Document the source inventory sheet and append findings

Point AGENTS.md at the live inventory spreadsheet and record the
FinePDFs / Sejm interpellation / journalism checks so the next
contributor does not redo them.

Co-authored-by: Cursor <cursoragent@cursor.com>

AGENTS.md

@@ -22,12 +22,15 @@ So datasheets and the README narrative are **hand/contributor-maintained**.
 Never run `make_docs.py` and commit the result blind — `git diff` first and
 restore anything it dropped or restamped.
 
-## Before you start: check the findings log
+## Source inventory
 
-`artifacts/source_findings.md` records sources already investigated, including
-the rejected ones and *why*. Check it before spending a day on a source someone
-already found to be blocked — and append your own finding there whether the
-answer was yes or no.
+Consult the live inventory **before** investigating or adding a source:
+https://docs.google.com/spreadsheets/d/1lFrY2GPRHp_zfwEAKvX2v8fxBwkgWkzdJvtgP41F5mA
+
+It is the first-stop index of shipped shards, open HF PRs, and investigated
+candidates (status, license, contributor, document-author field, blockers).
+After a new investigation, update the sheet and append the detailed argument
+to `artifacts/source_findings.md`.
 
 ## Add a source
 
@@ -150,8 +153,8 @@ before staging. `*.parquet` is LFS-tracked; commit the pointer, not the blob.
 
 ## Known drift / cleanup opportunities
 
-- `biblioteka_nauki`, `europeana`, `parlamint_pl` are in `SOURCES` but have no
-  `data/` shard (make_docs skips them with `! no stats`). Intentional placeholders
-  or stale — confirm before relying on `build_dynaword.py --all`.
+- `europeana` is in `SOURCES` but only has a preview shard under
+  `previews/v0.3.1/`. Confirm the inventory sheet before treating a registry
+  entry as a released parquet.
 - The global `ADDED` datasheet stamp is a footgun: give datasheets a per-source
   add date (or `custom_datasheet`) before making make_docs regenerate them.

