JEV
Bilingual research report on what Jev is, how its claims compare with independent evidence, and the open-source alternatives.
This repository is mostly documentation or write-up rather than runnable code.
Overview
JEV Research Report by Chorylee7 is a Chinese and English research report on TypeSafe's System One decision model and its open-source counterparts, with an evaluation scaffold and a changelog. Every number is tagged by evidence level: official, third-party, project self-reported, or unverified. Conclusions were checked on 21 September 2026, with star and licence snapshots from 29 September; the author ran no benchmarks.
Its quick reference card summarises Jev as a hosted decision model with noul, choice (up to 255 options) and score (2-10 levels) primitives at $0.042 per million input tokens with free output, a 64k request limit and text-only input. It notes official speed claims (193.6x versus 40-200x) contradict each other while independent tests show roughly 5-25x speed and 9-580x cost advantages, with Chinese tests near 2.7x cost advantage. Weak spots include invoices, Chinese customer service, high-cardinality intents, arithmetic, dates and indirect reasoning; abstention erodes cost advantages; and it advises using the max of the probabilities rather than the official confidence field. It names LAYA as the leading open alternative.
What the decision model decides
| Type | Question | Answers |
|---|---|---|
| Other | What is Jev, are its claims supported, and what are the alternatives? | — |
How it works
- Collect official, third-party and project claims
- Tag each with evidence level
- Synthesise a quick reference and detailed sections
- Provide an evaluation scaffold
Strengths
- Evidence-level tags on every claim
- Bilingual
- Practical advice on when not to use Jev
Caveats
- No original benchmarking by the author
- Snapshot dates; figures may change
At a glance
- Type
- Project
- Status
- research
- Has code
- yes
Use cases and integrations
How curated lists describe it
JEV 调研报告：TypeSafe System One 决策模型与开源对标（含原始核实记录）.
hellogumbo/awesome-jev
Frequently asked questions
Are Jev's speed claims accurate?
The report finds official figures contradictory and independent tests around 5-25x faster.
When should Jev be avoided?
High-cardinality classification, heavy Chinese workloads, numeric thresholds and very high-frequency control loops.
Which open alternative is recommended?
LAYA, per the report.
Related projects
Datawhale's Chinese Jev course: notebooks on the three primitives, translated recipes, evaluation and local Laya fine-tuning.
ArticleproductionPythonJev-compatible open System One engine with a Claude Code-style terminal REPL showing live probabilities, plus an experiment harness.
ProjectprototypePython33-chapter Chinese tutorial on agentic RL and GRPO, including a chapter on when Jev-style external decision models beat training with RL.
ProjectresearchPythonJapanese research deck, article and source ledger tracking TypeSafe's Jev and System One models over time.
ProjectdiscussionMarkdownHead-to-head eval page where TrustedRouter's gev-1.0 outperforms Jev on the same JevBench questions and also accepts images.
ArticlediscussionHacker News thread on Decision Index 0.2.1, which compared about 70 open reproductions of Jev's decision model.
Discussiondiscussion