If Jev cannot generate text, is this generating text or just selecting letters?

Selecting letters, and the author is explicit about it. Jev returns a typed decision with probabilities over a fixed option list; the text exists only because the controller composes those decisions. Nothing generates tokens — the README states there is no tokenizer, dictionary or canned response behind it, and that Jev still never returns freeform text. The reply is assembled by the controller, not emitted by the model.

How many API calls does one reply actually cost?

At least one per character, because each option is the reply prefix with exactly one character appended, plus a second call whenever several branches survive the first judgement, plus one final call to decide between complete, incomplete and invalid. The README gives no figure for a typical answer, so treat reply length as the cost driver rather than looking for a per-token number.

Why keep several branches instead of taking the most likely character each time?

Because greedy decoding gets stuck easily, in the author's words. The default mode keeps the high-probability continuations from Jev's distribution and treats generation as a small search problem, asking Jev a second time which surviving prefix can still become the best answer. That is what lets "I m" lose to an earlier branch and go on as "I am".

What happens when Jev picks EOS? Does the reply just end there?

Not blindly. EOS is one of the 98 options and is checked on every step, and once it wins the branch the controller makes one more typed decision between complete, incomplete and invalid, where only complete is treated as a finished reply. The README does not say what the loop then does with incomplete or invalid, so the termination rules beyond that check are not documented.

How do I run jev-freeform locally?

Node.js 22+ and a TypeSafe API key. Clone the repository, run npm install, export TYPESAFE_API_KEY, run npm start, then open 127.0.0.1:4177.

How many choices does each character decision offer?

98: the 95 printable ASCII characters plus newline, tab and EOS, and each option is the entire reply prefix with exactly one character appended.

What is the greedy mode versus the default mode?

Greedy just follows the selected character and gets stuck easily. The default keeps several high-probability continuations and treats generation as a search that can reject a bad path and backtrack.

CommunityJev Practices & Patterns

Jev Freeform

An experiment that composes TypeSafe Jev Choice decisions into free-form text: every next character becomes one typed decision over 98 options, with branch backtracking and a final complete/incomplete check.

See highlights

Open github.com

Stars: 3
Language: JavaScript
Published by: Community
Author: kesku
Last verified: Oct 2026

kesku/jev-freeform

User repository on GitHub

View on GitHub

An observable raw-character chat experiment powered entirely by TypeSafe Jev Choice

Forks: 0
Last pushed: Updated Sep 2026

Repo stats from the GitHub API, cached Oct 2026.

About this project

Can TypeSafe Jev Choice generate text through successive typed decisions? Well I tried it. Node.js 22+ and a TypeSafe API key. Jev can't generate text. It takes some state + a set of choices and returns a typed decision with probabilities over those choices. There are 95 printable ASCII characters, newline, tab, and EOS. 98 choices total. I ask Jev to judge continuations like: Each option is the entire reply prefix with exactly one character appended. The basic loop looks like this:

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

Tagscommunity jevbest catalog-import awesomejev

Works forJev for Classification

Used inJev for Software & DevTools

Languages: JavaScript

Highlights

The choice set is 95 printable ASCII characters plus newline, tab and EOS, so every option is a reply prefix with exactly one character appended.
The default mode keeps several high-probability continuations and re-asks Jev which prefix can still become the best answer, backtracking out of a bad branch.
A greedy mode follows the selected character directly; the README says it gets stuck pretty easily.
Stopping is a typed decision too: Jev picks complete, incomplete or invalid, and only complete is treated as a finished reply.

Watch out

The author frames this as an experiment in whether Choice can be composed into text, not a usable chat client. The reply only exists because the controller stitches character decisions together, and the README states plainly that Jev never returns freeform text and that no tokenizer, dictionary or canned response sits behind it. Cost tracks output length, since every character is one decision, plus a second call whenever several branches survive and a final call to judge the ending. The search mode is what makes it work at all: a prefix like "I m" can lose to an earlier one and continue as "I am". Needs Node.js 22+, a TypeSafe API key, and it serves on 127.0.0.1:4177. 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

jev-column-race

Live race rig that labels 1,000 real app reviews with Jev typed questions against Gemini JSON output on identical batches, then replays recordings and re-ranks rows locally without extra model calls.