artifacts/source_findings.md

@@ -541,3 +541,467 @@ Crawl the corpus already draws on.
 If the goal is raw scale: SAOS. If the goal is a corpus that is not two-thirds
 legalese, **there is no billion-scale option**, and the honest plan is a number of
 50–200M-token sources across different registers.
+
+> **Addendum 2026-08-31:** the last sentence is now too strong. `HuggingFaceFW/finepdfs`
+> `pol_Latn` (54.6B tokens claimed, ~43–45B measured) and its `-edu` subset (~12.6B)
+> are billion-scale, contemporary, PDF-sourced and not primarily legal — see the
+> FinePDFs entry below. They are still Common-Crawl-derived, so the "only
+> billion-scale contemporary source is crawl" half of the conclusion stands.
+
+---
+
+## FinePDFs — Polish PDFs from Common Crawl (`HuggingFaceFW/finepdfs`, `-edu`)
+
+**Investigated:** 2026-08-31
+**Verdict:** `hold` — mechanically build-ready and by far the largest non-legal
+Polish source found to date (**54.6B tokens claimed / ~43–45B measured**), but it
+is web-derived PDF text with only email+IP anonymisation upstream. Blocked on the
+same missing PII scrub as FineWeb-2, plus an overlap measurement and a deliberate
+threshold decision.
+
+This is a different corpus from FineWeb-2/HPLT: sourced from **PDFs** found in
+Common Crawl (CC-MAIN-2013-20 → CC-MAIN-2025-08), not HTML pages, with truncated
+PDFs refetched from the origin server.
+
+### Size — measured 2026-08-31
+
+Card table (`huggingface.co/datasets/HuggingFaceFW/finepdfs`, README):
+
+```
+pol_Latn   9,692,213 docs   54.63 B tokens   55.02 GB
+```
+
+`datasets-server/size` agrees on rows: **9,693,153 rows**, 79.70 GB parquet,
+23 columns. Independent check by sampling 80 rows across 12 evenly spread
+offsets: mean 12,653 chars/doc (median 2,668, p10 365, p90 26,824), mean
+`token_count` 4,486 → **43.5B** via `token_count`, **45.0B** at the repo's
+measured 0.367 tokens/char. The card's 54.6B is presumably a different
+tokenizer; either way this is ~4.5× the whole current corpus.
+
+`HuggingFaceFW/finepdfs-edu`, `pol_Latn`: **1,001,475 rows**, 16.87 GB. Sampled
+100 rows across 12 offsets: mean 34,402 chars (median 9,461) → **~12.6B tokens**
+(`token_count` mean 12,665 agrees: 12.7B). Card: filtered by a per-language
+educational classifier trained on Qwen3-235B annotations, and — unlike
+FineWeb-Edu — **globally deduplicated**.
+
+### The gate FineWeb-2 could not give us
+
+The earlier billion-scale entry rejected FineWeb-2 mainly because
+`clean_hplt_v3.py`'s signals (WDS score, register class, MT-probability) have no
+counterpart there. FinePDFs is better equipped. Columns:
+
+```
+text, id, dump, url, date, file_path, offset, token_count, language,
+page_average_lid, page_average_lid_score, full_doc_lid, full_doc_lid_score,
+per_page_languages, is_truncated, extractor, page_ends,
+fw_edu_scores, fw_edu_v2_scores, dclm_scores, ocr_quality_scores,
+minhash_cluster_size, duplicate_count
+```
+
+`fw_edu_scores` is populated per page in every sampled Polish row, and the
+classifier itself is published, so the quality threshold stays a tunable knob —
+or one can skip the work entirely and take the `-edu` subset.
+
+**Caveat found by sampling, not read off the card:** `dclm_scores` and
+`ocr_quality_scores` were `None` in *every* Polish row sampled (180 rows,
+including all 90 `rolmOCR` ones). Do not design an OCR-quality gate around that
+column without first confirming it is populated at scale for `pol_Latn`.
+
+### Extraction and text quality — verified on samples
+
+`extractor` splits digital-born (`docling`) from scanned (`rolmOCR`, GPU):
+50/80 vs 30/80 in `finepdfs`, 40/100 vs 60/100 in `finepdfs-edu`.
+
+The OCR half is **good**, which is the opposite of the `PleIAs/Polish-PD`
+finding above. Sampled `rolmOCR` openings — contemporary orthography, tables
+preserved as markdown, no character garbage:
+
+```
+PION PROREKTORA DS. NAUKI I WSPÓŁPRACY Z ZAGRANICĄ | CENTRUM PRZEDSIĘBIORCZOŚCI …
+OGŁOSZENIE O PRZETARGU | Działając na podstawie przepisów art. 38-41 ustawy …
+Analiza incydentów naruszających bezpieczeństwo teleinformatyczne … CERT Polska …
+```
+
+`is_truncated` is set on 7/80 sampled `finepdfs` rows (9%) and 49/100
+`finepdfs-edu` rows (49% — the subset selects long documents, which are the ones
