slashyear MCP server

slashyear

Every sentence published on slashyear.com is quoted verbatim from a numbered English Wikipedia revision, and the row carries that revision id, the article title and the section anchor it came from. No language model writes, paraphrases or summarises any published wording. This repository is the pipeline that produces those rows, the site that serves them, and the gates that prove the claim.

The rows are the point. Wikipedia publishes prose, one article per year. slashyear publishes the individual dated entries underneath that prose as records — tagged with year, calendar day, subject, century, decade, theme and source revision — so a year can be queried, joined, diffed and cited rather than only read.

What a row is

{
  "date": "January 20",
  "text": "January 20 – Richard Nixon is sworn in as the 37th president of the United States of America.",
  "cite": {
    "url": "https://en.wikipedia.org/w/index.php?oldid=1373164180#January",
    "title": "1969",
    "revid": 1373164180,
    "section": "Events › January"
  }
}

text is the source line with wiki markup mechanically removed and nothing else changed. revid is a frozen revision, not a live page, so the citation still resolves to the same words years from now even if the article is rewritten. That is what makes the corpus checkable instead of merely attributed.

The guarantee, and how it is checked

pipeline/verify.py walks the full chain for every bullet on every page:

1. the published text equals clean_line() applied to the stored wikitext line — nothing drifted in rendering;
2. that wikitext line appears verbatim in the stored article revision;
3. with --live, the stored revision is byte-identical to what Wikipedia serves for that revision id, so the local copy was never edited or corrupted.

A failure at any step means a published sentence exists that no source supports, which is the one defect this project cannot ship with, so the exit code is non-zero and a build can be gated on it. Nine further gates check the things verify cannot see — coverage against a fixed event list, famous-case recall, per-page-kind surface invariants, an independent re-render, life dates against Wikidata, section themes against the article's own headings, the exported HTML, and the live machine surface. See docs/PIPELINE.md and tests/.

Reproduce it

git clone https://github.com/ctrl-maud/slashyear && cd slashyear
python3 -m venv .venv && .venv/bin/pip install requests numpy
scripts/bootstrap.sh          # fetch the published data snapshot (no re-harvest)
scripts/reproduce.sh          # rebuild the pages from it and run the gates — prints PASS/FAIL

bootstrap.sh pulls the released snapshot from this repository's v1.0.0 release because the raw harvest is hundreds of megabytes of Wikipedia revisions and takes hours. To rebuild the corpus from Wikipedia itself instead, start at pipeline/harvest.py and follow the stage order in docs/PIPELINE.md — the order matters, and the document explains where it bites.

Layout

Getting the data

• Reproduction snapshot — slashyear-data-snapshot.tar.gz (105 MB) on the v1.0.0 release: the extracted claims, the Wikipedia revisions they came from, and the caches the gates need.
• Bulk dumps — https://slashyear.com/dump/events.ndjson.gz, years.ndjson.gz, entities.ndjson.gz, plus a Frictionless datapackage.json.
• Static JSON API — https://slashyear.com/api/year/1969.json, /api/date/july-20.json, /api/entity/byzantine-empire.json, and an OpenAPI description at /api/openapi.json. CORS-open, no key, no rate limit, no signup. See docs/API.md.
• MCP server — https://slashyear.com/mcp, so a model can query the corpus directly. No key; client setup in llms-install.md.

Beyond the year pages

• The world on a date — /api/asof/1969-07-20 (page): the entries dated to that day, every state's office holders that year, the Hot 100 No. 1 in force, U.S. prices, world population, official series, Federal Reserve exchange rates for that day and the year's Nobel Prizes, each part with its source. site/edge/asof.js.
• Claim checking — /api/check?claim=… (page). Leaders, No. 1 singles and Nobel Prizes get a structured verdict, because there the record is complete enough to say "contradicted". Any other claim gets the dated entries holding its key words and where the record dates them. No model writes a verdict.
• Where the sources disagree — pipeline/disagreements.py → /api/disagreements.json (page). Every date the build's cross-checks found in conflict: state-leader lists vs Wikidata, year-article births and deaths vs Wikidata, the Nobel Prize API vs Wikidata. Both sides quoted with their revision or item; no side picked.
• Nobel Prizes — /api/nobel.json, from the Nobel Prize API (CC0), each laureate kept only where Wikidata records the same prize in the same year.
• Official series — /api/official/index.json (page): U.S. and scientific series taken from the publishing agency and checked against a second copy before they go out.

Licensing

• Code — Apache-2.0 (LICENSE).
• Rows — CC BY-SA 4.0 (LICENSE-DATA). The text is derived from English Wikipedia, which is share-alike, so derivatives inherit that licence. Attribution goes to the Wikipedia contributors of the cited revisions; the citation shipped with every row is what makes that attribution possible.

Citing

CITATION.cff is present, so GitHub renders a "Cite this repository" button.