 629,454 (765,350,761, 62.4%), `CC-BY-SA-4.0` 120,074 (97,219,687, 7.9%). The
+last-revision labels of the sections above were 3.0 190,836 (81,022,938 tokens) and 4.0
+1,277,818 (1,145,889,253). Every 3.0 row stays 3.0; of the 1,277,811 4.0 rows still released,
+528,283 are now 3.0, 629,454 both and 120,074 stay 4.0. Each sidecar line gains `license_basis`
+and `pre_cutoff_revid`; the build record holds the stub dump and the count per basis; the card
+explains each basis.
+
+**Withheld: seven pages deleted from plwiki for copyright after the dump** —
+`wikipedia_4095632` Brzezina (opowiadanie), deleted 2026-09-01, log reason "Naruszenie praw
+autorskich" with a studocu.com link (not fetched); `wikipedia_6131122` Edmund Bielecki, `6131594`
+Ryszard Cyffer, `6131631` Stanisław Czarnomski, `6131646` Mikołaj Dmochowski (deleted
+2026-09-05), `6132108` Włodzimierz Daszkiewicz and `6132151` Konstanty Jankowski (oficer)
+(2026-09-06), log reason "naruszenie praw autorskich". 14,992 characters, 6,076 tokens, all
+labelled 4.0 by their last revision. None is in the stub dump, so their history cannot be
+checked. They are in `COPIED_PAGES` (9 pages now) with the day and the log reason.
+`fetch_wikipedia.py withhold` drops a page listed after the build from the committed release, as
+`prepare` would, and adds it to the copyright sample's `rows_dropped_since`, so the sample's draw
+still reproduces; it refuses to run after `relabel`.
+
+**No residue cleanup (v2#116)** — `strip_article` already drops category and file links, so
+`drop_residue` is not applied. Run on this release it would change 426 documents and remove
+477,849 characters, nearly all content: flattened medal tables whose line starts
+`Kategoria:Złoto:Srebro:Brąz:` and goes on with the results (`wikipedia_1515505`), and field
+labels such as `Kategoria: olej nieschnący` (`wikipedia_127177`). v2#116 read the 410 rows with a
+`Kategoria:` line as medal-table headers; most are, but 14 documents hold a category link on a
+line of its own (16 lines, e.g. `wikipedia_103834` `Kategoria:Wielcy marszałkowie zakonu
+krzyżackiego`), left by malformed links. Also left: 13 documents with an unclosed file link
+(`[[Plik:...|thumb|...`, `wikipedia_13010`) and 363 with `<ref`. Listed in the card, not cleaned
+in v0.2.6.
+
+**Enumeration** — `plwiki-20261001-stub-meta-history.xml.gz` (7,039,541,375 bytes, sha1
+`47fd7549…` matched to the dump's `dumpstatus.json`, 2026-10-18). plwiki's stub dump is written
+in parts; the whole file is listed under the `xmlstubsdumprecombine` job, which
+`wiki_history.download` now reads. 190,836 pages were last edited before the cutoff (3.0);
+120,074 have no revision before it (4.0); for 5,028 the shipped revision's sha1 equals the
+pre-cutoff one (3.0). For the other 1,152,624 the pre-cutoff revision was fetched from the
+pl.wikipedia.org API by revision id, converted with `strip_article` and compared as for
+wikivoyage: 492,757 give the shipped text and 30,498 lose words only (3.0); 629,338 gain at least
+one word (both); for 31 the pre-cutoff text is hidden (both). 85 pages edited after the cutoff
+are not in the stub dump (both, below). `compare` ran as 12 sidecar slices in parallel, joined in
+sidecar order (a test checks that joined slices equal one run); its labels were computed before
+the withholding and filtered to the sidecar after it. A rerun of relabel is byte-identical.
+
+The cutoff day: 76 pages created on 2023-06-07 are 4.0 alone. Of the 184 documents last edited
+that day, 37 gain words only through that day's edits (both) and 4 were created that day (4.0);
+if the switch came later that day, those 41 are 3.0 alone.
+
+**Deleted from plwiki since the dump** — 104 released pages (before the withholding) are missing
+from the stub dump. The deletion log of each was read from the pl.wikipedia API
+(`list=logevents` by title) and classed by its latest deletion: 7 for a copyright violation
+(withheld, above); 88 after a deletion discussion (Poczekalnia); 7 as a duplicate, needless
+disambiguation, advertising, spam or nonsense (`wikipedia_239449` Monceaux (Calvados), `1342574`
+Rhenus, `2550999` Odcinacz, `2977355` Feliks (męczennik), `5437048` Gropa (Połonina Czerwona),
+`5860110` Chór Akademicki Uniwersytetu Szczecińskiego, `5964373` Softswiss); 2 for a history
+merge or a move and then restored (`wikipedia_99795` Młotogłów wielkogłowy, 2026-09-18;
+`wikipedia_1105144` 103 męczenników koreańskich, 2026-09-21). Of the 97 still released, 85 are
+labelled both (no history to compare) and 12 were last edited before the cutoff (3.0 by their
+shipped revision). They are kept: their deletions are a quality call, not a rights one.
+`wikipedia_4033410` Teatr Piosenki Młyn, deleted after a discussion on 2026-09-30, was also
+deleted for copyright on 2017-05-23 and recreated by 2017-07-14; the released text was not
+compared with the deleted one. Counts per class are in the evidence's `deleted_since_dump`.
+
+**Blockers** — none to merge. Open:
+- Any added word makes a page 3.0 and 4.0; there is no threshold.
+- Text moved or merged from an older page into a page created after the cutoff is labelled 4.0
+  alone.
+- The time of day of the switch (the cutoff day above).
+- The comparison is on converted text, so words that `strip_article` drops (templates, `<ref>`
+  bodies) are not compared.
+- The 95 pages deleted after a discussion or as spam or duplicates stay in the release; Młyn is
+  not compared; pages deleted after 2026-10-01 and other pages by the authors of the seven are
+  not checked.
+- As in the sections above: pastes beyond the copyright sample, GFDL-era text, appendix headings,
+  the NERGAL pass.
+
+README, CHANGELOG and totals are left to v2#80.
+
 ## `studia_bas`: licence version resolved to CC BY 3.0 PL
 
 **Investigated:** 2026-10-06

data/wikipedia/wikipedia.attribution.jsonl

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:046067a7992e0f98950ba6d7e0e8b7499e75e58bb5aac0a64e0779e6d69e0a2e
-size 344192023
+oid sha256:aac58a47f87d548e0a42477e41a686a18705b86b95b25f2ce45d89d27bf38d09
+size 455694527

data/wikipedia/wikipedia.build.json

@@ -1,6 +1,6 @@
 {
-  "attribution_rows": 1468654,
-  "attribution_sha256": "046067a7992e0f98950ba6d7e0e8b7499e75e58bb5aac0a64e0779e6d69e0a2e",
+  "attribution_rows": 1468647,
+  "attribution_sha256": "aac58a47f87d548e0a42477e41a686a18705b86b95b25f2ce45d89d27bf38d09",
   "by_sa_4_cutoff": "2023-06-07T00:00:00+00:00",
   "copyvio_tagged": [
     {
@@ -199,6 +199,26 @@
       "url": "https://dumps.wikimedia.org/plwiki/20260901/plwiki-20260901-pages-articles6.xml-p4962394p6142332.bz2"
     }
   ],
+  "licence_labels": {
+    "cases": {
+      "changed_after_cutoff": 629338,
+      "created_after_cutoff": 120074,
+      "identical_text": 492757,
+      "no_text_added": 30498,
+      "page_not_in_dump": 85,
+      "pre_text_hidden": 31,
+      "same_revision_text": 5028,
+      "shipped_before_cutoff": 190836
+    },
+    "created_on_cutoff_day": 76,
+    "stub": {
+      "checked_at": "2026-10-18",
+      "name": "plwiki-20261001-stub-meta-history.xml.gz",
+      "sha1": "47fd7549026521bd02c174ee418c63a93f7e4118",
+      "size": 7039541375,
+      "url": "https://dumps.wikimedia.org/plwiki/20261001/plwiki-20261001-stub-meta-history.xml.gz"
+    }
+  },
   "prepare": {
     "copied_pages": [
       {
@@ -210,12 +230,47 @@
         "pageid": "1126780",
         "reason": "Czarnka: the lake paragraph (revision 42979631, 2015-06-19) closely paraphrases the Drawieński Park Narodowy page dpn.pl/czarne, archived 2012-01-25; whether Art. 4 of the Polish copyright act (official materials) applies was not assessed",
         "title": "Czarnka"
+      },
+      {
+        "pageid": "4095632",
+        "reason": "Brzezina (opowiadanie): deleted from plwiki on 2026-09-01 for a copyright violation, deletion log: \"Naruszenie praw autorskich: https://www.studocu.com/pl/document/uniwersytet-jagiellonski/filologia-polska/proza-poetycka-i-awangardowa-iwaszkiewicz-jasienski-choromanski/74568413\"",
+        "title": "Brzezina (opowiadanie)"
+      },
+      {
+        "pageid": "6131122",
+        "reason": "Edmund Bielecki: deleted from plwiki on 2026-09-05 for a copyright violation, deletion log: \"[[Wikipedia:Prawa autorskie|naruszenie praw autorskich]]\"",
+        "title": "Edmund Bielecki"
+      },
+      {
+        "pageid": "6131594",
+        "reason": "Ryszard Cyffer: deleted from plwiki on 2026-09-05 for a copyright violation, deletion log: \"[[Wikipedia:Prawa autorskie|naruszenie praw autorskich]]\"",
+        "title": "Ryszard Cyffer"
+      },
+      {
+        "pageid": "6131631",
+        "reason": "Stanisław Czarnomski: deleted from plwiki on 2026-09-05 for a copyright violation, deletion log: \"[[Wikipedia:Prawa autorskie|naruszenie praw autorskich]]\"",
+        "title": "Stanisław Czarnomski"
+      },
+      {
+        "pageid": "6131646",
+        "reason": "Mikołaj Dmochowski: deleted from plwiki on 2026-09-05 for a copyright violation, deletion log: \"[[Wikipedia:Prawa autorskie|naruszenie praw autorskich]]\"",
+        "title": "Mikołaj Dmochowski"
+      },
+      {
+        "pageid": "6132108",
+        "reason": "Włodzimierz Daszkiewicz: deleted from plwiki on 2026-09-06 for a copyright violation, deletion log: \"[[Wikipedia:Prawa autorskie|naruszenie praw autorskich]]\"",
+        "title": "Włodzimierz Daszkiewicz"
+      },
+      {
+        "pageid": "6132151",
+        "reason": "Konstanty Jankowski (oficer): deleted from plwiki on 2026-09-06 for a copyright violation, deletion log: \"[[Wikipedia:Prawa autorskie|naruszenie praw autorskich]]\"",
+        "title": "Konstanty Jankowski (oficer)"
       }
     ],
     "counts": {
       "CC-BY-SA-3.0": 284075,
-      "CC-BY-SA-4.0": 1421860,
-      "copied_page": 2,
+      "CC-BY-SA-4.0": 1421853,
+      "copied_page": 9,
       "ek_copyright": 0,
       "exact_duplicate": 382,
       "extracted": 1706319

