Built on Jev — a model that can't write, only judge

Ctrl+F for meaning.

Describe what you're looking for — a lie, a longing, a betrayal, a plot twist. Throughline reads every passage of the book and highlights everywhere it happens. In about a second.

Try

“someone tells a lie” in The Adventures of Pinocchio

301 passages · 1.09s · $0.0063

Chapter 121 passages with a lie — open the book →Chapter 36

Works on anything you read.

Contracts, transcripts, research interviews, your novel draft. Your document stays in this browser; passages are sent to Jev only to be judged, and Throughline never stores them.

  • Every clause that limits liability

    in a 90-page contract

  • Every time a user mentions onboarding pain

    across 40 interview transcripts

  • Every place you tell instead of show

    in your manuscript

  • Every answer that dodges the question

    in a hearing transcript

Drop a document here

PDF, EPUB, Word, Markdown or plain text · up to ~300,000 words

0 words

Search matches words. Chatbots summarize. Neither can find every lie.

Keyword & semantic search

Finds passages that mention a topic. It can't tell a lie from a passage about lying, or a hesitation from the word “hesitate”.

Chatbots

Read the book at once and hand you a few examples. They miss most instances, and sometimes quote lines that don't exist.

Throughline + Jev

Asks a yes/no question about every single passage, in parallel, and keeps the calibrated probability. Exhaustive, instant, and every highlight is real text.

That used to be the expensive part. Asking a large language model about each of Moby-Dick's 1,497 passages one by one takes minutes and real money. Jev judged all of them in 3.05s for $0.032.

How it works

  1. 01

    Split

    The text is cut into passages of about 120 words (about 240 characters in Chinese), never crossing a chapter.

  2. 02

    Ask

    For every passage, one request to Jev: the passage itself, plus one yes/no question per lens.

  3. 03

    Judge

    Jev answers every question in a single forward pass. No text is generated; it returns a calibrated probability.

  4. 04

    Draw

    Ordinary code turns those numbers into highlights, the minimap, and the throughline arc.

POST https://api.typesafe.ai/v1/systemone
{
  "model": "jev-latest",
  "state": {
    "book": "The Adventures of Pinocchio",
    "passage": "“I lost them,” answered Pinocchio,
      but he told a lie, for he had them
      in his pocket…"
  },
  "questions": {
    "lens_0": {
      "type": "noul",
      "instructions": "A reader is searching this
        book for: \"someone tells a lie\". Is this
        passage a genuine match?"
    }
  }
}
→ 91 ms
{
  "model": "jev-1.13.0",
  "answers": { "lens_0": { "type": "noul", "noul": 0.99 } }
}

Jev is TypeSafe AI's “System One” model: fast, calibrated decisions instead of prose. Throughline sends one of these per passage, dozens at a time, and the whole book comes back in about a second.