+Common Crawl truncates). Those documents stop mid-text; decide keep/drop
+explicitly rather than inheriting it.
+
+### Register — the honest read
+
+Domains are extremely dispersed: across 180 sampled rows no host appeared more
+than 3 times (`zgm-otwock.pl`, `sp9.wroclaw.pl`, `law.umk.pl`, `up.lublin.pl`,
+`pttk.torun.pl`, `1943.pl`, `cert.pl`, …). Genres seen: university monthly,
+municipal tender notices and land-lease listings, a CERT Polska incident report,
+training schedules, a solar-panel datasheet.
+
+So it is long-form **institutional/administrative** prose. That is new relative
+to HTML crawl, but the tender/notice slice pulls in the same direction as the
+corpus's existing legal-official skew. `finepdfs-edu` is the better register
+target for that reason, not just the smaller one.
+
+### License evidence — verified
+
+- README front matter: `license: odc-by`.
+- README body: *"As is tradition, the dataset is fully reproducible and released
+  under the **ODC-By 1.0 license**."* Same for `finepdfs-edu`.
+- Same argument shape as the accepted HPLT basis (packaging licensed, underlying
+  documents not owned by the publisher), but **ODC-By requires attribution**,
+  unlike HPLT's CC0. State the obligation in the datasheet.
+- Reproduction code: `github.com/huggingface/finepdfs`.
+
+### Blockers
+
+1. **PII.** The card's own pipeline step reads: *"PII Anonymization 🎭 — Kept
+   unchanged, emails and ip addresses are anonymized."* Emails and IPs only.
+   AGENTS.md requires phone numbers and PESEL/NIP/REGON/account numbers too, and
+   this is not hypothetical here: a sampled municipal *wykaz nieruchomości*
+   listed named private lessees. `sources.py` already claims a PII scrub for the
+   HPLT shard with no implementing code in `src/`; shipping a second such source
+   would be indefensible.
+2. **Overlap.** Dz.U./M.P., EUR-Lex and gov.pl publications circulate as PDFs and
+   are heavily crawled, so this shard plausibly re-contains `dziennik_ustaw`,
+   `eurlex` and `govpl` material. Upstream MinHash is per-language *within*
+   FinePDFs only. Repo still has no committed near-dedup tooling — measure before
+   ingesting, do not assume.
+3. **Balance.** Ingesting the full 43B+ would make DynaWord ~80% this one source.
+   `finepdfs-edu` (~12.6B) or an `fw_edu` threshold inside it is the sane
+   operating point, and even that roughly doubles the corpus.
+4. **Compute.** 55 GB (full) / 16.9 GB (edu) download plus tokenization.
+
+### Lesson
+
+FinePDFs answers the objection that killed FineWeb-2 here: a published
+per-language quality classifier and per-page scores ship *with* the data. But
+verify which score columns are actually populated for `pol_Latn` before
+designing a gate on one — two of the four were empty in every row sampled.
+
+---
+
+## Sejm API — interpelacje i zapytania poselskie (recheck)
+
+**Investigated:** 2026-08-31
+**Verdict:** `ready to build` — supersedes the 2026-08-27 `hold — recheck` above.
+The host was reachable, every endpoint in question exists, and the reply bodies
+are retrievable as HTML for about half the replies. **~115M words ≈ 305M tokens**
+with no new license work.
+
+### Enumeration — solved
+
+```
+GET https://api.sejm.gov.pl/sejm/term{N}/interpellations?limit=&offset=
+GET https://api.sejm.gov.pl/sejm/term{N}/writtenQuestions?limit=&offset=
+GET .../interpellations/{num}/body                  → server-rendered HTML
+GET .../interpellations/{num}/reply/{key}/body      → server-rendered HTML
+```
+
+Totals come back in headers, not the body — `X-Total-Count` and
+`Content-Range: items 0-1/19528`. Counts on 2026-08-31:
+
+| term | interpellations | written questions |
+|---|---:|---:|
+| 7 | 34,895 | 8,627 |
+| 8 | 34,043 | 9,955 |
+| 9 | 44,315 | 6,851 |
+| 10 | 19,528 | 3,913 |
+| **total** | **132,781** | **29,346** |
+
+Terms 1–6 were probed and every request **timed out at 20 s** (no HTTP status).
+Treat 7–10 as the served range until someone proves otherwise; the 2026-08-27
+entry above shows this host does go flaky, so a timeout is not a 404.
+
+### Reply availability — the number that decides the design
+
+Metadata-only sample of **600 interpellations** (terms 8/9/10, six offsets):
