How Plainly decides
Your device reads the letter. Fixed rules written in plain code decide the verdict, and each rule leaves a receipt. The explanations are fixed, pre-written templates: nothing is generated when you check a letter, and the same facts always give the same words.
The steps
- Your device reads the letter. A photo is shrunk to about 2000 pixels, turned grey for contrast and read by Tesseract, an open-source reader that runs inside your browser. A PDF made on a computer already carries its text, so that text is used as it is. The photo never leaves your device.
- You check the words. The text appears in a box, and you fix anything that was misread. Only this text is sent, when you press "Check this letter".
- The rules reader picks out the parts. On AWS Lambda, plain code finds who the letter says it is from, its date, deadlines, amounts, payment requests and threats, by keywords and patterns. Phone numbers, web addresses and emails are pulled out the same way.
- The rules and the official list. Over 20 rules run on the text, each based on a public source that is linked in the result. Phone numbers and web addresses are compared with a list of official contacts. Deadlines like "within 21 days of this notice" are worked out by date arithmetic.
- The verdict and the receipts. The rules give a score, the score gives the verdict, and every rule leaves a line in the receipts, whether it passed or not.
- The explanation. A fixed, pre-written template for each verdict and each warning sign, in English, हिन्दी and Español. Plainly fills in the checked facts: who the letter claims to be from, the official number, the deadline. Nothing is generated when you check a letter. The templates were drafted with the coding agent that built Plainly.
The rules
These are the rules in Plainly's code today, 21 in all. Each is a few lines of plain code that run on the letter's text and on the parts the rules reader picked out. A strong sign adds 3 points, a warning sign adds 1, a note adds nothing.
| Rule | What it looks for | Weight |
|---|---|---|
payment_gift_card | Payment by gift card: iTunes, Google Play, Steam, Amazon and similar cards | Strong (3) |
payment_crypto_wire | Bitcoin or other crypto, USDT, wire transfer, Western Union, MoneyGram, bitcoin ATM | Strong (3) |
payment_personal_upi | A UPI ID or payment app pointing to a personal handle or mobile number, when the sender claims to be a government body or utility | Strong (3) |
credential_request | Asks for an OTP, PIN, password, CVV, full Social Security number, full Aadhaar number or net-banking login | Strong (3) |
threat_arrest | Arrest, police, warrant, deportation, "digital arrest", FIR, CBI, or legal action "today" | Strong (3) |
video_call_demand | Demands a video call, Skype or WhatsApp video, or that you stay on the call | Strong (3) |
ai_instruction | Text aimed at AI tools, trying to change the verdict. The quote is never shown; the result says it is in the letter. The rule sees only text that reaches it: pasted messages, emails and PDFs with a text layer. Faint or hidden print in a photo may not be read by the reader on your device. | Strong (3) |
lookalike_domain | A web or email address that imitates an official one (a letter or two changed, look-alike characters, punycode, extra words like irs-gov-refund.com) without being it | Strong (3) |
unexpected_fee_to_release | A fee you must pay before a prize, refund or job is handed over ("registration charges", "processing fee before release") | Strong (3), or Warning (1) for a parcel, because real customs charges exist |
link_bait | Asks you to click, open or scan a link to claim money, pay, or "verify" or "update" details. Links to the official website don't count. | Strong (3) when the link is a short link, a free web host or a look-alike; when the letter claims to be an agency on our list but the link isn't its site or is hidden; when a hidden link comes with a threat to block your account; or when it tells you to reply "Y" or copy the link to make it work. Otherwise Warning (1) |
callback_unofficial | Asks you to call or WhatsApp a number that isn't an official line | Strong (3) for a personal mobile number from an "organisation" that threatens a cut-off, or any unofficial number with a cut-off within hours. Warning (1) for a personal mobile alone, a cut-off threat, or "if you didn't make this payment, call ..." |
kyc_update_threat | Your account or SIM will be blocked unless you update KYC or PAN | Strong (3) when it sends you to a link or number to do it, otherwise Warning (1) |
prize_or_refund_bait | A prize, lottery win, job you didn't apply for ("your CV has been selected", "daily salary"), cheap "scheme" loan, or a refund you must click, call or fill in details to receive | Strong (3) when the only way to claim it is a short link, WhatsApp, a personal mobile or a free email address, otherwise Warning (1) |
impersonation_mismatch | Claims to be an agency on our list but gives none of its contacts, or an email "From" an official name sent from a non-government address | Warning (1), or Strong (3) when it also asks you to click, pay, call, claim or verify through it |
urgency_short | Less than 72 hours to act: "today", "within 24 hours", "immediately", "tonight", "blocked today" | Warning (1) |
freemail_official | A Gmail, Yahoo, Outlook, Hotmail, Proton or Rediffmail address given as an official contact | Warning (1) |
secrecy | "Do not tell", "keep this confidential", "don't inform your bank or family", "you have to maintain confidentiality" | Warning (1) |
shortened_or_raw_link | A short link (bit.ly, tinyurl), a bare internet address, or a free web host (vercel.app, blogspot) | Warning (1) |
press_to_connect | "Press 1 to speak to an officer", "press 9 now". A language menu doesn't count. | Warning (1) |
unknown_contact | A phone number or website not on the official list for the organisation, next to at least one official one. On its own this never makes a letter "likely scam". | Note (0), or Warning (1) when the organisation is on our list |
injection_detected_model | The rules reader itself noticed text aimed at AI tools (a second, looser check than ai_instruction). The quote is never shown. | Note (0) |
The verdict
- Likely scam when the score is 3 or more: one strong sign, or three warning signs.