data/wikipedia/wikipedia.license-evidence.json

@@ -1,11 +1,11 @@
 {
- "license": "CC-BY-SA-3.0 or CC-BY-SA-4.0 per record, by the page's last revision",
+ "license": "CC-BY-SA-3.0, CC-BY-SA-4.0 or both per record, by contribution date",
  "spdx": [
   "https://spdx.org/licenses/CC-BY-SA-3.0.html",
   "https://spdx.org/licenses/CC-BY-SA-4.0.html"
  ],
  "status": "documented-open-cc-by-sa",
- "review_scope": "licence of contributors' text on pl.wikipedia and its version per page; copyright pastes estimated by sample; not independent legal advice",
+ "review_scope": "licence of contributors' text on pl.wikipedia and its version per contribution; copyright pastes estimated by sample; not independent legal advice",
  "checked_at": "2026-10-05",
  "evidence_url": "https://foundation.wikimedia.org/w/index.php?title=Policy:Terms_of_Use&oldid=554823",
  "sha256": "c96ef93b439ff9b8242dc3233f41ede57c0984ef9153fafa51aa5f8cf0678832",
@@ -25,25 +25,19 @@
    "url": "https://creativecommons.org/licenses/by-sa/4.0/deed.pl"
   }
  },
- "licence_text": {
-  "url": "https://creativecommons.org/licenses/by-sa/3.0/legalcode",
-  "sha256": "3f941b3b89cf7b8370ceb83cc76d2120d471b58735d8ca60238a751a48d7f72f",
-  "section": "4(b)",
-  "statement": "You may Distribute or Publicly Perform an Adaptation only under the terms of: (i) this License; (ii) a later version of this License with the same License Elements as this License"
- },
  "label_rule": {
-  "function": "src/fetch_wikipedia.py license_for_revision",
+  "function": "src/wiki_history.py compare; src/fetch_wikipedia.py relabel writes the labels",
   "cutoff": "2023-06-07T00:00:00+00:00",
-  "rule": "last revision before the cutoff -> CC-BY-SA-3.0; at or after -> CC-BY-SA-4.0; a timestamp without a time zone or missing stops the build",
-  "cutoff_day": "the Terms of Use give the day, not the time; midnight UTC labels the whole of 2023-06-07 as 4.0. If the switch came later that day, pages last edited before it are 3.0 text under a 4.0 label. The rights audit (wikimedia.md L3) suggested 2023-06-08; that would label 4.0 edits of 2023-06-07 as 3.0, which 4.0 §3(b)(1) does not allow. Neither choice is exact without the time of the switch",
+  "rule": "each contribution keeps the licence it was made under: a row whose text was all contributed before the cutoff -> CC-BY-SA-3.0; a page created at or after it -> CC-BY-SA-4.0; a row holding text from both sides, or whose pre-cutoff text cannot be compared, -> CC-BY-SA-3.0 AND CC-BY-SA-4.0. The basis is license_basis per row in wikipedia.attribution.jsonl, counted in wikipedia.build.json licence_labels",
+  "cutoff_day": "the Terms of Use give the day, not the time; midnight UTC counts edits made on 2023-06-07 as 4.0. If the switch came later that day, an older page that gained words before it is labelled both (3.0 is still named), and a page created before it is 3.0 text labelled 4.0 alone (pages_created_on_cutoff_day). The rights audit (wikimedia.md L3) suggested 2023-06-08; that would count 4.0 edits of 2023-06-07 as 3.0, which 4.0 §3(b)(1) does not allow. Neither choice is exact without the time of the switch",
   "licence_text_4": {
    "url": "https://creativecommons.org/licenses/by-sa/4.0/legalcode",
    "sha256": "28a9529c7d0bb4dc51f4bf5c116a3d16ef247a052f7591466768ddf563fd1cf5",
    "section": "3(b)(1)",
    "statement": "The Adapter's License You apply must be a Creative Commons license with the same License Elements, this version or later, or a BY-SA Compatible License."
   },
-  "pages_last_revised_on_cutoff_day": 184,
-  "tests": "src/test_wikipedia_contract.py test_license_for_revision, test_license_for_revision_fails_without_a_zoned_timestamp, test_every_row_label_follows_its_last_revision"
+  "pages_created_on_cutoff_day": 76,
+  "tests": "src/test_wikipedia_contract.py test_every_row_is_labelled_by_contribution; src/test_wiki_history.py"
  },
  "chain": [
   {
@@ -59,23 +53,20 @@
    "mark": "V"
   },
   {
-   "hop": "page last edited on or after 2023-06-07 -> released as 4.0",
-   "basis": "CC BY-SA 3.0 §4(b)(ii) lets an adaptation be licensed under a later version with the same License Elements",
-   "url": "https://creativecommons.org/licenses/by-sa/3.0/legalcode",
-   "mark": "I"
+   "hop": "page history -> licence label per row",
+   "basis": "stub dump plwiki-20261001-stub-meta-history.xml.gz (sha1 matched to dumpstatus.json) and the text of each page's last revision before 2023-06-07 from the MediaWiki API, converted with strip_article and compared with the shipped text (src/wiki_history.py); any added word -> CC-BY-SA-3.0 AND CC-BY-SA-4.0",
+   "url": "https://dumps.wikimedia.org/plwiki/20261001/plwiki-20261001-stub-meta-history.xml.gz",
+   "mark": "V"
   }
  ],
- "attribution": "author = \"Wikipedia contributors, <page URL>\"; the page history names the contributors; Terms of Use section 7 accepts a link or URL to the page. wikipedia.attribution.jsonl: id, pageid, revid, title, url, last_revision_at, license per row",
+ "attribution": "author = \"Wikipedia contributors, <page URL>\"; the page history names the contributors; Terms of Use section 7 accepts a link or URL to the page. wikipedia.attribution.jsonl: id, pageid, revid, title, url, last_revision_at, license, license_basis, pre_cutoff_revid per row",
  "intermediary_removed": "v0.2.0-v0.2.5 used the SpeakLeash plwiki snapshot (speakleash-ds-pub); the v0.2.6 rebuild reads the official dump directly",
- "finding": "Contributors' text is CC BY-SA; its version follows the page's last revision. Pages carrying the plwiki copyright-violation notice ({{NPA}}, {{Copyvio}}, {{Plagiat}}) are dropped, as are pages whose speedy-deletion notice {{ek}} gives a copyright reason (none in this dump) and the pages the copyright sample found copied (src/fetch_wikipedia.py COPIED_PAGES).",
+ "finding": "Contributors' text is CC BY-SA; each contribution keeps the version it was made under, and each row is labelled 3.0, 4.0 or both by the contributions its text holds. Pages carrying the plwiki copyright-violation notice ({{NPA}}, {{Copyvio}}, {{Plagiat}}) are dropped, as are pages whose speedy-deletion notice {{ek}} gives a copyright reason (none in this dump) and the pages found to hold copied text, by the copyright sample or by a plwiki deletion for copyright after the dump (src/fetch_wikipedia.py COPIED_PAGES).",
  "release_scope": {
   "labels": {
-   "CC-BY-SA-4.0": 1277818,
-   "CC-BY-SA-3.0": 190836
-  },
-  "tokens_by_label": {
-   "CC-BY-SA-4.0": 1145889253,
-   "CC-BY-SA-3.0": 81022938
+   "CC-BY-SA-3.0": 719119,
+   "CC-BY-SA-3.0 AND CC-BY-SA-4.0": 629454,
+   "CC-BY-SA-4.0": 120074
   },
   "pages_scanned": 2877278,
   "namespace_0": 2304425,
@@ -85,13 +76,18 @@
   "parser_failures": 0,
   "converted": 1706319,
   "exact_duplicates": 382,
-  "copied_pages": 2,
+  "copied_pages": 9,
   "ek_notices": 0,
   "ek_copyright": 0,
   "under_200_chars": 231717,
   "non_polish": 5564,
-  "documents_released": 1468654,
-  "last_revised_before_2009": 0
+  "documents_released": 1468647,
+  "last_revised_before_2009": 0,
+  "tokens": {
+   "CC-BY-SA-3.0": 364335667,
+   "CC-BY-SA-3.0 AND CC-BY-SA-4.0": 765350761,
+   "CC-BY-SA-4.0": 97219687
+  }
  },
  "copyvio_notice_recheck": {
   "checked_at": "2026-10-05",
@@ -135,7 +131,7 @@
  "copyright_sample": {
   "file": "wikipedia.copyvio-sample.jsonl",
   "checked_at": "2026-10-05",
-  "method": "random.seed(seed); per label in sorted order, random.sample of that label's released ids in page-id order; each sampled revision (revid) run through Earwig's Copyvio Detector API (copyvios.toolforge.org, action=search, use_engine=1, use_links=1); every page whose best source reached confidence 0.10 or a 'possible'/'suspected' verdict was compared by hand (Earwig compare detail, the page history, the source's date)",
+  "method": "random.seed(seed); per stratum in sorted order, random.sample of that stratum's released ids in page-id order; each sampled revision (revid) run through Earwig's Copyvio Detector API (copyvios.toolforge.org, action=search, use_engine=1, use_links=1); every page whose best source reached confidence 0.10 or a 'possible'/'suspected' verdict was compared by hand (Earwig compare detail, the page history, the source's date)",
   "seed": 20261006,
   "per_stratum": {
    "CC-BY-SA-3.0": 40,
@@ -216,18 +212,63 @@
     [
      "wikipedia_1126780",
      "CC-BY-SA-4.0"
+    ],
+    [
+     "wikipedia_4095632",
+     "CC-BY-SA-4.0"
+    ],
+    [
+     "wikipedia_6131122",
+     "CC-BY-SA-4.0"
+    ],
+    [
+     "wikipedia_6131594",
+     "CC-BY-SA-4.0"
+    ],
+    [
+     "wikipedia_6131631",
+     "CC-BY-SA-4.0"
+    ],
+    [
+     "wikipedia_6131646",
+     "CC-BY-SA-4.0"
+    ],
+    [
+     "wikipedia_6132108",
+     "CC-BY-SA-4.0"
+    ],
+    [
+     "wikipedia_6132151",
+     "CC-BY-SA-4.0"
     ]
    ],
    "note": "the sample was drawn from this earlier build; test_copyright_sample_matches_the_release reproduces it from the released ids plus rows_dropped_since, each of which must be a copyright drop"
-  }
+  },
+  "strata_are": "the labels by last revision the release carried before the relabel by contribution (v2#110): CC-BY-SA-3.0 for a page last edited before 2023-06-07, CC-BY-SA-4.0 from that day (src/fetch_wikipedia.py license_for_revision)"
  },
  "not_checked": [
   "pastes from copyrighted sources beyond the sample above",
-  "the 3.0 -> 4.0 label of pages edited after the cutoff (inferred from 3.0 §4(b)(ii), not reviewed by a lawyer), and the time of day of the switch (184 pages last edited on 2023-06-07)",
+  "whether a small post-cutoff edit is a contribution of its own: any added word makes the row CC-BY-SA-3.0 AND CC-BY-SA-4.0, with no threshold",
+  "text moved or merged from an older page into a page created after the cutoff (labelled CC-BY-SA-4.0 alone)",
+  "the time of day of the switch (76 pages created on 2023-06-07 labelled CC-BY-SA-4.0 alone)",
   "text imported from elsewhere under another licence or with extra attribution requirements (Terms of Use, Re-use); template notices are removed by the conversion",
   "text first contributed under GFDL before the 2009 move to CC BY-SA (0 pages last edited before 2009); GFDL dual licensing; only the CC BY-SA grant is relied on",
   "quotations inside articles (Art. 29 of the Polish copyright act)",
-  "deletion discussions and maintenance notices other than {{NPA}} and {{ek}}; their reasons are not read",
+  "deletion discussions and maintenance notices other than {{NPA}} and {{ek}}; their reasons are not read, except the deletion logs of pages deleted since the dump (deleted_since_dump)",
+  "pages deleted from plwiki after the stub dump (2026-10-01), and other pages by the authors of the 7 deleted for copyright",
+  "Teatr Piosenki Młyn (wikipedia_4033410), deleted for copyright on 2017-05-23 and recreated by 2017-07-14: the released text was not compared with the deleted one",
   "personal data (NERGAL scrub runs before release)"
- ]
+ ],
+ "deleted_since_dump": {
+  "checked_at": "2026-10-18",
+  "method": "released pages missing from the stub dump plwiki-20261001-stub-meta-history.xml.gz; the deletion log of each read from the pl.wikipedia API (list=logevents by title), classed by its latest deletion entry",
+  "missing_from_stub": 104,
+  "latest_deletion": {
+   "copyright violation (withheld, COPIED_PAGES)": 7,
+   "after a deletion discussion (Poczekalnia)": 88,
+   "duplicate, needless disambiguation, advertising, spam or nonsense": 7,
+   "for a history merge or a move, then restored": 2
+  },
+  "kept": "pages deleted for a reason other than copyright stay in the release: a quality call, not a rights one"
+ }
 }

