29 September 2026: Mini Jev is now LogitTrail. Renamed to avoid confusion with the distinct r-ms/mini-jev project. URLs and API identifiers remain compatible; the submitted arXiv version and historical records keep their original names. Name and version mapping.
An inspectable local interface for typed decisions from frozen language models.
日本語 · Project page · Current research source
LogitTrail turns a state and explicit criteria into a categorical Choice, the true-candidate probability (Noul), or an expected ordinal stage (Score). Its browser workbench exposes the candidate distribution, concentration, and request/response JSON. The released native system uses a frozen local language model with no trained decision head.
Status — 27 September 2026: a preprint has been submitted to arXiv and is awaiting moderation. There is no public arXiv identifier yet. The current conference manuscript is being prepared for NAACL 2027 System Demonstrations; it is not submitted, accepted, or peer reviewed. The current research source is on research/naacl2027-demo; use that branch for the instructions below.
| Start here | What you will find |
|---|---|
| Saved evidence explorer | Compare conditions, semantic candidate probabilities, and shared calls; export traceable JSON. Added September 28; no live inference |
| Current paper PDF | Manuscript, evidence, related work, and limitations |
| 60-second evidence walkthrough | Explanation of retained experiment results; not a new live inference run |
| Full interface recording | 145-second captioned recording of the local workbench; an earlier branch caption is retained |
| Reproduction guide and frozen source/evidence ZIP | CPU-only replay of retained analysis, plus native-build verification and exact artifact hashes |
| Native installation | Build the pinned runtime, download the model, and start the local API |
| Browser workbench guide | Run live inference and inspect typed outputs and JSON |
| Reviewer guide | Model-free integrity/accounting checks and the new LMQL diagnosis |
- 1,350 / 1,350 matched pairs have identical direct-readout and native one-token labels/logits. No latency advantage was established.
- 26,050 measured requests cover the matched study and two studies of presentation sensitivity and probability averaging. This is not a count of independent questions.
- Reduced sensitivity can coexist with incorrect or nearly constant predictions. Averaging needs multiple model calls and does not consistently improve task quality.
- 56 / 56 analysis outputs were reproduced byte for byte from retained records. This is analysis replay, not an independent experiment or second-machine replication.
Task-quality evidence uses Japanese public-data subsets on one Apple Silicon Mac. Training-data contamination is unknown; no human usability study has been performed. A separate 2,400-item AI-authored regression suite is historical background. See the paper guide for the full scope.
The September 27 revision adds a comparison appendix and a post-hoc analysis of the same 12 LMQL cases; all three original probability-tolerance failures remain. The public API/UI expose typed values, candidate probabilities, and semantic metadata; candidate token IDs and raw logits are retained internally and in research traces. No new inference or human-study results are added.
git clone --branch research/naacl2027-demo --single-branch https://github.com/UpHash-Network/mini-jev.git
cd mini-jevThen follow the native installation steps. The tested system is an Apple M5 Pro with 64 GB memory and macOS 26.4. It requires Python 3.10+, Xcode command-line tools, CMake, Git, and a roughly 20.4 GB model download. Weights and native binaries are not bundled; lower-memory hardware is not validated. The CPU-only analysis replay does not require the model.
Candidate probabilities are conditional on the supplied candidates. Concentration is not the probability that the answer is correct. LogitTrail is independently inspired by TypeSafe's Jev, with no affiliation or claim to reproduce its architecture, training, calibration, or speed. Candidate scoring and expected-value scoring have prior work; the contribution here is an inspectable implementation and auditable system evaluation.
Report an issue or reproduction result · Manuscript BibTeX · Software citation
Original code and authored data: MIT. Dataset-derived records and third-party artifacts retain their applicable licenses and attribution: NOTICE.