Today, from real use
Last 7 days
"False alarms?" counts warnings someone said were fine; "misses?" counts quiet verdicts someone said were a scam. Both exist only when people tap the buttons.
On the test corpus
Daily self-test
Every day at 04:00 UTC the live service checks a scam text, a genuine bank text, a QR image, an invoice PDF, an English and an Egyptian Arabic voice note, email headers, the wallet lists, an app file and a page render. You're told on Telegram only when something changes.
Investment warning list
The regulators' warning lists can't be fetched automatically (they block programs or build their pages in the browser), so this list is kept by hand. Copy each firm from the regulator's own warnings page — one line each: name, websites (space separated), regulator, date, source page. People checking a platform see the regulator, date and source page.
What people use
Every tool and channel, from the verdict log. People are distinct hashed accounts in the last 30 days.
Shared with Zala — waiting for review
A person's "it was a scam" is evidence, not a label. Your decision here enters the corpus the next time the Mac runs npm run review -- --sync.
Examples on the website
Only messages whose sharer said they may be shown, that you decided are scams. Remove every name and anything that could point to a person before publishing — the redactor does not remove names.
Invite people
Ready for pilot users?
What's set up in this deployment — yes or no, never a value.
Pilot invitations
Where the money goes
Deletion requests
Delete the Clerk user and the account's data (Forget everything), then mark it. Due within 30 days.