Campus Rumor Check (UNL)

Paste a rumor you saw about the University of Nebraska–Lincoln. The tool extracts the core claim, categorizes it, and checks it against verified sources — official university channels (emergency.unl.edu, Nebraska Today, UNL Police, the UNL Daily Crime & Fire Log) and established local news (Daily Nebraskan, 1011 News/KOLN, Lincoln Journal Star) — then returns a status:

• Confirmed by an official source
• Being reported by local news (no official statement matched)
• No matching official statement or news coverage found

For safety rumors (shooter, gun, lockdown, threat…), the official emergency links are pushed to the top of the result, right under the status, with a standing reminder that no alert is not an all-clear.

There's also a Past incidents timeline (curated + best-effort UNL Police crime-log scrape) and a Recent checks log.

Crime-log rows aren't just displayed — scripts/scrape_crimelog.py's output (data/crimelog_raw.json) is also matched against your claim like any other source, so a claim can come back "confirmed by an official source" off a raw police-log line, not just news/official-feed items.

Run it

python server.py

Opens http://localhost:8000/. The page needs the server — it calls /api/check.

Works with no API key: claim extraction, matching against the cached source feed, and verified-source search links. Add an Anthropic key to also run a live AI web check (Claude searches the verified domains and classifies):

pip install anthropic
# then set ANTHROPIC_API_KEY, or run: ant auth login

The AI check is on by default when available; the checkbox on the form turns it off per check.

Keep the data fresh

Script	Does	Writes
scripts/fetch_sources.py	Pulls recent items per verified domain (Google News RSS, key-free) + a snapshot of emergency.unl.edu. Daily Nebraskan uses its own site RSS instead (every article, not just what Google indexed); Nebraska Today's own feed supplements Google with same-day items	data/sources.json
scripts/scrape_crimelog.py	Parse of the UNL Daily Crime & Fire Log landing page; --backfill also walks the calendar archive	data/crimelog_raw.json (feeds both the Past-incidents panel and rumor matching)
scripts/check_rumor.py "..."	Run a check from the terminal (heuristic + optional AI)	appends data/checks.json

python scripts/fetch_sources.py --days 45
python scripts/scrape_crimelog.py --backfill   # walks the crime log's calendar archive

POST /api/refresh (or a button-free curl -X POST localhost:8000/api/refresh) runs both fetchers with their default settings. Each result shows the cache age, so a stale run is visible, not silent.

Automated refresh (GitHub Actions)

.github/workflows/refresh-sources.yml runs both scripts on a schedule (every 6 hours by default) on GitHub's own servers and commits the results — no need to run anything by hand, or have your machine on. It also has a manual "Run workflow" button on the repo's Actions tab. Both scripts are standard-library only, so the workflow needs no dependencies.

This doubles as the data-refresh story for a Vercel deploy: Vercel's filesystem is read-only at runtime, so instead of writing data/*.json from a live request, this workflow refreshes and commits it on a schedule, and a Vercel project connected to the repo auto-redeploys on every push — no database needed for the read side. (It doesn't cover data/checks.json, which is written by live /api/check calls — that part still needs real storage if it should persist on Vercel.)

Since the bot pushes to main on its own, get in the habit of git pull before you push local changes, so you don't hit a rejected push.

Data files

File	Contents
data/config.json	School, verified-source list, safety keywords, emergency panel text, model
data/sources.json	Cached verified-source items the checker matches against (seeded)
data/incidents.json	Curated timeline of notable campus-safety incidents (seeded)
data/crimelog_raw.json	Unreviewed crime-log rows from the scraper — promote notable ones into incidents.json
data/checks.json	Log of past checks

Design choices / limits

• "No match" is not "safe." Official alerts can lag a real event, and the tool only sees what has been published to the sources above. In a possible emergency: emergency.unl.edu, your UNL Alert texts/email, and 911.
• Verified-source allowlist. The AI check passes allowed_domains so Claude can only cite the sources in config.json. Widen or narrow that list there.
• Heuristic claim extraction. The result shows "we read your rumor as …" — if it misread, rephrase and re-run.
• Crime-log scrape is best-effort. The log is an ASP.NET page; the default run only parses the server-rendered landing entries, but --backfill drives the calendar's day-by-day archive directly (the currently-visible month, not the site's full 60-day window). Rows are leads, not confirmed facts, even when the checker surfaces one as a "confirmed" match — the title/summary are auto-extracted from raw log text, not written up like the curated incidents.
• Not affiliated with the University of Nebraska–Lincoln or any news outlet.
• Some seeded incident details are marked unverified — confirm against the linked sources before relying on them.