Chat without generating any free text

01 · Role in the system

What Jev does here

Jev answers typed questions with at most 255 options, so it cannot pick from 30,000 words directly. The word tree splits the vocabulary by meaning - each word glossed in one line by a cheap model, embedded, then recursively clustered - into 254 top groups. For each word of the reply, Jev picks a group (shown as its most common words), keeps the top three, then picks the best word inside each; the reply ends when "the reply is complete" outranks every word. Two deterministic anti-loop rules stop greedy repetition. The verdict is honest: short replies work (the author reports about $0.003 and 2.6 seconds each), longer ones fall apart into word salad - one calibrated decision at a time is what Jev is built for, not a chain of dependent ones.

02 · Control boundary

Where Jev sits

The vocabulary is clustered offline into choice-sized groups; at reply time each word is a group Choice then a word Choice, with deterministic anti-loop rules and a completion option.

Code owns the loop, permissions, thresholds, validation, and side effects. Jev owns only the bounded judgments described above.

03 · Known limits

What this evidence does not prove

• The author's own finding: replies beyond a few words fall apart into word salad, because Jev is built for one calibrated decision at a time, not chains of dependent ones.
• Calls go through OpenRouter (model typesafe/jev-1.13) rather than the native TypeSafe endpoint.
• Per-reply cost and latency figures are author-reported.

04 · Attribution

Public sources

This is a Community record: the project was published by a third-party community author.

• finetuningsingh ↗Community · github · public · checked 2026-09-20
• finetuningsingh ↗Community · demo · public · checked 2026-09-20