Speak in sentences compiled from decisions

01 · Role in the system

What Jev does here

Each turn, one request asks Jev roughly thirteen parallel questions - intent, topic, emotion, which speech act fits, which way a judgment leans, which finite claim, what caveat, what tone, how long - each answered as a distribution over finite options plus scores for confidence and emotion intensity. Those answers normalize into a typed Semantic IR, and a deterministic compiler assembles the reply from layered phrase banks: acknowledgement, short answer, claim with hedge, qualification, follow-up. Confidence is made visible in the wording itself - below 0.5 Jev declines or asks which part you mean, 0.6 becomes "I think", 0.97 becomes "yes" - so the model cannot bluff. The README's cleanest evidence that language is only a rendering layer: switch the locale pack and the same decisions come out in Chinese, with nothing about the questions, IR, or planner changed.

Language as the rendering layer

JevSpeak removes the generative model from conversation entirely. Each turn, one request asks roughly thirteen parallel questions — intent, topic, emotion, speech act, judgment lean, finite claim, caveat, tone, length — each answered as a distribution over finite options. The answers normalize into a typed semantic IR, and a deterministic compiler assembles the reply from layered phrase banks: acknowledgement, short answer, claim with hedge, qualification, follow-up.

Confidence becomes grammar: below 0.5 the system declines or asks which part you mean, 0.6 surfaces as I think, 0.97 as yes — so the model cannot bluff, because hedging is a rendering decision driven by the calibrated number. The cleanest proof that wording is only a renderer: switch the locale pack and the same decisions come out in Chinese, with questions, IR, and planner untouched.

Bounded by design, and honest about it

What can be said is bounded by the phrase banks and the IR vocabulary — the record names that as the system's ceiling, not a bug. Conversation memory is structured state over the last N turns, not open-ended recall, and the no-generative-model claim covers the reply path specifically; the author used Claude as a development tool to build the codebase, which the README discloses.

Read against the index's two word-choice experiments, this record completes an argument: five words works, 30,000 words collapses into word salad, and the middle path is to compile language from a typed IR — decisions decide meaning, templates decide words.

02 · Control boundary

Where Jev sits

One batched request of parallel questions per turn; answers normalize into a typed Semantic IR and a deterministic, locale-swappable compiler renders it - templates decide wording, the IR decides meaning.

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

• Templates decide wording, so what can be said is bounded by the phrase banks and the IR vocabulary.
• Conversation memory is structured state over the last N turns, not open-ended recall.
• The README states no generative model runs in the reply path; the author used Claude as a development tool to build the codebase.

04 · Attribution

Public sources

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

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