JevSpeak
GitHub ↗

Jev can't generate text.
So I made it talk anyway.

JevSpeak 0.1

A conversational interface built from probabilistic decisions, deterministic language composition, and no generative LLM.

Language is the rendering layer. Jev makes the decisions.

Talk to Jev →How it works ↓
How a reply is produced
JevSpeak 0.1 — pipeline▤
  1. 01 USER"Will AI replace programmers?"
  2. 02 JEVintent question 100% · stance mixed 68% · claim replace_tasks 70%
  3. 03 SEMANTIC IR{ stance: "mixed", confidence: 0.68, mainClaim: "replace_tasks", qualification: "context_dependent" }
  4. 04 LANGUAGE COMPILERstance(mixed) → "yes and no" + claim template + qualification clause
  5. 05 JEV SAYS"Yes and no. AI is going to take over some programming tasks, not the whole job, though it depends on the specifics."

*The words are produced by code. The meaning is chosen by Jev. No GPT / Claude / Gemini model runs at any point in this pipeline.

Decisions, Not Tokens
Jev answers 13 semantic questions in parallel with probability distributions — intent, stance, claim, tone…
Deterministic Compiler
Plan slots → pick phrases by seed → grammar. Same IR, same sentence, every time.
Calibrated Wording
0.54 → “maybe”. 0.86 → “probably”. Below 0.50 Jev declines instead of bluffing.

Meaning by Jev.
Words by code.

*Same decisions, two languages: switch EN / 中文 and only the rendering layer changes.