Skip to content

Jev Evaluation

A pre-registered adversarial evaluation of jev: nine experiments, 28 predictions fixed before any request was sent, and a report plus prompting guide regenerated from the measured logs.

Stars
0
Language
Python
License
MIT
Published by
Community
Author
willkelly
Last verified
Oct 2026

About this project

An adversarial evaluation of jev, the decision model sold by TypeSafe. It follows a plan written before any request was sent. That plan fixed nine experiments, the sample size of each, and twenty-eight predictions, each stated with the result that would prove it wrong. Nothing reported here was chosen after seeing an outcome. One run produced every number below: 123,805 requests, 138 minutes, $12.69, five failures, all answered by jev-1.13.0.

Quoted from the project's README. Read the full README.

Highlights

  • One run answered everything: 123,805 requests, 138 minutes, $12.69 and five failures, all served by jev-1.13.0.
  • Batching really is free here: 60 questions in one state cost 20x fewer tokens and finished 8x sooner, with identical answers.
  • Calibration holds in distribution and collapses outside it: routing ECE is 0.075, but random 3-SAT is answered satisfiable at every clause ratio.
  • verify_citations.py indexes every number in the prose and exits non-zero if a figure cannot be traced to a measurement.

Watch out

Twelve of twenty-five testable predictions held and thirteen were wrong, which the author calls the useful half. Ground truth always comes from a solver or from construction, never from the model, and every number here comes from jev-1.13.0 measured against one plan. Phase 0 is a hard gate and a support-ticket control scored 0.980 throughout, which is what makes the rest readable. Derived figures such as a drift or a rate of decline are rebuilt from the log only for the calibration experiment, the largest remaining gap against the plan. The raw logs are 1.3 GB and published separately compressed with zstd to 47 MB, with a sha256 and record count per file in manifest.json. Python 3.11 and an API key are needed to re-run, and the digest we reviewed ends mid-sentence in the closing section. Copy reviewed against the repository README on 2026-10-04, not from running it.

Questions about this entry

Short answers drawn from the linked project or page.

More Jev resources like this

24GitHub stars
Unofficial laptop study of parallel constrained decoding: prefill once, share the KV cache across fields, and decide all schema fields in one batched forward pass without writing JSON.
Jev Practices & Patterns#community#jevbest#catalog-import
16GitHub stars
Vendor-neutral harness that tests whether Jev reranking improves a small RAG stack, using frozen hybrid BM25 plus BGE-M3 candidates and isolating retrieval, rerank, and generation stages.
Jev Practices & Patterns#community#jevbest#catalog-import
11GitHub stars
Python browser agent that reads one lean page view per state and answers with a single JSON action, using any OpenAI-compatible model with no decision API or browser daemon.
Jev Practices & Patterns#community#jevbest#catalog-import
39GitHub stars
Reproduction kit that rebuilds a frozen typed-decision benchmark suite from public sources and runs any engine through it with resume, scoring, and index computation.
Jev Practices & Patterns#community#jevbest#catalog-import
7GitHub stars
Benchmark harness that tests whether models can forecast federal motion-to-dismiss outcomes from the written record, scored per claim-defendant pair with micro-Brier and clustered intervals.
Jev Practices & Patterns#community#jevbest#catalog-import
Frozen 100-question-per-cell check of TypeSafe Jev on Korean versus English reading, paraphrase, and medical exam items, with recorded responses and uncertainty intervals.
Jev Practices & Patterns#community#jevbest#catalog-import

From the Jev guides

Original write-ups on what this entry is for.

Most Jev accuracy problems are state problems. What belongs in a state, what is a pre-made judgment to leave out, and why naming the slice you ask about matters more than the wording.
Jev Practices & Patterns#state#design#primitives
Read article
A Noul returns one number and no separate confidence. How to read the value, pin its boundary with criteria, split compound questions, and set a per-action threshold that matches the cost of being wrong.
Jev Practices & Patterns#noul#primitives#confidence
Read article
A Score is a weighted position, not a ruler. Why the number can land between levels, why levels are judged separately, and how to combine several atomic scores in code instead of one crowded rubric.
Jev Practices & Patterns#score#primitives#scoring
Read article
Back to all resources

From the Jev community

Posts from builders shipping with Jev right now.

Follow @typesafeai