Skip to content
JevDirectory.org
Evaluation & Training

Jaggedness

The documented set of jev-1.13 failure modes — literal reading, counting, dates, indirection, context rot — each with a guardrail.

Open docs.typesafe.aiAll 117 terms
Category
Evaluation & Training
Also known as
—
Related terms
6
Directory entries
1
Docs
docs.typesafe.ai
Added
2026-09-24

Definition

TypeSafe maintains a jaggedness page for the current model: nine failure modes with a do-this-instead for each. The page frames them as design constraints rather than bugs, reviewed periodically rather than continuously.

The list is unusually honest for a vendor doc, and it doubles as a checklist for new integrations: state the exact condition, keep arithmetic in code, extract dates as components, reduce indirection, filter state, and test adversarial input.

Tagslimitationsevaluation

Related terms

Definitions that connect to this one.

Literal readingJev answers the question you wrote, not the one you meant; scoping words, negations, and implied conditions are read at face value.Evaluation & TrainingUnreliable countingJev does not count reliably — characters, occurrences, or list items — and error grows with the size of the thing counted.Evaluation & TrainingDates as textJev reads dates as strings, not ordered quantities; extract components with Choice and do the ordering and math in code.Evaluation & TrainingIndirectionQuestions about a property of a property, or with double negatives, cost accuracy; write instructions as directly as possible.Evaluation & TrainingStructural invariantsArithmetic identities you might assume — complementary probabilities summing to 1, agreement across question types — are not guaranteed.Evaluation & TrainingContext rotThe accuracy loss that comes from padding state with detail unrelated to the question; filter state before sending it.Core Concepts

From the directory

Jev 1.13 Jaggedness
A maintained list of jev-1.13's known failure modes, literal reading, unreliable counting, dates as text, indirection, context rot, and contradictory criteria, each with a guardrail.
Practices & PatternsDocs#official#docs#evaluation
Official

From the community

Posts from builders shipping with Jev right now.

Follow @typesafeai

Jev plays itself at 3D chess

Khalil
Khalil
@BuildWithKhalil
·

got access to jev and instead of doing anything useful i made it play itself at 3D chess same model on both sides, slowly developing beef with itself you can watch jev vs jev or jump in and ruin its day chess-jev.loomens.com

Watch on X
Diogo Almeida
Diogo Almeida
TypeSafe AI
@CompleteSkeptic

After co-inventing ChatGPT, I kept asking myself: why have superhuman chat models not led to AGI? I’ve spent the last 2 years in stealth building a new way to train models (RLCD), and a new type of frontier AI model that we are releasing today: Jev • 20-200x faster • 40-400x

Watch on X
Reply
Read 3 replies

dmx.to reshapes the timeline in plain language

GPT and Jev compared live

LitJev: Jev-style decisions on any local LLM

Zhengxu Yu
Zhengxu Yu
@yuzxfred
·

开源了 LitJev：我们对 Jev 类型化决策 API 实现方式的猜想，可接入任意现成 LLM。 小测试：Qwen3.8-27B，10 道 MMLU-Pro 一次完成，0.3 s。比让模型只生成答案快 2.9x；比生成完整 schema JSON 快更多（显然）。更多测试马上来。 非官方，基于公开信息。

Reply
Read 1 reply

Secret detection with Jev, tested

Jev Driving Lab: a playground for System One models