data/wikipedia/wikipedia.md

@@ -6,7 +6,7 @@ Polish Wikipedia
 - **Source (upstream):** https://dumps.wikimedia.org/plwiki/20260901/
 - **Domain:** encyclopedic
 - **Language:** Polish (pl)
-- **License:** `CC-BY-SA-3.0` or `CC-BY-SA-4.0` per record (by last-revision date)
+- **License:** `CC-BY-SA-3.0`, `CC-BY-SA-4.0` or both per record (by contribution date)
 - **Created (range):** 2010-11-12, 2026-09-01 (date of the page's last revision in the dump)
 - **Added:** 2026-06-15; rebuilt 2026-10-05 (UTC) from the official Wikimedia dump
 
@@ -14,25 +14,45 @@ Polish Wikipedia
 Wikipedia contributors license their text under CC BY-SA 4.0 under the Wikimedia
 [Terms of Use](https://foundation.wikimedia.org/w/index.php?title=Policy:Terms_of_Use&oldid=554823) in effect since 7 June 2023, and under CC BY-SA 3.0 under the
 [2014-2023 Terms of Use](https://foundation.wikimedia.org/w/index.php?title=Policy:Terms_of_Use&oldid=98524) (both read 2026-10-05 UTC). The pl.wikipedia site licence
-(`rightsinfo`) links BY-SA 4.0. Each row is labelled by its page's last revision in the
-dump:
-
-| last revision (UTC) | `license` | documents | tokens | share of tokens |
-|---|---|---:|---:|---:|
-| before 2023-06-07 | `CC-BY-SA-3.0` | 190,836 | 81,022,938 | 6.6% |
-| 2023-06-07 or later | `CC-BY-SA-4.0` | 1,277,818 | 1,145,889,253 | 93.4% |
-
-A page last edited on or after the cutoff can still hold 3.0 text from earlier edits;
-[CC BY-SA 3.0](https://creativecommons.org/licenses/by-sa/3.0/legalcode) §4(b)(ii) lets such an adaptation be licensed under "a later
-version of this License with the same License Elements", so it is labelled 4.0. The Terms
-of Use give the day of the switch, not the time. The cutoff is midnight UTC, so the
-184 pages last edited on 2023-06-07 itself are labelled 4.0. If the switch came later
-that day, those last edited before it are 3.0 text under a 4.0 label; a 2023-06-08 cutoff
-(the rights audit's suggestion) would instead put that day's 4.0 edits under a 3.0 label,
-which [4.0](https://creativecommons.org/licenses/by-sa/4.0/legalcode) §3(b)(1) ("this version or later") does not allow. Neither choice is
-exact without the time of the switch. The rule is `license_for_revision` in `src/fetch_wikipedia.py`,
-tested at the cutoff second, across time zones, and on a missing timestamp, which stops the
-build. v0.2.0-v0.2.5 labelled every row `CC-BY-SA-3.0`.
+(`rightsinfo`) links BY-SA 4.0.
+Each contribution keeps the licence it was made under: `CC-BY-SA-3.0` before 2023-06-07, `CC-BY-SA-4.0` from that day.
+Each row is labelled by the contributions its text holds:
+
+| `license` | documents | tokens | share of tokens |
+|---|---:|---:|---:|
+| `CC-BY-SA-3.0` | 719,119 | 364,335,667 | 29.7% |
+| `CC-BY-SA-3.0 AND CC-BY-SA-4.0` | 629,454 | 765,350,761 | 62.4% |
+| `CC-BY-SA-4.0` | 120,074 | 97,219,687 | 7.9% |
+
+A row labelled `CC-BY-SA-3.0 AND CC-BY-SA-4.0` holds text under each licence. It is not split by contribution;
+its page history tells the two apart. The basis of each label is the sidecar's `license_basis`:
+
+| `license_basis` | `license` | documents | rule |
+|---|---|---:|---|
+| `shipped_before_cutoff` | `CC-BY-SA-3.0` | 190,836 | the shipped revision predates the cutoff |
+| `same_revision_text` | `CC-BY-SA-3.0` | 5,028 | the shipped revision has the sha1 of the last pre-cutoff one in the stub dump (e.g. a revert) |
+| `identical_text` | `CC-BY-SA-3.0` | 492,757 | the last pre-cutoff revision converts to the shipped text, whitespace aside |
+| `no_text_added` | `CC-BY-SA-3.0` | 30,498 | later edits only removed words: the shipped words are a subsequence of the pre-cutoff ones |
+| `created_after_cutoff` | `CC-BY-SA-4.0` | 120,074 | the page has no revision before the cutoff |
+| `changed_after_cutoff` | `CC-BY-SA-3.0 AND CC-BY-SA-4.0` | 629,338 | later edits added at least one word |
+| `page_not_in_dump` | `CC-BY-SA-3.0 AND CC-BY-SA-4.0` | 85 | the page is not in the stub dump (deleted since), so its history is not known |
+| `pre_text_hidden` | `CC-BY-SA-3.0 AND CC-BY-SA-4.0` | 31 | the API returns no text for the last pre-cutoff revision (hidden or missing) |
+| `pre_text_unparseable` | `CC-BY-SA-3.0 AND CC-BY-SA-4.0` | 0 | the last pre-cutoff revision does not convert |
+
+The page histories come from the stub dump `plwiki-20261001-stub-meta-history.xml.gz` (7,039,541,375 bytes, sha1 `47fd7549026521bd02c174ee418c63a93f7e4118`,
+matched against its `dumpstatus.json`; revision metadata, no text). For each row shipped from a revision made on or
+after 2023-06-07, the text of the page's last revision before that day (`pre_cutoff_revid`) was fetched from the
+MediaWiki API, converted with the same `strip_article` as the release and compared with the shipped text, whitespace
+folded. Where they differ, a word diff (Python `difflib`) counts the words the later edits added; any added
+word, however few, makes the row `CC-BY-SA-3.0 AND CC-BY-SA-4.0`. Text moved into a page created after the cutoff from an
+older page is not traced, so that page is `CC-BY-SA-4.0` alone. `src/wiki_history.py` (`download`, `scan`, `fetch`,
+`compare`) does this; `src/fetch_wikipedia.py relabel` writes the labels.
+
+The Terms of Use give the day of the switch, not the time. The cutoff is midnight UTC, so revisions made on
+2023-06-07 itself count as `CC-BY-SA-4.0`. Pages labelled `CC-BY-SA-4.0` alone and created that day: 76.
+If the switch came later that day, their first edits may be 3.0 text under a 4.0 label.
+v0.2.0-v0.2.5 labelled every row `CC-BY-SA-3.0`; before this relabel, main labelled each row by its page's last
+revision alone (`CC-BY-SA-4.0` from 2023-06-07).
 