+743 replies, of which **389 (52.4%) are `onlyAttachment: true`** — the reply
+exists solely as a linked PDF (`/interpellations/attachment/<KEY>/i00501-o1.pdf`).
+22/600 interpellations had no reply at all. So roughly half the ministry answers
+are behind a PDF extraction step; the other half come as HTML from the `body`
+endpoint.
+
+### Lengths — small samples, treat as ±30%
+
+| document | n | mean words | median |
+|---|---:|---:|---:|
+| interpellation body | 24 | 388 | 280 |
+| reply body (text ones) | 13 | 552 | 269 |
+| written question body | 8 | 350 | 191 |
+
+Size at 2.654 tokens/word: 132,781 × 388 ≈ 51.5M words, 29,346 × 350 ≈ 10.3M,
+text replies (≈0.59 per interpellation × 552) ≈ 53M → **≈115M words ≈ 305M
+tokens**. The attachment-only replies are a further ~50M words if PDF extraction
+is added. Worth checking whether the FinePDFs `pol_Latn` shard already carries
+those `sejm.gov.pl` PDFs before writing an extractor.
+
+### License
+
+Unchanged from the shipped `sejm_api` shard: official parliamentary materials
+outside copyright under art. 4 pkt 2 pr. aut., reusable under the Open Data Act.
+No new license question.
+
+### Notes for the fetcher
+
+- The `body` HTML repeats a metadata header — title, `int-recipient`,
+  `intAuthor` (*Zgłaszający:*), `intDateTresc` (*Data wpływu:*). Those belong in
+  columns, not in `text`; the same strings recur on every document, so they are
+  exactly the frequency-detected boilerplate AGENTS.md asks for.
+- Signatories are MPs and ministry officials acting officially → names stay.
+  Citizen correspondence is sometimes quoted inside interpellations, so spot-check
+  for private-individual data before the parquet is written.
+- Register is genuinely additive: question/answer argumentation and ministerial
+  justification, not statute text.
+
+---
+
+## Contemporary Polish journalism — three candidates checked, all closed
+
+**Investigated:** 2026-08-31
+**Verdict:** `blocked` / `rejected`. Recording these so the next person does not
+re-check the same footers.
+
+### naukawpolsce.pl (Fundacja PAP) — `blocked`, permission notice ≠ license
+
+Science journalism, contemporary, editorially reviewed — exactly the register the
+corpus lacks. The reuse page (`https://naukawpolsce.pl/node/28615`,
+*Wykorzystywanie treści*, updated 13.02.2024) says, verbatim:
+
+> *"Fundacja PAP zezwala na bezpłatny przedruk większości artykułów z Serwisu
+> Nauka w Polsce pod warunkiem podania źródła artykułu i mailowego poinformowania
+> redakcji raz w miesiącu o fakcie korzystania z serwisu. Informacja o możliwości
+> kopiowania/przedruku znajduje się w prawym górnym rogu stron, na których
+> zamieszczone są artykuły."*
+
+Three reasons this does not ship as-is: **"większości artykułów"** leaves the
+licensed subset undefined at document level (the page points to a per-article
+marker instead, so any build would have to be per-record license-gated, MIXED
+shape); the grant is conditioned on a **monthly notification email**, an ongoing
+obligation a redistributed corpus cannot carry for its downstream users; and the
+footer still reads *"Copyright Fundacja PAP 2025"* with no CC license anywhere.
+
+Would unblock: written permission from Fundacja PAP, or confirmation that the
+per-article marker is server-rendered *and* names a real license — plus a
+decision on the notification clause.
+
+### oko.press — `rejected`
+
+Footer, fetched directly: *"Copyright © 2016 - 2026 . Made with love by OKO.press
+team. All Rights Reserved."* No Creative Commons statement anywhere on the front
+page. Closed.
+
+### klubjagiellonski.pl — `rejected`
+
+Long-form political essays, and the homepage HTML contains **no** license,
+copyright, "licencja" or Creative Commons string at all. Silence is not a grant;
+`/regulamin/` is 404. Closed unless someone finds an actual terms page.
+
+(`culture.pl` returned HTTP 403 to scripted requests and was not pursued.)
+
+### Lesson
+
+The same test that worked for `govpl`/`samorzad_gov_pl` — fetch the page, grep the
+server-rendered HTML for *licencj|Creative Commons|CC BY|prawa autorskie* — closes
+most media candidates in one request each. Polish outlets that permit reprinting
+tend to do it by **permission notice with conditions**, not by CC license, and a
