Skip to content

Latest commit

 

History

2 Commits

Folders and files

NameName
Last commit message
Last commit date
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

Jevals: independent benchmark data for Jev and Jev-type models

The data behind jevals.com, an independent, ground-truth benchmark of TypeSafe AI's Jev (the first System One model, called as typesafe-ai/jev through Vercel AI Gateway) and the LLMs that answer the same typed questions: noul (yes or no), choice (pick one of N) and score (place the state on a rubric). Every answer carries probabilities and is scored against human labels for Decision Score, accuracy, calibration, cost and speed. New to Jev? What is Jev? The System One model, explained.

Findings, release 2026-09-18: Jev ties the best of six LLMs on yes/no questions (PubMedQA) at 1/28 of the price, ties for second on Banking77 intents, and no model beats guessing on HelpSteer2. Read the findings.

Latest results

Decision Score, release 2026-09-18 (100 = perfect, 0 = guessing the label base rates, below 0 = worse than that). Full boards with accuracy, calibration, cost per 1,000 decisions and latency at the links.

Model noul (PubMedQA) choice (Banking77) score (HelpSteer2 helpfulness)
Gemini 3.8 Flash 73.0 74.1 4.6
Jev 69.0 67.8 9.2
GLM-5.3 60.6 66.8 7.8
DeepSeek V4.1 Flash 47.5 63.9 −19.0
Qwen3.8 Flash 62.4 62.4 −1.4
Mistral Medium 3.5 58.0 59.5 −13.7
Mercury 2.5 55.7 54.0 −5.5
Label prior 0.0 0.0 0.0
  • noul: Gemini 3.8 Flash (73.0) and Jev (69.0) tie for first in Decision Score on PubMedQA. Jev costs $0.029 per 1,000 decisions, p95 653 ms. Release 2026-09-18.
  • choice: Gemini 3.8 Flash (74.1) ranks first in Decision Score on Banking77. Jev (67.8, tied for rank 2 of 8): $0.043 per 1,000 decisions, p95 693 ms. Release 2026-09-18.
  • score: No model clearly beats guessing on HelpSteer2 helpfulness. Jev (9.2, tied for rank 1 of 8): $0.036 per 1,000 decisions, p95 670 ms. Release 2026-09-18.

Head to head: Jev vs Gemini 3.8 Flash · Jev vs GLM-5.3 · Jev vs DeepSeek V4.1 Flash · Jev vs Qwen3.8 Flash · Jev vs Mistral Medium 3.5 · Jev vs Mercury 2.5.

Releases

Layout

  • releases/<release>/board.json: the published board, one row per system and question type. The numbers on the site are these rows.
  • runs/<system>__<task>__<suite>.jsonl: one log per system and task. The first line describes the run: system configuration, prices, harness version, where and when it ran. Every other line is one decision: item_id, epoch (repeat 0 to 4), target (index of the true label in the task's options), order_seed (which option order was shown), output (the typed answer with its stated probabilities), usage, cost_usd, seconds, malformed, refusal and retries.
  • suites/<suite>/: suite.json and one file per task: the source dataset and revision, the options, the instructions, and each item's id, upstream row index and label. Item text is not included; every item points at its row in the public dataset.

Option order

Choice questions show their options shuffled; noul and score questions keep the task file's order. Epochs 0 and 1 use order_seed 0 (identical requests, for the repeat flip rate), epochs 2 to 4 use seeds 1 to 3. The order for an item is shuffled(options, fnv1a(`${item_id}:${order_seed}`)), where options is the task file's options array:

function fnv1a(s) {
  let h = 0x811c9dc5;
  for (let i = 0; i < s.length; i++) { h ^= s.charCodeAt(i); h = Math.imul(h, 0x01000193); }
  return h >>> 0;
}
function mulberry32(seed) {
  let a = seed >>> 0;
  return () => {
    a = (a + 0x6d2b79f5) >>> 0;
    let t = a;
    t = Math.imul(t ^ (t >>> 15), t | 1);
    t ^= t + Math.imul(t ^ (t >>> 7), t | 61);
    return ((t ^ (t >>> 14)) >>> 0) / 4294967296;
  };
}
function shuffled(xs, seed) {
  const a = xs.slice(), r = mulberry32(seed);
  for (let i = a.length - 1; i > 0; i--) { const j = Math.floor(r() * (i + 1)); [a[i], a[j]] = [a[j], a[i]]; }
  return a;
}

Recompute the numbers

The formulas (Decision Score, accuracy, ECE, confidence intervals and ranks, the confidence gate, flip rates, cost and latency) are at https://jevals.com/methodology/. Every number on a board can be recomputed from the logs here.

The same data is also served as JSON at https://jevals.com/api/v1/<endpoint> (releases, board, models, tasks, items, item, compare; e.g. https://jevals.com/api/v1/board), and item by item at https://jevals.com/explore/.

Licence and citation

CC BY 4.0. Cite as: Jevals (jevals.com), release 2026-09-18, suite 0.1.0. CC-BY-4.0. The source datasets keep their own licences; see NOTICE.

Corrections

If a number is wrong, a label looks wrong, or a model deserves a row, open an issue with the release and the row or item id. Every correction lands in the changelog.

About

Independent benchmark data for TypeSafe's Jev (System One model) vs LLMs: accuracy, calibration, cost. Boards + per-decision logs, CC-BY-4.0

Topics

Resources

Stars

1 star

Watchers

0 watching

Forks

Releases

Packages

Contributors