Jev Choice vs Score vs Noul

Choice

What it is. One option out of a fixed set. You declare the set as criteria, a map of option key to description, and the answer comes back as one of your keys plus a probability for every option.

When to use it. Unordered categories. Routing a reply to a queue, tagging a post by topic, picking which of your accounts should answer. Reach for it whenever the answer is a name your code will switch on.

{
  "state": "Been using this for three weeks and the scheduler silently dropped two posts. Anyone else seeing that?",
  "model": "jev-1.13.0",
  "questions": {
    "reply_intent": {
      "type": "choice",
      "instructions": "What does this reply want from the account owner",
      "criteria": {
        "bug_report": "Describes something that is broken or behaved unexpectedly",
        "question": "Asks for information or help",
        "praise": "Positive feedback with nothing requested",
        "promotion": "Promotes the replier's own product, link or account",
        "troll": "Insult or provocation with no substance"
      }
    }
  }
}

{
  "model": "jev-1.13.0",
  "answers": {
    "reply_intent": {
      "type": "choice",
      "choice": "bug_report",
      "probabilities": {
        "bug_report": 0.871,
        "question": 0.104,
        "troll": 0.020,
        "praise": 0.003,
        "promotion": 0.002
      },
      "confidence": 0.744
    }
  },
  "usage": { "input_tokens": 96, "output_tokens": 12 }
}

Gotchas.

• criteria is a map, not an array. There is no options field. If you copied an options: [...] example from a launch write-up, it will not validate.
• Descriptions can be null. {"calm": null, "frustrated": null, "angry": null} is valid when the key says everything. Write a description the moment two options could overlap.
• 255 options is the ceiling. Past that, do two stages: a Choice over groups, then a Choice inside the winning group.
• probabilities covers every option and sums to 1.0. Read it. A top answer at 0.44 with the runner-up at 0.42 is a coin flip with a label on it.
• The key is the contract. choice returns your key verbatim, so name keys the way your switch statement wants them.
• confidence is derived. It is a statistic computed from the distribution you already have, not an independent second opinion.

Noul

What it is. A yes or no statement, answered with one float from 0 to 1. You write instructions as a statement rather than a question, and the float is the probability that the statement is true of the state.

When to use it. Flags. Is this engagement bait, does this contain a slur, does this read as AI written, is this off topic for the account. Nouls are cheap to add, so a feed filter usually runs several at once.

{
  "state": { "draft_post": "Drop a 🔥 if you agree and I'll DM you the template. Follow for part 2.", "platform": "X (Twitter)" },
  "model": "jev-1.13.0",
  "questions": {
    "slop": {
      "type": "noul",
      "instructions": "This reads like it was generated by an AI rather than written by a person.",
      "criteria": {
        "true": "Uniformly long sentences, triadic lists, words like delve or landscape.",
        "false": "Uneven rhythm, specific detail, an identifiable voice."
      }
    },
    "bait": {
      "type": "noul",
      "instructions": "This post explicitly asks the reader for likes, reposts, replies, follows, or bookmarks."
    }
  }
}

{
  "model": "jev-1.13.0",
  "answers": {
    "slop": { "type": "noul", "noul": 0.21 },
    "bait": { "type": "noul", "noul": 0.97 }
  },
  "usage": { "input_tokens": 118, "output_tokens": 4 }
}

Gotchas.

• Write a statement, not a question. "This post asks for engagement" beats "Does this post ask for engagement?" A statement makes it obvious which direction high means.
• criteria is optional and takes true and false keys. Use it when the statement is fuzzy, which is most of the time. Put the boundary cases in it.
• You set the threshold, and it is not 0.5. Pick it from your own labelled examples, and keep a dead band in the middle where nothing acts automatically.
• Threshold per consequence. Blurring a post and deleting a post should not share a cutoff.
• Avoid negations. "This is not off topic" is two negations deep by the time your code reads it, and TypeSafe documents that indirection degrades accuracy.

All three in one request

TypeSafe documents that all declared questions are evaluated in the same parallel pass, so mixing types adds no extra round trip. They also share one copy of the state, which is why TypeSafe's fan-out pattern exists: a cookbook that batched 13 questions over one document measured 12.2x cheaper and 10.0x faster than asking them one at a time, with no change in the answers.

This is the shape the free post scorer uses. Seven questions, one request: one Score for reach strength on a five-level rubric, three Scores for hook, clarity and specificity on three levels each, and three Nouls for AI-slop, engagement bait and toxicity. All seven questions are answered in the same parallel pass.

import { TypeSafeClient, noul, score } from '@typesafe-ai/sdk';

const client = new TypeSafeClient({ timeout: 4000 });

const { answers, usage } = await client.systemOne({
  state: { draft_post: text, platform: 'X (Twitter)' },
  questions: {
    strength: score('How far will this post travel compared to a typical post from the same author?', STRENGTH_LEVELS),
    hook: score('How well does the opening line stop someone mid scroll?', HOOK_LEVELS),
    clarity: score('How easily does a reader with no context understand this?', CLARITY_LEVELS),
    specificity: score('How concrete is this?', SPECIFICITY_LEVELS),
    slop: noul('This reads like it was generated by an AI rather than written by a person.'),
    bait: noul('This post explicitly asks the reader for likes, reposts, replies, follows, or bookmarks.'),
    toxicity: noul('This text contains slurs, harassment, sexual content, or abuse targeted at a person or group.'),
  },
});

answers.strength.score;          // 2.41
answers.strength.confidence;     // 0.63
answers.strength.probabilities;  // { "0": 0.03, "1": 0.14, ... }
answers.hook.legend['2'];        // the rubric text the draft landed on
answers.bait.noul;               // 0.04
answers.bait.confidence;         // undefined. Nouls do not have one.

Two practical notes from running this. Pin the model string: log jev-1.13.0 rather than jev-latest, or the thresholds you tuned drift the day the alias moves. And keep the state small. TypeSafe documents that irrelevant context degrades accuracy, so send the draft and the platform, not the whole account history.