 Text pasted into an article from a copyrighted source is under neither licence. The
 fetcher drops the 35 pages that carry plwiki's copyright-violation
@@ -47,9 +67,10 @@ template's codes `npa`, `plagiat`, `copypaste`, `kopiuj-wklej`, or free text wit
 scan found 3 inside HTML comments, which MediaWiki does not show and the build does not
 count. A paste without either notice cannot be detected by the build.
 
-To measure undetected pastes, 40 released pages per label (seed
+To measure undetected pastes, 40 released pages per stratum (seed
 20261006, page-id order) were run through Earwig's Copyvio Detector (web search and
-the page's external links) on 2026-10-05. Every page whose best match reached
+the page's external links) on 2026-10-05. The strata are the labels by last revision the release
+carried then: `CC-BY-SA-3.0` for a page last edited before 2023-06-07, `CC-BY-SA-4.0` from that day. Every page whose best match reached
 confidence 0.10 or a "possible"/"suspected" verdict was compared by hand
 with the page history and the date of the matching text (see `wikipedia.copyvio-sample.jsonl`):
 
@@ -66,11 +87,19 @@ why each is classed so). Both are dropped from this release: `COPIED_PAGES` in
 `src/fetch_wikipedia.py` lists them by page id with the reason, and the build stops if one is
 not found. The sample was drawn before that drop, from the 1,468,656 documents of Hub PR 68
 commit 3393e05; the test reproduces the draw from the release plus the dropped rows. The share of such pages is below 12.9% (`CC-BY-SA-3.0`) and
-12.9% (`CC-BY-SA-4.0`), 95% Wilson upper bounds; the labels hold
+12.9% (`CC-BY-SA-4.0`), 95% Wilson upper bounds; the strata hold
 6.6% and 93.4% of the tokens. The bounds cover only
 what the detector can find: print sources, translations and paraphrase scoring below the
 hand-check threshold cannot match.
 
+Pages deleted from plwiki for a copyright violation after the dump are dropped too. Of the released
+pages, 104 are missing from the stub dump `plwiki-20261001-stub-meta-history.xml.gz`; their deletion logs, read
+on 2026-10-18, give a copyright violation as the reason for 7. `COPIED_PAGES` lists them with the log's
+reason, and `fetch_wikipedia.py withhold` dropped them from this release as a rebuild would, adding them
+to the sample's dropped rows. Of the other 97, 88 were deleted after a deletion discussion,
+7 as duplicates, needless disambiguations, advertising, spam or nonsense, and 2 were deleted
+for a history merge or a move and restored. They are kept: a quality call, not a rights one.
+
 Evidence: `wikipedia.license-evidence.json`.
 
 ## Attribution
@@ -78,7 +107,9 @@ Each row's `author` is `Wikipedia contributors, <page URL>`, e.g.
 `Wikipedia contributors, https://pl.wikipedia.org/wiki/Warszawa`. The page history names the contributors; the Terms of Use
 (section 7) accept a link or URL to the page as attribution.
 `wikipedia.attribution.jsonl` has one line per row: `id`, `pageid`, `revid` (the revision
-in the dump), `title`, `url`, `last_revision_at`, `license`. The row `id` is
+in the dump), `title`, `url`, `last_revision_at`, `license`,
+`license_basis` (how the label was decided, see Licensing) and `pre_cutoff_revid` (the page's last revision
+before the cutoff, null for a page created after it or edited after it and since deleted; its text is at `https://pl.wikipedia.org/w/index.php?oldid=<pre_cutoff_revid>`). The row `id` is
 `wikipedia_<page id>` and `created` is the date of `last_revision_at`.
 
 The text is modified: wiki markup is converted to plain text (see Provenance).
@@ -102,19 +133,23 @@ pages with the same text (sha1 of the stripped text), the lower page id is kept.
 2,877,278 pages scanned → 2,304,425 in namespace 0 → 598,071 redirects,
 0 empty pages and 35 with a copyright-violation notice skipped
 (0 parser failures) → 1,706,319 converted →
-382 exact duplicates, 2 copied pages (`COPIED_PAGES`) and 0 copyright `{{ek}}`
-pages dropped → 1,705,935 read by the build →
-1,468,654 released.
+382 exact duplicates, 9 copied pages (`COPIED_PAGES`: 2 found by the copyright sample,
+7 deleted from plwiki for a copyright violation after the dump) and 0 copyright `{{ek}}`
+pages dropped → 1,705,928 read by the build →
+1,468,647 released.
 
 Not cleaned: appendix headings ("Zobacz też", "Uwagi", "Przypisy", "Bibliografia", "Linki
-zewnętrzne") stay in 828,388 documents (56.405%), with what is left of those sections
+zewnętrzne") stay in 828,386 documents (56.405%), with what is left of those sections
 after the conversion (2.96% of all characters from the first such heading to the end);
-174,668 documents (11.893%) have whitespace-only lines.
+174,668 documents (11.893%) have whitespace-only lines. Malformed links leave markup: 14
+documents hold a category link on a line of its own (`wikipedia_103834`), 13 an unclosed file
+link (`[[Plik:...|thumb|...`, `wikipedia_13010`); 363 contain `<ref`. Other lines starting
+`Kategoria:` are content, mostly flattened medal tables (`Kategoria:Złoto:Srebro:Brąz:...`).
 
 ## Statistics
 | documents | characters | tokens (tiktoken proxy) |
 |---:|---:|---:|
-| 1,468,654 | 3,140,838,409 | 1,226,912,191 |
+| 1,468,647 | 3,140,823,417 | 1,226,906,115 |
 
 v0.2.5 (SpeakLeash snapshot) had 1,171,897 documents, 1,844,707,761 characters and
 707,194,207 tokens. The SpeakLeash extraction is not documented. Its text drops the
@@ -125,30 +160,38 @@ in v0.2.5 and 59,133 here.
 ## Per-document license metadata
 | license | documents |
 |---|---:|
-| `CC-BY-SA-3.0` | 190,836 |
-| `CC-BY-SA-4.0` | 1,277,818 |
+| `CC-BY-SA-3.0` | 719,119 |
+| `CC-BY-SA-3.0 AND CC-BY-SA-4.0` | 629,454 |
+| `CC-BY-SA-4.0` | 120,074 |
 
-Author metadata present for **1,468,654** documents.
+Author metadata present for **1,468,647** documents.
 
 ## Filters applied (build_dynaword.py)
 - drop documents < 200 chars: **231,717**
 - drop non-Polish (diacritic ratio): **5,564**
 - exact cross-source dedup (sha1): **0** (the fetcher already dropped 382)
 - OCR alpha-ratio < 0.70 (OCR sources only): **0**
-- read 1,705,935 → kept 1,468,654
+- read 1,705,928 → kept 1,468,647
 
 Token counts are a fast tiktoken (cl100k) proxy (~1% off Llama-3); the canonical
 Llama-3 count is computed at release.
 
 ## Not checked
 - pastes from copyrighted sources beyond the sample above
-- the 3.0 → 4.0 label of pages edited after the cutoff (a reading of 3.0 §4(b)(ii), not a
-  lawyer's opinion), and the time of day of the switch (184 pages last edited on 2023-06-07)
+- whether a small post-cutoff edit is a contribution of its own: any added word makes the row
+  `CC-BY-SA-3.0 AND CC-BY-SA-4.0`, with no threshold
+- text moved or merged from an older page into a page created after the cutoff (labelled `CC-BY-SA-4.0` alone)
+- the time of day of the switch (see Licensing)
 - text imported from elsewhere under another licence or with extra attribution
   requirements (Terms of Use, Re-use); notices kept in templates are removed by the conversion
 - text first contributed under GFDL before the 2009 move to CC BY-SA, and GFDL dual
   licensing; only the CC BY-SA grant is relied on
 - quotations inside articles (Art. 29 of the Polish copyright act)
 - deletion discussions and maintenance notices other than `{{NPA}}` and `{{ek}}`; their
-  reasons are not read
+  reasons are not read, except the deletion logs above
+- pages deleted from plwiki after the stub dump (2026-10-01), and other pages by the authors
+  of the 7 deleted for copyright
+- "Teatr Piosenki Młyn" (`wikipedia_4033410`), deleted for copyright on 2017-05-23 and recreated
+  by 2017-07-14: the released text was not compared with the deleted one (it is among the
+  88 deleted after a discussion in 2026)
 - personal data (the NERGAL scrub runs before release)

data/wikipedia/wikipedia.parquet

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:f4abfff711b207691d1278efc28d8af18551454143f3d8f2292ec3ddcefc314a
-size 1462846393
+oid sha256:6fec912de221285b1d3de1e813b66aca9ebfb62782e64a73bf48783e8a013fa9
+size 1462886095