- Consistent with a genuine letter only when all of these hold: the organisation is on our official list, at least one of its web addresses or phone numbers in the letter matches the list, the score is 0, and no text in it is aimed at AI tools.
- Can't tell in every other case.
Quotes. Every warning sign is shown with the words from the letter that set it off. They are copied from the text you sent, so you can find each one in your letter.
The official contacts list. Each organisation has its official web domains and phone numbers, the page we took them from and the date we checked. Web domains are matched first, phone numbers second, because real organisations publish many numbers.
Why it never says "safe"
A scammer can copy a real letter exactly, down to the layout and the wording. A photo cannot prove who sent it. So the best Plainly can honestly say about a letter is that it is consistent with a genuine one.
That is why every result, including the reassuring ones, ends with the official number and where we found it. The last step is always a call you place yourself, to a number you did not get from the letter.
When the evidence is thin, Plainly says "Can't tell" rather than guessing. A wrong "this is fine" costs far more than an extra phone call.
Privacy
- The photo never leaves your device. Tesseract reads it inside your browser. Plainly never receives the picture, and nor does anyone else.
- Only the words are sent, after you have checked them. Plainly's code on AWS Lambda (us-east-1) checks them in memory and drops them when the check ends. They are not saved to a database, a file or a log.
- No AI service sees your letter. Amazon Textract and Amazon Bedrock are switched off for this account (why).
- Logs hold only a request id, the verdict, which rules fired and timings. Never the words of the letter.
- There is no share link and no account, so there is nothing about your letter to look up later.
- To stop abuse, we count requests per network address for an hour, and total checks per day, in Amazon DynamoDB. Those counters hold no letter text and expire on their own.
- Your language and reading-level choices are remembered in your own browser. So is the reader's language data, so your next letter is read sooner.
Limits
- Reading on your device isn't perfect. Tesseract does well with clear, flat, printed pages and worse with glare, creases, handwriting or dim light. A missed word can hide a warning sign, which is why you see the words and can fix them before the check. For letters with Hindi, choose "English and हिन्दी"; it reads more slowly.
- Reading on your device needs a recent browser. Safari 16 (iOS 16) or later, Chrome or Edge 97 or later, or Firefox 102 or later. On an older phone the reader can't start; type or paste the letter's words instead, and the check works the same.
- Keywords, not understanding. The rules reader finds things by words and patterns. A scam worded in an unusual way can slip past it. On 12 real scam messages published by governments, which the rules were never tuned on, it marked 1 as "Likely scam" and the other 11 as "Can't tell". It called none of them genuine, and it called none of 11 genuine letters a scam.
- A short official list. We cover a small set of organisations in India, the United States and the United Kingdom. For anyone else, the best answer may be "Can't tell".
- Three languages. Explanations exist in English, हिन्दी and Español only, because each one is a fixed template written in advance. They were drafted with the coding agent, and the हिन्दी and Español text has not yet been reviewed by a native speaker.
- Scams change. The rules catch common patterns, not every new one. A clean result is not a guarantee.
- Not legal, tax or financial advice. For anything with money or legal steps at stake, use the official number or a trusted adviser.
Amazon Textract and Amazon Nova
Plainly was designed to use Amazon Textract and Amazon Nova; the code is built and tested but switched off because this account is on the AWS Free plan, which doesn't include them. On an upgraded account, one setting turns Nova back on for the text your device reads; Textract would also need the page to upload the photo again, which reading on the device replaced.
Until then, Plainly calls no AI service. Everything described on this page runs on services the Free plan includes: API Gateway, Lambda, DynamoDB and CloudWatch. The same Lambda serves the site's pages, because AWS hasn't yet verified this new account for CloudFront; the CloudFront and S3 setup is ready in the template for when it does.