33 online·4645 visitors
COMMUNITY OBSERVED

LogJev

An open-source Jev-style decision layer that accepts text, images, or audio and returns bounded choices or scores.

CODEX / CLAUDE CODEOFFICIAL TYPESAFE SKILLWORKFLOW CODE + DOCS

Overview & Result

MajoSayo describes LogJev as a Node.js and Python implementation for Jev-style bounded decisions using the author's own models, with text, image, or audio input and choices or scores from logprobs.

How Jev fits in the loop

  1. The application supplies a bounded decision question and its input state.
  2. The decision layer scores the available choices or outputs a score.
  3. The surrounding agent or application consumes the typed result and owns the next action.

How to reproduce

  1. Open the repository and verify the current Node.js and Python setup instructions.
  2. Run one harmless text classification with a fixed option set and inspect the returned choice or score.
  3. Only then test image or audio inputs if the selected model supports the required modality and logprobs.

Why this build matters

It gives JevForAgents a concrete local/open implementation to compare with hosted decision APIs, while keeping the evidence boundary explicit.

Limitations

  • The reviewed post does not provide a benchmark or a reproducible latency measurement.
  • Support depends on the selected model exposing the required modality and logprob behavior.
  • Local-model parity with TypeSafe Jev is not established by this post.

Patterns