data/wikipedia/wikipedia.stats.json

@@ -1,17 +1,18 @@
 {
-  "read": 1705935,
-  "kept": 1468654,
+  "read": 1705928,
+  "kept": 1468647,
   "drop_short": 231717,
   "drop_lang": 5564,
   "drop_dup": 0,
   "drop_ocr": 0,
-  "chars": 3140838409,
-  "tokens": 1226912191,
+  "chars": 3140823417,
+  "tokens": 1226906115,
   "licenses": {
-    "CC-BY-SA-4.0": 1277818,
-    "CC-BY-SA-3.0": 190836
+    "CC-BY-SA-3.0": 719119,
+    "CC-BY-SA-3.0 AND CC-BY-SA-4.0": 629454,
+    "CC-BY-SA-4.0": 120074
   },
-  "authors_with_value": 1468654,
+  "authors_with_value": 1468647,
   "secs": 34.6,
-  "license": "CC BY-SA (3.0 or 4.0 per record, by last-revision date)"
+  "license": "CC BY-SA (3.0, 4.0 or both per record, by contribution date)"
 }

src/fetch_wikimedia.py

@@ -453,9 +453,23 @@ def relabel(out: Path, data_dir: Path, key: str) -> None:
     `license` column, the sidecar (wiki_history.apply_labels), stats `licenses` and `license`,
     build.json (`attribution_sha256`, `licence_labels`) and the licence evidence's labels and tokens.
     Text, ids and every other column are unchanged; rerunning changes nothing."""
+    site = SITES[key]
+    docs, tokens = write_labels(out, data_dir, key, site["wiki"], site["cutoff"])
+    path = data_dir / f"{key}.license-evidence.json"
+    ev = load(path)
+    ev["license"] = f"{site['before']}, {site['after']} or both per record, by contribution date"
+    ev["review_scope"] = f"licence of contributors' text on {site['host']} and its version per contribution; not independent legal advice"
+    ev["release_scope"].update(labels=docs, tokens=tokens)
+    save(path, ev)
+    print(f"{key}: relabelled, {docs}")
+
+
+def write_labels(out: Path, data_dir: Path, key: str, wiki: str, cutoff) -> tuple[dict, dict]:
+    """relabel's rewrite of the parquet, sidecar, stats and build record, shared with
+    fetch_wikipedia.py relabel; `wiki` names the stub dump, `cutoff` is the licence switch.
+    Returns (documents, tokens) per label."""
     import pyarrow.parquet as pq
     from wiki_history import STUB_DATE, apply_labels
-    site = SITES[key]
 
     def path(name):
         return data_dir / f"{key}.{name}"
@@ -470,21 +484,16 @@ def relabel(out: Path, data_dir: Path, key: str) -> None:
     stats = load(path("stats.json"))
     stats.update(licenses=docs, license=SOURCES[key]["license"])
     path("stats.json").write_text(json.dumps(stats, indent=2))  # build_dynaword's format
-    day = site["cutoff"].date().isoformat()
+    day = cutoff.date().isoformat()
     build = load(path("build.json"))
     build["attribution_sha256"] = digest(path("attribution.jsonl").read_bytes())
     build["licence_labels"] = {
-        "stub": load(out / f"{site['wiki']}-{STUB_DATE}-stub-meta-history.xml.gz.json"),
+        "stub": load(out / f"{wiki}-{STUB_DATE}-stub-meta-history.xml.gz.json"),
         "cases": dict(Counter(lab["case"] for lab in labels)),
         "created_on_cutoff_day": sum(lab["case"] == "created_after_cutoff" and lab["first_at"].startswith(day) for lab in labels),
     }
     save(path("build.json"), build)
-    ev = load(path("license-evidence.json"))
-    ev["license"] = f"{site['before']}, {site['after']} or both per record, by contribution date"
-    ev["review_scope"] = f"licence of contributors' text on {site['host']} and its version per contribution; not independent legal advice"
-    ev["release_scope"].update(labels=docs, tokens=tokens)
-    save(path("license-evidence.json"), ev)
-    print(f"{key}: relabelled, {docs}")
+    return docs, tokens
 
 
 def share(part: int, whole: int) -> str:
@@ -528,12 +537,14 @@ def licensing_text(key: str, ev: dict) -> str:
 
 
 def contribution_text(key: str, build: dict, ev: dict, tok: int) -> str:
-    """The labels of a release relabelled by contribution, and how they were decided."""
-    from wiki_history import CASES
-    site, lab, scope, stub = SITES[key], build["licence_labels"], ev["release_scope"], build["licence_labels"]["stub"]
-    if "SA" not in site["after"]:  # wikinews: 2.5 and the known switch time need their own text (v2#110)
+    """The labels of a release relabelled by contribution, and how they were decided (also in
+    data/wikipedia/wikipedia.md, which is written by hand)."""
+    from wiki_history import CASES, CONVERTERS, WIKIS
+    cfg, lab, scope, stub = WIKIS[key], build["licence_labels"], ev["release_scope"], build["licence_labels"]["stub"]
+    if "SA" not in cfg["after"]:  # wikinews: 2.5 and the known switch time need their own text (v2#110)
         raise NotImplementedError(f"{key}: no card text for contribution labels yet")
-    before, after, day = site["before"], site["after"], site["cutoff"].date().isoformat()
+    before, after, day = cfg["before"], cfg["after"], cfg["cutoff"][:10]
+    module, convert = CONVERTERS[key]
     licence = {"before": before, "after": after, "mixed": f"{before} AND {after}"}
     labels = "\n".join(f"| `{lic}` | {scope['labels'].get(lic, 0):,} | {scope['tokens'].get(lic, 0):,} | {share(scope['tokens'].get(lic, 0), tok)} |"
                        for lic in (before, licence["mixed"], after))
@@ -555,11 +566,11 @@ its page history tells the two apart. The basis of each label is the sidecar's `
 The page histories come from the stub dump `{stub['name']}` ({stub['size']:,} bytes, sha1 `{stub['sha1']}`,
 matched against its `dumpstatus.json`; revision metadata, no text). For each row shipped from a revision made on or
 after {day}, the text of the page's last revision before that day (`pre_cutoff_revid`) was fetched from the
-MediaWiki API, converted with the same `strip` as the release and compared with the shipped text, whitespace
+MediaWiki API, converted with the same `{convert}` as the release and compared with the shipped text, whitespace
 folded. Where they differ, a word diff (Python `difflib`) counts the words the later edits added; any added
 word, however few, makes the row `{licence['mixed']}`. Text moved into a page created after the cutoff from an
 older page is not traced, so that page is `{after}` alone. `src/wiki_history.py` (`download`, `scan`, `fetch`,
-`compare`) does this; `src/fetch_wikimedia.py relabel` writes the labels.
+`compare`) does this; `src/{module}.py relabel` writes the labels.
 
 The Terms of Use give the day of the switch, not the time. The cutoff is midnight UTC, so revisions made on
 {day} itself count as `{after}`. Pages labelled `{after}` alone and created that day: {lab['created_on_cutoff_day']}.

src/fetch_wikipedia.py

@@ -13,6 +13,18 @@ checked against the sha1 in dumpstatus.json and in the dump's sha1sums.txt.
                                                          # COPIED_PAGES, copyright {{ek}} pages
     python3 src/build_dynaword.py --sources wikipedia --speakleash-dir $B --out .
     python3 src/fetch_wikipedia.py attribution --out $B  # data/wikipedia sidecars
+    python3 src/fetch_wikipedia.py withhold    --out $B  # drops pages added to COPIED_PAGES since
+
+Labels by contribution (slayerlabs/polish-dynaword-v2#110, option B) replace the last-revision
+labels on the committed release; the pages-articles dump is not re-read. H=<history dir, never committed>
+    python3 src/wiki_history.py download --dump plwiki --out $H
+    python3 src/wiki_history.py scan --stub $H/plwiki-20261001-stub-meta-history.xml.gz \\
+        --sidecar wikipedia=data/wikipedia/wikipedia.attribution.jsonl --out $H
+    python3 src/wiki_history.py fetch --source wikipedia --history $H/wikipedia.history.jsonl --out $H
+    python3 src/wiki_history.py compare --source wikipedia --sidecar data/wikipedia/wikipedia.attribution.jsonl \\
+        --history $H/wikipedia.history.jsonl --fetched $H/wikipedia --parquet data/wikipedia/wikipedia.parquet \\
+        --out $H/wikipedia.labels.jsonl
+    python3 src/fetch_wikipedia.py relabel --out $H  # data/wikipedia; the card is edited by hand
 """
 from __future__ import annotations
 
@@ -59,15 +71,33 @@ EK_RE = re.compile(r"\{\{\s*(?:(?i:szablon|template)\s*:\s*)?"
 EK_NAMES = {"Ek", "EK", "Ekspresowe kasowanko", "Usuń", "Speedy", "Delete"}
 # The build records every template redirect to these, so the alias lists above can be checked against the dump.
 TAG_TEMPLATES = {"Szablon:NPA", "Szablon:Ek"}
-# Pages the copyright sample (data/wikipedia/wikipedia.copyvio-sample.jsonl) found to hold text
-# copied or closely paraphrased from a copyrighted source. Keyed on the wiki page id, which is
-# stable across dumps (the dataset id is derived from it). prepare fails if one is not reached.
+
+
+def _deleted(day: str, reason: str = "[[Wikipedia:Prawa autorskie|naruszenie praw autorskich]]") -> str:
+    """A COPIED_PAGES reason for a page deleted from plwiki for copyright: the day and the log's reason."""
+    return f'deleted from plwiki on {day} for a copyright violation, deletion log: "{reason}"'
+
+
+# Pages that hold text copied or closely paraphrased from a copyrighted source: found by the
+# copyright sample (data/wikipedia/wikipedia.copyvio-sample.jsonl), or deleted from plwiki for a
+# copyright violation after the dump (Special:Log/delete). Keyed on the wiki page id, which is
+# stable across dumps (the dataset id is derived from it). prepare fails if one is not reached;
+# `withhold` drops one added after the build from the committed release.
 COPIED_PAGES = {
     833396: "Obnażyć prawdę: the 'Opis fabuły' paragraph closely paraphrases Filmweb's production "
             "description of the film; added in revision 15302011 (2008-12-26), which cites Filmweb",
     1126780: "Czarnka: the lake paragraph (revision 42979631, 2015-06-19) closely paraphrases the "
              "Drawieński Park Narodowy page dpn.pl/czarne, archived 2012-01-25; whether Art. 4 of the "
              "Polish copyright act (official materials) applies was not assessed",
+    4095632: "Brzezina (opowiadanie): " + _deleted("2026-09-01", (  # the link's tracking query left out
+        "Naruszenie praw autorskich: https://www.studocu.com/pl/document/uniwersytet-jagiellonski/"
+        "filologia-polska/proza-poetycka-i-awangardowa-iwaszkiewicz-jasienski-choromanski/74568413")),
+    6131122: "Edmund Bielecki: " + _deleted("2026-09-05"),
+    6131594: "Ryszard Cyffer: " + _deleted("2026-09-05"),
+    6131631: "Stanisław Czarnomski: " + _deleted("2026-09-05"),
+    6131646: "Mikołaj Dmochowski: " + _deleted("2026-09-05"),
+    6132108: "Włodzimierz Daszkiewicz: " + _deleted("2026-09-06"),
+    6132151: "Konstanty Jankowski (oficer): " + _deleted("2026-09-06"),
 }
 # {{ek}} reasons that state a copyright violation: the template's own codes that render as
 # "naruszenie praw autorskich" (npa, plagiat, copypaste, kopiuj-wklej) and free text naming one
@@ -352,9 +382,80 @@ def attribution(out: Path, data_dir: Path) -> None:
     print(f"attribution rows: {n}")
 
