Corpus sourcing sample // one row per source

Provenance Sheet Workbench

A small, genuinely license-cleared, pre-2020, human-written English text sample, documented the way the full corpus would be: date proof, license clause, verification trail, register tag, dedup status. Publicly visible is not the same as licensed to train on; two famous sources below fail that screen, and the sheet says exactly why.

All license checks
run 2026-07-17
every claim linked
8sources screened
5cleared into the batch
2excluded by license screen
1cross-source duplicate caught

The provenance sheet

Click any row for its verification trail: the page checked, the clause quoted, the check date, and a short attributed excerpt. Excluded sources carry the clause that disqualified them and contribute no ingested text.

IDSourceRegisterText dateLicenseStatusDedup

Dedup, demonstrated on a real cross-source duplicate

The same 1813 text ships in two different Project Gutenberg editions (#1342, 1998 and #42671, the 2013 Chapman edition). They differ in punctuation only (a comma and the quote marks), the classic shape of a cross-site duplicate. Word-level shingles see through that: the recorded run measured a 5-word shingle Jaccard of 1.0000 between their openings. The button recomputes it from the raw strings embedded in this page, in your browser.

A. Gutenberg #1342 (1998)

B. Gutenberg #42671, Chapman ed. (2013)

The seven registers, honestly mapped

Three registers are already sampled above. The rest get a sourcing method, not a hand-wave; where the well-known source fails the license screen, the sheet says so, because the screen-out is the skill.

screen shown

Product reviews

The famous free option, the Yelp Open Dataset, is licensed for academic purposes only (row Y-01). Cleared paths: direct platform licensing, or pre-1929 printed reviews that are public domain by age.

method defined

Marketing copy

Pre-1929 US print advertising (public domain by age) in digitized periodicals, taken page-by-page with issue dates as date proof. Volume is real; each periodical still gets its own row.

sampled above

Cryptocurrency and finance forums

Bitcoin Stack Exchange (row SE-01): CC BY-SA dumps, per-post timestamps as date proof, attribution and share-alike conditions logged per row.

sampled above

Customer support messages

The Enron corpus (row EN-01): half a million public-record business emails, including customer-facing threads, with message headers as date proof.

sampled above

Academic writing

PLOS journals (row PL-01): CC BY 4.0 by policy, article dates as date proof, DOI as citation. Scales to tens of thousands of pre-2020 articles.

screen shown

Casual social media

Bulk Reddit and Twitter dumps fail the screen (row RD-01): platform terms never licensed redistribution for commercial reuse. Cleared paths are narrower and need per-source review; I would not pad the batch with murky text to hit a quota.

method defined

Non-native (ESL) business English

The hardest register. Learner corpora are usually research-only, so they get screened. The workable path: non-native-authored segments inside public-record business archives, identified by author metadata and cleared row by row.

Sheet integrity selftest

Runs in your browser against the same data object that renders the sheet and generates the CSV. Nothing here is a screenshot of a claim; it is the claim, recomputed.

The method, in six steps

01 / locateFind pre-2020 candidates in archives, dumps and public records.
02 / prove the dateEdition headers, revision IDs, message headers, issue dates.
03 / read the licenseQuote the clause that permits or blocks commercial training use.
04 / tag the registerMap every source to the domain list it serves.
05 / dedupShingle overlap against everything already in the batch.
06 / one row per sourceDate, license, origin, citation. A documented trail, not a bulk dump.