Community · E2 · artifact verified
Make Jev talk one word at a time
A demonstration pushing the judgment-only model past its envelope: Jev 'writes' by answering which-word-comes-next in 250-word batches - each option shown as the whole reply so far plus the word - top-3 shortlist, final pick, until sentence end.
01 · Role in the system
What Jev does here
Jev never generates text, so the script generates around it: the word list (~1,100 words) splits into batches of 250 - Jev refuses more than 255 options - all batches are asked in parallel for the top 3 each, one final question picks among the ~15 shortlisted words, and the loop repeats until it chooses a period. The trick that made it coherent: every option is displayed as the WHOLE reply so far with the word appended ('hi I'm Jev nice to'), so Jev judges which phrase reads best - the kind of decision it is built for - instead of picking reply-ish lone words that fall apart grammatically. Stdlib-only Python, no install; routes through OpenRouter so the model is reachable without the native endpoint.
02 · Control boundary
Where Jev sits
Jev as a forced-choice next-word selector: batched choice questions over whole-sentence continuations, a shortlist final, and a loop to sentence end - a study of how far calibrated decisions stretch before generation becomes necessary.
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
- Routes through OpenRouter rather than the native systemone endpoint.
- Throughput is inherently low: thousands of calls per sentence.
- 1 star; a demonstration of the envelope, not a practical text generator.
04 · Attribution
Public sources
This is a Community record: the project was published by a third-party community author.
- Zaceface ↗Community · github · public · checked 2026-10-08