 
+def withhold(data_dir: Path) -> None:
+    """Drop from the committed release the pages added to COPIED_PAGES after the build, as prepare
+    would on a rebuild, without re-reading the dump: the parquet and sidecar rows, stats, build.json
+    (`prepare.counts` and `copied_pages`, attribution rows and sha) and the evidence's release scope.
+    Each is added to the copyright sample's `rows_dropped_since` with its stratum, so the sample's
+    draw still reproduces. Run before `relabel`; rerunning changes nothing."""
+    import pyarrow as pa
+    import pyarrow.compute as pc
+    import pyarrow.parquet as pq
+    from fetch_wikimedia import label_totals  # not at the top: fetch_wikimedia imports this module
+
+    def path(name):
+        return data_dir / f"{SOURCE}.{name}"
+
+    side = [json.loads(line) for line in path("attribution.jsonl").open(encoding="utf-8")]
+    held = [rec for rec in side if rec["pageid"] in COPIED_PAGES]
+    if not held:
+        print(f"{SOURCE}: nothing to withhold")
+        return
+    build = load(path("build.json"))
+    if "licence_labels" in build:  # ponytail: relabel's case counts would keep the withheld rows
+        raise ValueError(f"{SOURCE}: withhold before relabel, then filter the labels to the sidecar")
+    held_ids = {rec["id"] for rec in held}
+    table = pq.read_table(path("parquet"))
+    table = table.filter(pc.invert(pc.is_in(table["id"], pa.array(sorted(held_ids)))))
+    groups = pq.ParquetFile(path("parquet")).metadata.row_group(0).num_rows  # build_dynaword's batches, kept
+    pq.write_table(table, path("parquet"), compression="zstd", row_group_size=groups)
+    side = [rec for rec in side if rec["id"] not in held_ids]
+    path("attribution.jsonl").write_text("".join(json.dumps(r, ensure_ascii=False) + "\n" for r in side), encoding="utf-8")
+
+    docs, tokens = label_totals(table["license"].to_pylist(), table["token_count"].to_pylist())
+    stats = load(path("stats.json"))
+    stats.update(read=stats["read"] - len(held), kept=table.num_rows, chars=pc.sum(pc.utf8_length(table["text"])).as_py(),
+                 tokens=sum(tokens.values()), licenses=docs, authors_with_value=table.num_rows)
+    path("stats.json").write_text(json.dumps(stats, indent=2))  # build_dynaword's format
+    prep = build["prepare"]
+    prep["counts"]["copied_page"] += len(held)
+    for rec in held:  # prepare's per-row label is the last revision's
+        prep["counts"][license_for_revision(rec["last_revision_at"])] -= 1
+    prep["copied_pages"] += [{"pageid": str(rec["pageid"]), "title": rec["title"], "reason": COPIED_PAGES[rec["pageid"]]}
+                             for rec in held]
+    build.update(attribution_rows=len(side), attribution_sha256=digest(path("attribution.jsonl").read_bytes()))
+    save(path("build.json"), build)
+    ev = load(path("license-evidence.json"))
+    scope = ev["release_scope"]
+    scope.update(labels=docs, tokens_by_label=tokens, copied_pages=len(prep["copied_pages"]), documents_released=table.num_rows)
+    ev["copyright_sample"]["population"]["rows_dropped_since"] += [
+        [rec["id"], license_for_revision(rec["last_revision_at"])] for rec in held]
+    path("license-evidence.json").write_text(json.dumps(ev, ensure_ascii=False, indent=1) + "\n", encoding="utf-8")  # the file's format
+    print(f"{SOURCE}: withheld {sorted(held_ids)}, {stats['kept']} rows left")
+
+
+def relabel(out: Path, data_dir: Path) -> None:
+    """Label each row by contribution (v2#110) from out/wikipedia.labels.jsonl (wiki_history.py
+    compare): fetch_wikimedia.write_labels rewrites the parquet `license` column, the sidecar,
+    stats and build record; this updates the licence evidence's labels and tokens. Text, ids and
+    every other column are unchanged; rerunning changes nothing."""
+    from fetch_wikimedia import write_labels  # not at the top: fetch_wikimedia imports this module
+    docs, tokens = write_labels(out, data_dir, SOURCE, "plwiki", BY_SA_4_CUTOFF)
+    path = data_dir / f"{SOURCE}.license-evidence.json"
+    ev = load(path)
+    ev["license"] = "CC-BY-SA-3.0, CC-BY-SA-4.0 or both per record, by contribution date"
+    ev["review_scope"] = ("licence of contributors' text on pl.wikipedia and its version per contribution; "
+                          "copyright pastes estimated by sample; not independent legal advice")
+    scope = ev["release_scope"]
+    scope.pop("tokens_by_label", None)  # the wikimedia sources' name, so their card text applies
+    scope.update(labels=docs, tokens=tokens)
+    path.write_text(json.dumps(ev, ensure_ascii=False, indent=1) + "\n", encoding="utf-8")  # the file's format
+    print(f"{SOURCE}: relabelled, {docs}")
+
+
 def main() -> None:
     ap = argparse.ArgumentParser()
-    ap.add_argument("step", choices=["discover", "acquire", "extract", "prepare", "attribution"])
+    ap.add_argument("step", choices=["discover", "acquire", "extract", "prepare", "attribution", "withhold", "relabel"])
     ap.add_argument("--out", required=True, type=Path)
     ap.add_argument("--data-dir", type=Path, default=Path(__file__).resolve().parent.parent / "data" / SOURCE)
     ap.add_argument("--workers", type=int, default=8)
@@ -362,7 +463,8 @@ def main() -> None:
     a.out.mkdir(parents=True, exist_ok=True)
     {"discover": lambda: discover(a.out), "acquire": lambda: acquire(a.out),
      "extract": lambda: extract(a.out, a.workers), "prepare": lambda: prepare(a.out),
-     "attribution": lambda: attribution(a.out, a.data_dir)}[a.step]()
+     "attribution": lambda: attribution(a.out, a.data_dir), "withhold": lambda: withhold(a.data_dir),
+     "relabel": lambda: relabel(a.out, a.data_dir)}[a.step]()
 
 
 if __name__ == "__main__":

src/sources.py

@@ -131,24 +131,31 @@ SOURCES = {
                       "dump's sha1sums.txt) by src/fetch_wikipedia.py. Articles (namespace 0) "
                       "only, no redirects; one document per page, id from the page id. "
                       "v0.2.0-v0.2.5 used the SpeakLeash plwiki snapshot; that intermediary "
-                      "was removed in the v0.2.6 rebuild.",
+                      "was removed in the v0.2.6 rebuild. The licence labels come from the "
+                      "page histories (stub-meta-history dump plwiki-20261001, "
+                      "src/wiki_history.py).",
         "legal_note": "Pages carrying the plwiki copyright-violation notice {{NPA}} (or its "
                       "redirects {{Copyvio}}, {{Plagiat}}) in the dump are dropped, as are "
                       "pages whose speedy-deletion notice {{ek}} gives a copyright reason and "
-                      "the pages the copyright sample found copied (COPIED_PAGES in "
+                      "the pages found to hold copied text, by the copyright sample or by a "
+                      "plwiki deletion for copyright after the dump (COPIED_PAGES in "
                       "src/fetch_wikipedia.py). Other text pasted into an article cannot be "
                       "detected; the stratified web sample is recorded in "
                       "wikipedia.license-evidence.json.",
-        "license": "CC BY-SA (3.0 or 4.0 per record, by last-revision date)",
+        "license": "CC BY-SA (3.0, 4.0 or both per record, by contribution date)",
         "license_spdx": "MIXED",
-        "traceable": "Contributors license their text under CC BY-SA 4.0 since the Wikimedia "
-                     "Terms of Use of 2023-06-07, under CC BY-SA 3.0 before. A page whose "
-                     "last revision predates 2023-06-07 is labelled CC-BY-SA-3.0; a page "
-                     "revised on or after that day is an adaptation of the 3.0 text, which "
-                     "3.0 §4(b)(ii) allows under 4.0, and is labelled CC-BY-SA-4.0. Each "
-                     "row's author field links its page, whose history names the "
-                     "contributors; wikipedia.attribution.jsonl gives page id, revision id, "
-                     "title and last-revision time per row.",
+        "traceable": "Contributors license their text under CC BY-SA 4.0 since the "
+                     "Wikimedia Terms of Use of 2023-06-07, under CC BY-SA 3.0 before, "
+                     "and each contribution keeps the licence it was made under. A row "
+                     "whose text was all contributed before 2023-06-07 is labelled "
+                     "CC-BY-SA-3.0, a page created on or after that day CC-BY-SA-4.0, "
+                     "and a row holding text from both sides, or whose text before that "
+                     "day cannot be compared (a hidden revision or a page deleted "
+                     "since), 'CC-BY-SA-3.0 AND CC-BY-SA-4.0'. Each row's author field "
+                     "links its page, whose history names the contributors; "
+                     "wikipedia.attribution.jsonl gives page id, revision id, title, "
+                     "last-revision time, the basis of the label and the page's last "
+                     "revision before 2023-06-07 per row.",
         "upstream": "https://dumps.wikimedia.org/plwiki/20260901/",
         "domain": "encyclopedic",
         "created": "2010-11-12, 2026-09-01",

