Guards a published contact address so you can delete the contact form. Header forensics reject forged mail, first-time senders confirm with one reply, and everything else is routed and digested. No puzzles, no third party, no account.

Provided free and as is, without warranty of any kind — including merchantability, fitness for a particular purpose, and the accuracy or completeness of any result. See the licence. You are responsible for checking what this produces before relying on it.

Better Than CAPTCHA

Publish a real email address again, and delete the contact form.

What it does

Watches the mailbox behind your published address and decides, per message, whether a human should ever see it — using four checks, cheapest first.

Header forensics. email-authenticity-check reads the SPF, DKIM and DMARC results the receiving server already recorded, compares the visible From domain against Return-Path, Reply-To and the DKIM signing domain, and flags link domains that don't match the sender. Deterministic Python: no model, no tokens, the same answer every time for the same headers. Forged mail dies here having cost nothing.
Known sender. Anyone who has written before goes straight through. A person is challenged at most once, ever.
Confirm reply. An unknown but authentic sender gets one automatic reply asking them to reply once — anything, even a single word. Their reply arrives as ordinary mail and adds them to the known list permanently.
Route. Real mail reaches you. Everything held back is listed in a digest, never silently dropped.

Why this instead of a CAPTCHA

Vendors hide behind CAPTCHA forms for one reason: publish an address and the inbox drowns. So the customer pays — clicking traffic lights to ask a simple question — to solve the vendor's spam problem.

A puzzle asks a human to prove they're not a robot. This asks them to send one reply, which a human does in two seconds and a bulk mailer almost never does at all, because completing a round trip per address destroys the economics of blasting. Nothing is outsourced to a third party, no visitor is fingerprinted, and there is no account to create.

Why it isn't a spam filter

Spam filters score spamminess. Most of what makes a published address unusable is perfectly authenticated, perfectly legitimate bulk pitching — it is simply not addressed to a problem you have. Rungs 2 and 3 sort that out by whether a human on the other end will spend two seconds, which no content score can measure.

Use case

You want [email protected] on your website instead of a form. A prospective client emails you and gets through on their second message, which they send within a minute of the first. A vendor blasting 40,000 addresses never replies to anything, so you never see them.

Setup

Credential — one email-read credential named contact-inbox: IMAP host, port, the address, and an app password. Read-only by construction: the folder is opened READ_ONLY, so this workflow can never mark, move or delete your mail.

Recipient — the last task sends the digest to [email protected]. Change it to a real address, a contact name, or user:<name> for an in-app inbox.

Schedule — every 5 to 15 minutes. A first-time sender waits that long for the confirmation reply, so shorter is friendlier.

State — two plain text files under $INTOUCH_HOME/state/better-than-captcha/, one entry per line, meant to be edited by hand:

file	what it holds
`known-senders.txt`	addresses that skip the challenge. Add colleagues and existing clients here before you go live.
`seen-ids.txt`	message ids already handled, so nobody is challenged twice.

What it does NOT do

It does not read replies. Promoting a sender to known when they reply is not automated yet: today you add them to known-senders.txt. Until that lands, the same person is challenged once per new thread rather than once ever.
It does not judge intent. The optional fifth rung — an AI step reading the body to separate a real enquiry from an SEO pitch — is not in this version. Everything authentic and unknown gets the confirmation reply.
It does not touch your mailbox. No flags, no moves, no deletes. "Held back" means listed in the digest and not replied to; the message is still sitting in your inbox.
It has not been run against a live mailbox. The steps and their contracts are verified and the workflow validates; the behaviour against real mail at volume is not yet measured.

Why the batch is small

maxMessages is 10. A published property is truncated at 10,000 characters, and the decision step reads the message JSON from an environment variable — a larger batch would be cut mid-object and silently drop the tail. Ten messages every five minutes is 120 an hour, far more than a contact address receives.