Paste any privacy policy. See what it lets them take.
Red Flags reads a privacy policy and makes a WANTED poster for the company, listing what the policy lets it take. Every charge quotes the exact sentence from the policy, so you can check it yourself.
It also catches weasel words: phrases that sound harmless ("trusted partners", "improve our services", "affiliates") and what they actually allow.
Built for Hackyard Yard #3, theme One Screen. The whole app lives on one page. No routes, no second page.
Live: https://redflags-yard.vercel.app
- Open the page. TikTok's poster is already there, from a real scan of its policy.
- Pick one of eight real policies (TikTok, Temu, Snapchat, Spotify, Discord, Zoom, LinkedIn, Google), paste your own, or paste a link to one.
- Open any charge to see the sentence it came from, highlighted in the full policy.
- Split. The policy is cut into sentences, and each keeps its exact position in the text.
- Rules. Seventeen kinds of harm, each with its own patterns, run over every sentence. A filter drops sentences that deny the harm ("Mozilla does not track users…"), make it conditional on your consent, or describe a legal requirement.
- Weasel words. Sixty-five phrases that sound harmless are matched and translated into what they permit.
- AI check. If a Groq key is set,
gpt-oss-120blooks for what the rules missed. Its findings go through the same gate as the rules: the quote has to be found word for word in the policy, and a denial or a consent-gated sentence is dropped. Without a key, the rules run on their own. - Score. Each kind of harm counts once at full weight; repeats add a little, up to double. The total becomes a score out of 100 and a letter grade.
A charge appears only if its quote matches the policy character for character. Every quote carries its start and end position, and the tests check text.slice(start, end) === quote for every charge across all eight policies. The AI cannot invent a clause.
We measured the rules against 240 sentences from the eight policies (eval/gold.json). Separate AI reviewers labeled each sentence without seeing what the scanner said. A second AI then re-read a random 15 and disagreed with none. No human has labeled them, so treat these as a careful estimate rather than ground truth. Half the sample came from sentences an early version flagged, so it leans toward hard cases.
| first version | now | |
|---|---|---|
| precision | 31.7% | 54.3% |
| recall | 74.5% | 86.3% |
An independent re-run from scratch reproduced the first column and a mid-point of 53.7%; one more fix since took precision to 54.3%. The scanner still over-flags; it leans toward showing you a borderline sentence rather than hiding a real one. eval/results.tsv lists every change we tried, including the ones we threw away.
Next.js 16, TypeScript, Tailwind CSS 4, Vercel AI SDK 7 with Groq, Vitest.
pnpm install
cp .env.example .env.local # optional: add GROQ_API_KEY for the AI check
pnpm dev
pnpm testThe tests never call Groq.
The idea predates this week. The code does not: every file here was written between Sep 21 18:00 UTC and Sep 25 18:00 UTC 2026, per the Yard #3 rule "The idea can be old. The code cannot." Built with Claude (Claude Code); the commits say which model wrote what.
MIT. Not legal advice: a policy says what a company reserves the right to do, not what it has done.