src/test_wiki_history.py

@@ -247,6 +247,12 @@ def test_compare(tmp_path):
     assert out[3]["first_at"] == "2025-01-01T00:00:00Z" and out[2]["first_at"] is None  # page 2 is not in the dump
     # page 1 was deleted after the build: its shipped revision is still its last before the cutoff
     assert (out[1]["pre_revid"], out[1]["pre_at"]) == (19, "2020-01-01T00:00:00Z") and out[2]["pre_revid"] is None
+    # sidecar slices, run apart and joined in order, give the same records
+    lines, joined = side.read_text().splitlines(keepends=True), []
+    for k, part in enumerate((lines[:5], lines[5:])):
+        (piece := tmp_path / f"side{k}.jsonl").write_text("".join(part))
+        joined += wh.compare("wikinews", piece, hist, fetched_dir, parquet, convert)
+    assert joined == list(out.values())
     recs.pop(80)
     (fetched_dir / "00000.jsonl.zst").unlink()
     wh.fetch("wikinews", [(0, r) for r in sorted(recs)], tmp_path, session=API(), log=lambda m: None)
@@ -281,6 +287,25 @@ def test_download_checks_a_present_file(tmp_path):
         wh.download("plwikivoyage", tmp_path, Status("0" * 40, "in-progress"))
 
 
+def test_download_takes_a_big_wiki_file_from_the_recombine_job(tmp_path):
+    body = b"stub"
+    name = "plwiki-20261001-stub-meta-history.xml.gz"
+    (tmp_path / name).write_bytes(body)
+
+    class Status:
+        def __init__(self, state):
+            self.state = state
+
+        def get(self, url, headers, timeout):
+            return FakeResp({"jobs": {
+                "xmlstubsdump": {"status": "done", "files": {"plwiki-20261001-stub-meta-history1.xml.gz": {}}},
+                "xmlstubsdumprecombine": {"status": self.state, "files": {name: {"size": 4, "sha1": hashlib.sha1(body).hexdigest()}}}}})
+
+    assert wh.download("plwiki", tmp_path, Status("done"))["name"] == name
+    with pytest.raises(ValueError, match="plwiki xmlstubsdumprecombine not done"):
+        wh.download("plwiki", tmp_path, Status("in-progress"))
+
+
 def test_apply_labels():
     import pyarrow as pa
     side = [{"id": f"w_{i}", "last_revision_at": f"2025-01-0{i}T00:00:00Z", "license": "old"} for i in (1, 2)]

src/test_wikimedia_contract.py

@@ -526,6 +526,22 @@ def test_attribution_fails_when_the_parquet_has_a_row_the_dump_does_not(tmp_path
 
 # --- the contract on a released source -----------------------------------------------------
 
+def check_contribution_labels(key, cols, meta, build):
+    """The labels by contribution (v2#110) on a released source relabelled from wiki_history.compare:
+    each row's label follows its recorded basis, and the build record counts the bases."""
+    cfg, labels = wh.WIKIS[key], build["licence_labels"]
+    licence = {"before": cfg["before"], "after": cfg["after"], "mixed": f"{cfg['before']} AND {cfg['after']}"}
+    for row_id, label in zip(cols["id"], cols["license"]):
+        m = meta[row_id]
+        basis = m["license_basis"]
+        assert label == m["license"] == licence[wh.CASES[basis]]
+        assert (basis == "shipped_before_cutoff") == (m["last_revision_at"] < cfg["cutoff"])
+        assert (m["pre_cutoff_revid"] is None) == (basis in ("created_after_cutoff", "page_not_in_dump"))
+        assert basis != "shipped_before_cutoff" or m["pre_cutoff_revid"] == int(m["revid"])
+    assert labels["cases"] == dict(Counter(m["license_basis"] for m in meta.values()))
+    assert labels["stub"]["name"] == f"{cfg['dump']}-{wh.STUB_DATE}-stub-meta-history.xml.gz" and SHA1.fullmatch(labels["stub"]["sha1"])
+
+
 def check_released(key, tmp_path, dropped=(0, 0, 0)):
     """Checks that hold for any source built by fetch_wikimedia.py; counts specific to a source live in its own test file.
     `dropped` is (documents, characters, tokens) of the v0.2.5 rows that DROPPED_PAGES[key] withholds."""
@@ -575,16 +591,11 @@ def check_released(key, tmp_path, dropped=(0, 0, 0)):
     # every row's date, author and id follow the page's last revision; its label too, until the
     # release is relabelled by contribution (relabel), then the basis recorded per row
     by_contribution = build.get("licence_labels")
-    licence = {"before": site["before"], "after": site["after"], "mixed": f"{site['before']} AND {site['after']}"}
+    if by_contribution:
+        check_contribution_labels(key, cols, meta, build)
     for row_id, created, label, author in zip(cols["id"], cols["created"], cols["license"], cols["author"]):
         m = meta[row_id]
-        if by_contribution:
-            basis = m["license_basis"]
-            assert label == m["license"] == licence[wh.CASES[basis]]
-            assert (basis == "shipped_before_cutoff") == (m["last_revision_at"] < wh.WIKIS[key]["cutoff"])
-            assert (m["pre_cutoff_revid"] is None) == (basis in ("created_after_cutoff", "page_not_in_dump"))
-            assert basis != "shipped_before_cutoff" or m["pre_cutoff_revid"] == int(m["revid"])
-        else:
+        if not by_contribution:
             assert label == m["license"] == fm.label(key, m["last_revision_at"]) and label in (site["before"], site["after"])
         assert created == m["last_revision_at"][:10]
         assert m["url"] == fm.page_url(key, m["title"]) and author == f"{site['author']}, {m['url']}"
@@ -594,10 +605,6 @@ def check_released(key, tmp_path, dropped=(0, 0, 0)):
     for label, tok in zip(cols["license"], cols["token_count"]):
         tokens[label] += tok
     assert stats["licenses"] == dict(Counter(cols["license"])) == scope["labels"] and scope["tokens"] == dict(tokens)
-    if by_contribution:
-        assert by_contribution["cases"] == dict(Counter(m["license_basis"] for m in meta.values()))
-        stub = by_contribution["stub"]
-        assert stub["name"] == f"{site['wiki']}-{wh.STUB_DATE}-stub-meta-history.xml.gz" and SHA1.fullmatch(stub["sha1"])
     assert scope["created_range"] == [min(cols["created"]), max(cols["created"])]
     day = site["cutoff"].date().isoformat()
     on_day = [m["last_revision_at"] for m in meta.values() if m["last_revision_at"].startswith(day)]

src/test_wikipedia_contract.py

@@ -13,8 +13,10 @@ import pyarrow as pa
 import pyarrow.parquet as pq
 import pytest
 
+import fetch_wikimedia as fm
 import fetch_wikipedia as fw
 from sources import SOURCES
+from test_wikimedia_contract import check_contribution_labels
 from verify_paper_claims import license_family
 
 ROOT = Path(__file__).resolve().parents[1]
@@ -195,13 +197,84 @@ def test_prepare_fails_when_a_listed_page_is_not_reached(tmp_path, monkeypatch):
         fw.prepare(tmp_path)
 
 
+def test_withhold_drops_pages_added_after_the_build(tmp_path, monkeypatch):
+    monkeypatch.setattr(fw, "COPIED_PAGES", {20: "Kret: copied", 30: "Mysz: deleted"})
+    rows = [("wikipedia_10", 10, "Ala ma kota.", "2010-01-01T00:00:00Z", 4),
+            ("wikipedia_30", 30, "Mysz é", "2024-01-01T00:00:00Z", 3),
+            ("wikipedia_40", 40, "Kret", "2025-01-01T00:00:00Z", 2)]
+    lic = [fw.license_for_revision(r[3]) for r in rows]
+    pq.write_table(pa.table({"id": [r[0] for r in rows], "text": [r[2] for r in rows], "license": lic,
+                             "token_count": [r[4] for r in rows]}), tmp_path / "wikipedia.parquet")
+    (tmp_path / "wikipedia.attribution.jsonl").write_text("".join(json.dumps(
+        {"id": i, "pageid": p, "title": t[:4], "last_revision_at": at, "license": label}) + "\n"
+        for (i, p, t, at, _), label in zip(rows, lic)))
+    (tmp_path / "wikipedia.stats.json").write_text(json.dumps(
+        {"read": 5, "kept": 3, "chars": 22, "tokens": 9, "licenses": {}, "authors_with_value": 3, "secs": 1}))
+    fw.save(tmp_path / "wikipedia.build.json", {"prepare": {
+        "counts": {"extracted": 6, "exact_duplicate": 0, "copied_page": 1, "ek_copyright": 0,
+                   "CC-BY-SA-3.0": 2, "CC-BY-SA-4.0": 3},
+        "copied_pages": [{"pageid": "20", "title": "Kret", "reason": "Kret: copied"}]}, "attribution_rows": 3})
+    fw.save(tmp_path / "wikipedia.license-evidence.json", {
+        "release_scope": {"labels": {}, "tokens_by_label": {}, "copied_pages": 1, "documents_released": 3},
+        "copyright_sample": {"population": {"documents": 4, "rows_dropped_since": [["wikipedia_20", "CC-BY-SA-3.0"]]}}})
+
+    fw.withhold(tmp_path)
+    assert pq.read_table(tmp_path / "wikipedia.parquet")["id"].to_pylist() == ["wikipedia_10", "wikipedia_40"]
+    assert [json.loads(line)["id"] for line in (tmp_path / "wikipedia.attribution.jsonl").read_text().splitlines()] == [
+        "wikipedia_10", "wikipedia_40"]
+    stats = json.loads((tmp_path / "wikipedia.stats.json").read_text())
+    assert stats == {"read": 4, "kept": 2, "chars": 16, "tokens": 6, "authors_with_value": 2, "secs": 1,
+                     "licenses": {"CC-BY-SA-3.0": 1, "CC-BY-SA-4.0": 1}}
+    build = fw.load(tmp_path / "wikipedia.build.json")
+    counts = build["prepare"]["counts"]
+    assert counts == {"extracted": 6, "exact_duplicate": 0, "copied_page": 2, "ek_copyright": 0,
+                      "CC-BY-SA-3.0": 2, "CC-BY-SA-4.0": 2}
+    # the release's read arithmetic still holds
+    assert stats["read"] == counts["extracted"] - counts["exact_duplicate"] - counts["copied_page"] - counts["ek_copyright"]
+    assert [p["pageid"] for p in build["prepare"]["copied_pages"]] == ["20", "30"]
+    assert build["prepare"]["copied_pages"][1]["reason"] == "Mysz: deleted"
+    assert build["attribution_rows"] == 2
+    assert build["attribution_sha256"] == fm.digest((tmp_path / "wikipedia.attribution.jsonl").read_bytes())
+    ev = fw.load(tmp_path / "wikipedia.license-evidence.json")
+    assert ev["release_scope"] == {"labels": {"CC-BY-SA-3.0": 1, "CC-BY-SA-4.0": 1},
+                                   "tokens_by_label": {"CC-BY-SA-3.0": 4, "CC-BY-SA-4.0": 2},
+                                   "copied_pages": 2, "documents_released": 2}
+    assert ev["copyright_sample"]["population"] == {
+        "documents": 4, "rows_dropped_since": [["wikipedia_20", "CC-BY-SA-3.0"], ["wikipedia_30", "CC-BY-SA-4.0"]]}
+
+    before = {p.name: p.read_bytes() for p in tmp_path.iterdir()}
+    fw.withhold(tmp_path)  # rerun: nothing left to withhold
+    assert {p.name: p.read_bytes() for p in tmp_path.iterdir()} == before
+
+    monkeypatch.setattr(fw, "COPIED_PAGES", {10: "Ala: deleted"})
+    fw.save(tmp_path / "wikipedia.build.json", {**build, "licence_labels": {}})
+    with pytest.raises(ValueError, match="before relabel"):
+        fw.withhold(tmp_path)
+
+
 # --- released data -------------------------------------------------------------------------
 
