Factle
A daily "spot the fake fact" game. Every day has 5 rounds. Each round shows 5 surprising facts from Wikipedia's "Did you know…?": 4 are true, and 1 is a fake made by changing one detail of a real fact (a number, a year, a country or a nationality). Pick the fake as fast as you can: up to 100 points a round for speed, 0 for a wrong pick, out of 500 a day. Everyone gets the same rounds each UTC day, and there's a shared leaderboard.
Play: https://biccyanzac-pixel.github.io/factle/ (also on the jacob.gg hub)
How it works
pipeline/ Python: Wikipedia DYK archive -> facts -> fakes -> daily schedule (runs offline, free)
site/ the static game (GitHub Pages, gh-pages branch = site/)
worker/ leaderboard API (Cloudflare Worker + D1), https://factle-leaderboard.jacob-gg-leaderboard-worker.workers.dev
docs/ decisions: ingest.md (Phase 1), fakes.md (Phase 2), difficulty.md, architecture.md
- Facts (
pipeline/ingest.py): 121k DYK hooks are converted to statements and filtered. Facts about living people are excluded using Wikipedia + Wikidata, as are unsuitable topics and fast-changing claims. About 38k facts remain. See docs/ingest.md.
- Fakes (
pipeline/mutate.py, pipeline/checks.py) are rule-based, with no LLM. Each fake must pass a local contradiction model (MiniLM NLI cross-encoder), a "not true elsewhere" check against the source article and Wikipedia search, and grammar rules. See docs/fakes.md.
- Schedule (
pipeline/schedule.py) writes site/data/days/YYYY-MM-DD.json and site/data/index.json. Days are pre-generated a year ahead. They're public: anyone can read ahead in the repo, which the owner accepted.
- Leaderboard: the browser sends only (day, round, pick, time taken). The worker judges the pick and computes the points with the same code the site uses (
site/js/daily.js).
Develop
python -m venv .venv && .venv/Scripts/pip install requests mwparserfromhell spacy pytest torch transformers sentencepiece protobuf
.venv/Scripts/python -m spacy download en_core_web_lg
node tools/serve.mjs 8081 # http://localhost:8081/ (?api=http://127.0.0.1:8811 for a local worker)
npm test # JS: dates, timezones, scoring
.venv/Scripts/python -m pytest # Python: hook parsing, filters, living people, mutations
cd worker && npm install && npm run db:init:local && npm run dev:local # then: npm test
Rebuild the data: python -m pipeline.fetch_archive && python -m pipeline.parse_hooks && python -m pipeline.ingest, then
python -m pipeline.schedule --start <tomorrow> --days N --keep-until <today>. Never regenerate past days: --keep-until keeps them.
Deploy
- Site:
git subtree split --prefix site -b gh-pages && git push -f origin gh-pages
- Worker:
cd worker && npx wrangler deploy. Redeploy whenever site/data/index.json or site/js/daily.js changes; the worker bundles them.
Credits and licences
Facts come from English Wikipedia (CC BY-SA 4.0). Every fact links to its source articles, and the derived dataset (data/facts.jsonl, site/data/days/) is CC BY-SA 4.0 (data/README.md). Wikidata is CC0. The NLI model is cross-encoder/nli-MiniLM2-L6-H768 (Apache-2.0). Game code: no licence chosen yet.
