Crate jev_bridge 

Bridge a generic OpenAI-compatible inference API into a Jev-style decision scoring service. Jev-style decisions are not a model format; they are a readout. This crate implements that readout on top of any runtime that can return next-token log probabilities:

1. render the direct-options-v1 prompt through the serving runtime’s own chat template, with thinking disabled;
2. read the log probabilities of the fixed uppercase answer letters at the final position;
3. softmax only over the declared options, so no other vocabulary can leak into the distribution;
4. answer in the System One wire format with choice, noul and score questions.

No model weights are copied and no answer text is generated: a decision costs one prompt evaluation and zero output tokens.

Layout

• strategy probes the upstream service and parses the response shapes different runtimes use for the same distribution.
• prompt owns the prompt contract, the single-token answer-slot check and the softmax.
• server holds server::Bridge and the HTTP surface.
• wire validates System One requests and builds their responses.
• evaluate turns --score predictions plus a gold file into offline calibration metrics (accuracy, NLL, Brier, ECE, reliability curve) and recomputes a written report for verification.
• readout runs the startup self-check: a few questions with obvious answers, read through the same slot channel every request uses, so a model that is not answering the question at all is visible instead of silently confident.

The jev-bridge binary is a thin CLI over server::Bridge.

Documentation

Every public item is documented, and the crate builds with #![warn(missing_docs)] so that stays true.

Modules

evaluate

Offline calibration evaluation for --score predictions.

perturb

Perturbation flips: a decision that changes when nothing relevant changed is not a decision about the evidence.

prompt

The direct decision prompt contract and the single-token answer-slot check.

readout

The readout self-check: does this model actually answer with slot letters?

render

Local chat-template rendering.

server

The bridge: one resident upstream model behind a Jev-style scoring API.

strategy

Runtime probing of a generic OpenAI-compatible inference service.

tokenizer

Local tokenization through the Hugging Face tokenizers crate.

wire

System One wire format: request validation and response construction.