+notice scoped to "most articles" cannot be turned into a shard.
+
+---
+
+## Polish-language Q&A / forum dumps — no Stack Exchange site exists
+
+**Investigated:** 2026-08-31
+**Verdict:** `rejected` — the resource does not exist, so stop looking for it.
+
+Technical Q&A would fill the conversational gap with CC-BY-SA-licensed text, as
+it does for other languages. It is not available for Polish: the official Stack
+Exchange dump (`archive.org/metadata/stackexchange`, files last written 2024-04)
+contains **no `pl.*` item at all** — the localized Stack Overflows are only
+`es.`, `ja.`, `pt.`, `ru.`. `pl.stackoverflow.com` does not resolve in DNS.
+
+Conversational Polish therefore still depends on the CLARIN-PL PELCRA route,
+which is `blocked — temporary` above.
+
+---
+
+## Python documentation — Polish community translation (`python-docs-pl`)
+
+**Investigated:** 2026-08-31
+**Verdict:** `hold — highest-priority new candidate, pending a no-machine-
+translation sample audit.` The provenance, reproducibility, and rights trail are
+unusually clean; the remaining question is whether the selected material meets
+this repository's stricter human-authorship rule.
+
+### Enumeration
+
+At repository revision `29fde82726566209944a0692a8d1ab879ae14bc7`, the official
+`python/python-docs-pl` translation repository contains **501 `.po` files**
+(16.4 MB). Its generated coverage report says 5.76% of source words and 12.49%
+of strings are translated, with core articles at 100%. Rebuild from the pinned
+CPython source revision and `Doc/locales/pl/LC_MESSAGES/`; emit only non-empty
+Polish `msgstr` content, assembled back into page-level documents, rather than
+crawling a rendered site.
+
+### License evidence
+
+- The official Polish translation repository: `https://github.com/python/python-docs-pl`
+  says translator submissions are donated to the PSF under CC0.
+- The underlying documentation is under the PSF License Version 2:
+  `https://docs.python.org/3/license.html`.
+- Python's translation guide identifies the Polish team and says machine-
+  translated material must be reviewed by a human:
+  `https://devguide.python.org/documentation/translations/translating/`.
+
+### Blockers
+
+The guide's review requirement is not proof that every historical Polish string
+was human-authored or reviewed. Sample the core-guide slice for English fallback,
+code/UI leftovers, and translation quality; do not label it ready before that
+audit. Strip reStructuredText markup, code/configuration blocks, and navigation
+boilerplate before the normal build gates.
+
+### Lesson
+
+A versioned translation repository plus an official render is a better source
+contract than HTML scraping, but an upstream policy is not a substitute for the
+corpus's own no-machine-translation check.
+
+---
+
+## Kubernetes documentation — Polish localization
+
+**Investigated:** 2026-08-31
+**Verdict:** `hold — legally clear but low-yield; validate authorship and text
+quality before treating it as a source.`
+
+### Enumeration
+
+Repository `kubernetes/website`, revision
+`7b31f0c8a465f19bc808872939968ed4dd393c69`, has **86** Markdown files under
+`content/pl/`, totaling 44,769 whitespace-delimited words before front matter,
+shortcodes, and code blocks are removed. The content guide documents the
+`content/<language_code>/docs` layout:
+`https://kubernetes.io/docs/contribute/style/content-guide/`.
+
+### License evidence and blockers
+
+The repository declares CC-BY-4.0:
+`https://github.com/kubernetes/website`. Build page-level records from the pinned
+source, preserving canonical URLs and contributor attribution, but first sample
+for untranslated English and evidence of machine translation. Even before
+cleaning, this is only a small technical/instructional shard, so it should not
+displace higher-value register coverage.
+
+### Lesson
+
+A real language directory is not enough: measure prose after code removal and
+verify the repository's human-text policy separately.
+
+---
+
+## OpenStax Poland — academic textbooks
+
+**Investigated:** 2026-08-31
+**Verdict:** `blocked — do not ingest without written clarification of AI-use
+rights.`
+
+OpenStax Poland's catalog would add valuable academic textbook prose (physics,
