Every email, judged by Jev before it reaches you

Jev is not a language model. It cannot write a sentence. You hand it some state and a typed question, and it hands back a typed answer: a yes/no probability, one choice out of a list, or a score — each with a calibrated confidence your code can branch on.

“Is this spam?” is exactly that kind of question. We don’t need prose about the email; we need one number we can trust, fast and cheap enough to ask about every single message. With an LLM that check cost us about $0.002 per email, so it lived behind the Pro plan. With Jev it is roughly a hundredth of that, so it doesn’t need a paywall at all.

One request per email, two questions at once: a noul for the spam probability and a six-way choice for the category. The sender’s SPF/DKIM/DMARC result goes in as part of the state, because a failed check is one of the strongest phishing signals there is.

POST https://api.typesafe.ai/v1/systemone
{
  "model": "jev-latest",
  "state": "From: ...\nSender authentication: pass\nSubject: ...\n\n<body>",
  "questions": {
    "spam":     { "type": "noul",   "instructions": "Is this email unsolicited spam, a scam or phishing…?" },
    "category": { "type": "choice", "instructions": "What kind of email is this?",
                  "criteria": { "customer": "…", "transactional": "…", "newsletter": "…",
                                "cold_outreach": "…", "phishing": "…", "junk": "…" } }
  }
}

What came back for a fake “PayPaI Security” message in our first live test:

{
  "model": "jev-1.13.0",
  "answers": {
    "spam":     { "type": "noul", "noul": 0.99 },
    "category": { "type": "choice", "choice": "phishing", "confidence": 1.0 }
  },
  "usage": { "input_tokens": 507 }
}

The same test on a genuine Turkish-language quote request returned 0.04 and customer. Both round trips took about 0.8 seconds. Those are two hand-made examples, not a benchmark — which is exactly why the default below never hides anything.

• Label it (default). At 80% or above, the notification still reaches you, marked as likely spam with its score. A wrong call costs you a glance, never a customer.
• Hold it. Opt in per inbox and spam stops pinging you. Every held message is listed in Activity with sender, subject, score and category — no silent drops, ever.
• Off. One click, and that inbox’s mail is never sent to Jev.

Flagged mail also never triggers an AI auto-reply, and the score travels in the email.received webhook and the API as spam_score and category, so your own code can act on it. See the API docs.

Once every email is a cheap typed question away, filters stop being worth building. In each inbox’s settings you write a sentence — “invoices and receipts”, “a customer complaining”, “job applications” — and pick what happens: forward it to an address from your business identity, mark the notification important, or don’t notify at all.

Under the hood it is one more Jev request per email with one noul question per rule (up to ten), threshold 80%. In our first live test a Turkish e-invoice scored 0.95 against the Turkish rule “faturalar ve makbuzlar” and 0.01 against “a customer complaining”, in 0.86 seconds. Every firing is written to Activity with its score, and spam never triggers a rule — a phishing mail saying “invoice” must not reach your accountant.

To be scored, the sender, subject and up to the first 6,000 characters of the body are sent to TypeSafe AI’s API. We store only the resulting number and category; the body itself is still deleted from our side once your notification is delivered. We are a customer of TypeSafe AI, not affiliated with them. Details in our privacy policy.

The pattern that worked for us: one piece of state, several small typed questions, and a threshold on the calibrated number — then keep a human-visible log of every decision so a wrong one is cheap to catch. If your decisions are about email, you can skip the plumbing: connect an address, point a webhook at your code, and read spam_score from the payload. All Your Inboxes starts free with one inbox.