 def _released():
-    table = pq.read_table(DATA / "wikipedia.parquet", columns=["id", "source", "created", "license", "author"])
+    table = pq.read_table(DATA / "wikipedia.parquet", columns=["id", "source", "created", "token_count", "license", "author"])
     return table, table.to_pydict()
 
 
+def _meta():
+    with (DATA / "wikipedia.attribution.jsonl").open(encoding="utf-8") as fh:
+        return {r["id"]: r for r in map(json.loads, fh)}
+
+
+def _read(name):
+    return json.loads((DATA / f"wikipedia.{name}.json").read_text(encoding="utf-8"))
+
+
+# the release's labels by contribution (v2#110), from the stub dump plwiki-20261001-stub-meta-history.xml.gz
+LABELS = {"CC-BY-SA-3.0": 719119, "CC-BY-SA-3.0 AND CC-BY-SA-4.0": 629454, "CC-BY-SA-4.0": 120074}
+CASES = {"shipped_before_cutoff": 190836, "same_revision_text": 5028, "identical_text": 492757, "no_text_added": 30498,
+         "created_after_cutoff": 120074, "changed_after_cutoff": 629338, "page_not_in_dump": 85, "pre_text_hidden": 31}
+STUB_SHA1 = "47fd7549026521bd02c174ee418c63a93f7e4118"
+
+
 def test_registry_entry():
     entry = SOURCES["wikipedia"]
     assert entry["file_key"] == "wikipedia" and "speakleash_key" not in entry
@@ -253,23 +326,29 @@ def test_tag_template_aliases_match_the_dump():
         assert fw.ek_tag("{{%s|npa}}" % name) == {"reason": "npa", "source": ""}
 
 
-def test_every_row_label_follows_its_last_revision():
+def test_every_row_is_labelled_by_contribution():
     table, cols = _released()
-    meta = {}
-    with (DATA / "wikipedia.attribution.jsonl").open(encoding="utf-8") as fh:
-        for line in fh:
-            row = json.loads(line)
-            meta[row["id"]] = row
+    meta = _meta()
     assert len(meta) == table.num_rows
-    for row_id, created, label, author in zip(cols["id"], cols["created"], cols["license"], cols["author"]):
+    for row_id, created, author in zip(cols["id"], cols["created"], cols["author"]):
         m = meta[row_id]
-        assert label == m["license"] == fw.license_for_revision(m["last_revision_at"])
         assert created == m["last_revision_at"][:10]
         assert author == f"{fw.AUTHOR}, {m['url']}"
         assert row_id == f"wikipedia_{m['pageid']}"
-    stats = json.loads((DATA / "wikipedia.stats.json").read_text(encoding="utf-8"))
-    evidence = json.loads((DATA / "wikipedia.license-evidence.json").read_text(encoding="utf-8"))
-    assert stats["licenses"] == dict(Counter(cols["license"])) == evidence["release_scope"]["labels"]
+    stats, build, ev = _read("stats"), _read("build"), _read("license-evidence")
+    check_contribution_labels("wikipedia", cols, meta, build)
+    assert build["attribution_sha256"] == fm.digest((DATA / "wikipedia.attribution.jsonl").read_bytes())
+    tokens = Counter()
+    for label, tok in zip(cols["license"], cols["token_count"]):
+        tokens[label] += tok
+    scope = ev["release_scope"]
+    assert stats["licenses"] == dict(Counter(cols["license"])) == scope["labels"] and scope["tokens"] == dict(tokens)
+    assert stats["license"] == SOURCES["wikipedia"]["license"]
+    assert stats["licenses"] == LABELS and build["licence_labels"]["cases"] == CASES
+    assert build["licence_labels"]["stub"]["sha1"] == STUB_SHA1
+    # the card's account of the labels is the one the wikimedia cards carry
+    card = (DATA / "wikipedia.md").read_text(encoding="utf-8")
+    assert fm.contribution_text("wikipedia", build, ev, stats["tokens"]) in card
 
 
 def test_copyright_sample_matches_the_release():
@@ -282,11 +361,13 @@ def test_copyright_sample_matches_the_release():
     sample = json.loads((DATA / "wikipedia.license-evidence.json").read_text(encoding="utf-8"))["copyright_sample"]
     rows = [json.loads(line) for line in (DATA / sample["file"]).read_text(encoding="utf-8").splitlines()]
     _, cols = _released()
-    prepare = json.loads((DATA / "wikipedia.build.json").read_text(encoding="utf-8"))["prepare"]
+    meta = _meta()
+    prepare = _read("build")["prepare"]
     copyright_drops = {f"wikipedia_{p['pageid']}" for p in prepare["copied_pages"] + prepare["ek_copyright"]}
     dropped = dict(sample["population"]["rows_dropped_since"])
     assert set(dropped) <= copyright_drops and not set(dropped) & set(cols["id"])
-    population = list(zip(cols["id"], cols["license"])) + list(dropped.items())
+    # the strata are the labels by last revision the release carried before v2#110
+    population = [(i, fw.license_for_revision(meta[i]["last_revision_at"])) for i in cols["id"]] + list(dropped.items())
     random.seed(sample["seed"])
     expected = []
     for label in sorted(sample["per_stratum"]):

src/wiki_history.py

@@ -84,9 +84,12 @@ def download(dump: str, out: Path, session=None) -> dict:
     session = session or requests.Session()
     base = f"https://dumps.wikimedia.org/{dump}/{STUB_DATE}"
     name = f"{dump}-{STUB_DATE}-stub-meta-history.xml.gz"
-    status = session.get(f"{base}/dumpstatus.json", headers={"User-Agent": UA}, timeout=60).json()["jobs"]["xmlstubsdump"]
+    jobs = session.get(f"{base}/dumpstatus.json", headers={"User-Agent": UA}, timeout=60).json()["jobs"]
+    # a big wiki (plwiki) is dumped in parts, and the whole file comes from the recombine job
+    job = "xmlstubsdumprecombine" if name in jobs.get("xmlstubsdumprecombine", {}).get("files", {}) else "xmlstubsdump"
+    status = jobs[job]
     if status["status"] != "done":
-        raise ValueError(f"{dump} xmlstubsdump not done: {status['status']}")
+        raise ValueError(f"{dump} {job} not done: {status['status']}")
     meta, dst = status["files"][name], out / name
     if not dst.exists() or dst.stat().st_size != meta["size"]:
         subprocess.run(["curl", "-fsSL", "--retry", "8", "-C", "-", "-A", UA, "-o", str(dst), f"{base}/{name}"], check=True)
@@ -383,8 +386,11 @@ def compare(source: str, sidecar: Path, history: Path, fetched: Path, parquet: P
     cfg = WIKIS[source]
     convert = convert or _converter(source)
     side = [json.loads(line) for line in open(sidecar, encoding="utf-8")]
-    hist = {r["pageid"]: r for r in map(json.loads, open(history, encoding="utf-8"))}
-    pre = {r["revid"]: r for r in read_texts(fetched)}
+    # only this sidecar's pages and revisions, so a big source can run as sidecar slices in parallel
+    pageids = {int(r["pageid"]) for r in side}
+    hist = {r["pageid"]: r for r in map(json.loads, open(history, encoding="utf-8")) if r["pageid"] in pageids}
+    revids = {h["pre_revid"] for h in hist.values()}
+    pre = {r["revid"]: r for r in read_texts(fetched) if r["revid"] in revids}
     need = [r["id"] for r in side if r["last_revision_at"] >= cfg["cutoff"]]
     shipped = dict(zip(*pq.read_table(parquet, columns=["id", "text"], filters=[("id", "in", need)] if need else None)
                        .to_pydict().values())) if need else {}