+psychology, economics, marketing, and nursing). Chapter HTML and PDFs are
+enumerable, but a current book page simultaneously displays a CC BY statement,
+an instruction that the textbook may not be used for AI training or processing
+without OpenStax/OpenStax Poland permission, and a default CC-BY-NC-SA footer:
+`https://openstax.org/books/mikroekonomia-podstawy/pages/1-podsumowanie`.
+
+The contradictory on-page terms are not a clean license chain for a training
+corpus. Ask OpenStax/OpenStax Poland which terms govern the Polish editions and
+whether redistribution in a training-data parquet is permitted before spending
+time on a fetcher.
+
+### Lesson
+
+Never treat a CC badge as sufficient where the same current page adds an
+AI-specific restriction or a different default license.
+
+---
+
+## `reactjs/pl.react.dev` — false Polish-localization lead
+
+**Investigated:** 2026-08-31
+**Verdict:** `rejected` — it currently serves English documentation, not Polish
+prose.
+
+At revision `005aa3fd7e75d0d5f16dc83acc6e88ed89b5e85f`, the repository has no
+Polish content tree, and the body text at `https://pl.react.dev/` is English
+despite some Polish interface labels. The repository's CC-BY-4.0 declaration
+(`https://github.com/reactjs/pl.react.dev`) is therefore immaterial until actual
+Polish content is published.
+
+### Lesson
+
+A language subdomain and a localized navigation shell are not evidence that the
+underlying documents are in that language.
+
+---
+
+## Mozilla Common Voice — Polish text corpus
+
+**Investigated:** 2026-08-31
+**Verdict:** `rejected` — incompatible distribution terms and weakly additive,
+fragment-level text.
+
+The Polish v26 datasheet reports 253,879 sentences (230,301 validated), but
+204,981 validated sentences (89.0%) are marked `selected-europarl-v7-pl`:
+`https://mozilladatacollective.com/datasets/cmqinmu2a00winq07gyrtri0q`.
+It declares CC0-1.0 while also forbidding re-hosting or re-sharing the dataset.
+That prohibition conflicts with committing and redistributing a parquet shard.
+The remaining records are short prompt sentences, mostly parliamentary-derived,
+which neither fit the 200-character document gate nor materially improve the
+current official/parliamentary-heavy mix.
+
+### Lesson
+
+Do not select the permissive-looking license line while ignoring a dataset
+portal's explicit redistribution restriction or its source-composition field.
+
+---
+
+## Tatoeba — Polish sentences
+
+**Investigated:** 2026-08-31
+**Verdict:** `hold — reusable, but not yet a credible DynaWord-shaped source.`
+
+Tatoeba lists **137,190** Polish sentences in its live language statistics
+(`https://tatoeba.org/en/stats/sentences_by_language`) and publishes text dumps
+under CC BY 2.0 FR, with some records additionally CC0
+(`https://tatoeba.org/en/downloads`). Its sentence IDs and contributor metadata
+make a reproducible, attributed export possible.
+
+Nearly every record is too short for the existing 200-character build gate.
+Grouping unrelated user-contributed examples into artificial documents would
+misrepresent the source, while retaining individual sentences would produce a
+large fragment shard. Resolve the record model, contributor-attribution form,
+human-authorship/quality sampling, and language/duplicate filtering before
+building.
+
+### Lesson
+
+An open dump is not automatically an appropriate document corpus; document
+boundaries and attribution can be the real blocker.
+
+---
+
+## Otwórz Książkę — historic academic books
+
+**Investigated:** 2026-08-31
+**Verdict:** `hold — attractive genre, but no current, per-book license audit.`
+
+The former ICM/University of Warsaw portal is described as hosting 102 books,
+with licenses selected per author/book, in the free-content directory at
+`https://wolnelektury.pl/biblioteki/polskie-serwisy-z-wolnymi-tresciami/otworzksiazke/`.
+The `otworzksiazke.pl` site returned 502 during this check. Do not infer a
+source-level license from a directory listing: wait for the portal to return or
+locate the preserved per-book files and license notices, then enumerate only the
+clearly reusable works.
+
+### Lesson
+
+For collections with per-work licenses, availability of a catalog is not a
+license grant for a bulk shard.