LCT
Open decision models that derive calibrated confidence from recurrent looped-layer dynamics instead of RL, with early exit and a Jev API.
Overview
Jev-LCT (Looped Calibration Transformer) is an open-source family of small System One decision models aligned with Jev's 'decisions, not strings' idea. Rather than reinforcement learning for calibration, it recurrently recomputes only the top two layers of a causal transformer backbone and extracts confidence from internal trajectory dynamics: how hidden states converge, how decision entropy falls and how the softmax margin evolves across loops.
An adaptive early exit lets easy queries finish on the first loop when the margin is at least 0.90, which the README says covers 85% or more of simple queries in about 45 ms, while ambiguous ones iterate up to four loops until convergence. It supports Jev's POST /v1/systemone format with choice, noul and score through a FastAPI server and a typed Python SDK. Weights are on Hugging Face, including a 0.49B Qwen2.5-based model reported at 50.8 ms on an RTX 3090 Ti, as full models or adapters. Documentation is in English and Chinese.
What the decision model decides
| Type | Question | Answers |
|---|---|---|
| Choice | Which option fits the state? | — |
| Yes / no | Is this statement true? | — |
| Score | Where does the state fall on the scale? | — |
How it works
- Send state and typed questions to the /v1/systemone server
- Model runs a first pass and checks the decision margin
- Easy cases exit early; hard cases loop up to four times
- Confidence is computed from trajectory dynamics
- Calibrated probabilities are returned
Strengths
- Calibration without RL training
- Adaptive compute via early exit
- Jev-compatible API
- Small models for edge devices
Caveats
- Novel method; independent validation not shown in the excerpt
- Latency figures from one GPU
- Research-stage project
Getting started
Download a model from huggingface.co/CaoHaoWei and run the FastAPI server described in the README.At a glance
- Type
- Project
- Status
- research
- Language
- Python
- Runtime
- PyTorch with FastAPI server
- Has code
- yes
Use cases and integrations
How curated lists describe it
Jev-LCT: Open System-One Decision Engine with Free Calibrated Confidence from Recurrent Trajectories.
hellogumbo/awesome-jev
Frequently asked questions
How is confidence calibrated?
From recurrent hidden-state dynamics across loops rather than RL training.
Is it compatible with Jev clients?
Yes, it supports the /v1/systemone format.
How fast is it?
The README reports about 50.8 ms for the 0.49B model on an RTX 3090 Ti, with many easy queries exiting early.
Related projects
Exact-target audit finding Jev's Choice probabilities are a step function on stated base rates while per-option Noul tracks them closely.
ProjectresearchPythonResearch replicating Jev-style typed decisions with open LLMs via logprobs; Qwen3.5-9B reached 94% with good calibration in ~1 s.
ProjectresearchPythonLocal reproduction of Jev-style inference that reads option-token logits from a small LLM in one pass, with an honest failure log.
ProjectresearchPythonLaptop-scale reimplementation of RL for Calibrated Decisions: parallel constrained decoding trained with GRPO and a Brier reward.
ProjectresearchPython- AnyJev★ 1,255
Nokia's training-free library that turns open LLMs into Jev-style deciders robust to option order, plus self-distilled Tacit models.
ModelresearchPython Jev-style replica: frozen Qwen3-1.7B plus a ~0.8M-parameter trained head returns calibrated Choice, Boolean and Score distributions.
ServiceresearchPython