mini-jev
LogitTrail (formerly Mini Jev): local workbench that reads typed Choice, Noul and Score decisions from a frozen LLM's logits.
Overview
LogitTrail, renamed from Mini Jev in September 2026, is an inspectable local interface for typed decisions from frozen language models. Given a state and explicit criteria, it produces a categorical Choice, a true-candidate probability (Noul), or an expected ordinal stage (Score). A browser workbench shows the candidate distribution, its concentration and the raw request/response JSON. The native system uses a frozen local model with no trained decision head.
It is a research project: a preprint was submitted to arXiv and a manuscript is being prepared for NAACL 2027 System Demonstrations, not yet peer reviewed. The README reports that 1,350 of 1,350 matched pairs gave identical direct-readout and native one-token labels and logits, that no latency advantage was established, and that reduced sensitivity to presentation can coexist with incorrect or near-constant predictions. A saved-evidence explorer, a reproduction ZIP and a reviewer guide are provided. It suits researchers studying how faithful and stable logit-based decisions are.
What the decision model decides
| Type | Question | Answers |
|---|---|---|
| Choice | Which candidate best satisfies the stated criteria for this state? | — |
| Yes / no | Probability that the candidate is true | true · false |
| Score | Expected ordinal stage | — |
How it works
- Build the pinned native runtime and download the model.
- Send a state and criteria to the local API or browser workbench.
- Read candidate probabilities directly from the frozen model's logits.
- Inspect distribution, concentration and JSON; export traceable evidence.
Strengths
- Fully local inference with inspectable distributions.
- Reproduction package with exact artifact hashes and CPU-only replay of analysis.
- Honest reporting of negative findings, such as no established latency advantage.
Caveats
- Research status; the paper is not peer reviewed.
- Instructions live on a separate research branch, not main.
- No trained decision head, so accuracy depends on the frozen base model.
Getting started
Use the research/naacl2027-demo branch and follow its native installation guide to build the runtime, download the model and start the local API.At a glance
- Type
- Project
- Status
- research
- Runtime
- local native service, browser workbench
- Has code
- yes
Use cases and integrations
How curated lists describe it
LogitTrail (formerly Mini Jev): inspectable local typed decisions, candidate…
wh000wh000/awesome-jev-live
Frequently asked questions
What is the difference between LogitTrail and Mini Jev?
It is the same project renamed to avoid confusion with r-ms/mini-jev; URLs and API identifiers stay compatible.
Does it train a decision head?
No. The released native system reads decisions from a frozen local language model.
Can I try it without installing?
The project page offers a saved-evidence explorer and recordings, but they show retained results rather than live inference.
Related projects
Local Jev-style decision model on Qwen3.5-4B with four stress tests that show where one-pass decisions break.
ProjectresearchPythonRunnable yardstick that tests Jev's three headline claims on a small local model: typed output, no hallucination, calibration.
ProjectresearchPythonpytest-style tool that pins fixtures and expected Jev answers, then fails a model upgrade when answers flip or confidence drops.
ProjectbetaPythonTest harness for any Jev-API decision model: accuracy, calibration and robustness probes for reordering, injection and padding.
ServicebetaPythonStudy of Jev's choices in multi-armed bandits, testing whether Bayesian instructions and supplied statistics improve its exploration.
ProjectresearchExperiments testing whether Jev can spot positive-expected-value dice bets from payout tables, and whether recent wins or losses sway it.
Projectresearch