The latest articles on DEV Community by CaoHaoWei (@caohaowei). https://dev.to/caohaowei https://media2.dev.to/dynamic/image/width=90,height=90,fit=cover,gravity=auto,format=auto/https:%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Fuser%2Fprofile_image%2F4166339%2F15cc10bc-fff7-4b97-bb07-a6274e82e7cd.png https://dev.to/caohaowei en CaoHaoWei Tue, 06 Oct 2026 11:57:59 +0000 https://dev.to/caohaowei/jev-lct-free-calibrated-confidence-from-looped-transformer-trajectories-2ooo https://dev.to/caohaowei/jev-lct-free-calibrated-confidence-from-looped-transformer-trajectories-2ooo <p>Small decision models have two bad options today: generate text token-by-token (slow, fragile to parse), or "introspect" their confidence verbally (systematically miscalibrated). Jev-LCT is an open-source System-One decision engine that takes a third route — extracting calibrated probabilities directly from a transformer's internal recurrent dynamics.</p> <h2> How it works </h2> <ul> <li> <strong>Parallel Looped Prefill</strong>: recurrently recomputes only the top k=2 layers of a causal transformer, preserving full representation fidelity.</li> <li> <strong>Endogenous Trajectory Confidence</strong>: reads calibrated probabilities from hidden-state convergence dynamics (cosine delta, entropy reduction, softmax margin trajectories) — no RL, no extra generation tokens.</li> <li> <strong>Adaptive Dual-Channel Early Exit</strong>: 85%+ of simple queries exit after a single loop in ~45ms; ambiguous ones iterate up to 4 loops.</li> </ul> <h2> Benchmarks </h2> <p>Measured on an RTX 3090 Ti, across a 300-item benchmark (Banking77, ARC, TruthfulQA, BoolQ, MMLU):</p> <ul> <li>Jev-LCT-Qwen2.5-0.5B — 50.8ms</li> <li>Jev-LCT-Qwen2.5-1.5B — 61.9ms, ~70% overall accuracy (recommended balance)</li> <li>Jev-LCT-Qwen3-8B — 89.2ms</li> </ul> <p>It's compatible with TypeSafe AI's Jev <code>POST /v1/systemone</code> spec and ships with a FastAPI server plus a typed Python SDK. Apache 2.0, 52/52 tests passing, weights on HuggingFace. Built for agent decision-making, security routing, fraud detection, and edge robotics — anywhere you need a fast decision instead of a paragraph.</p> <p>I'm an independent developer — try it out, and issues are welcome: <a href="https://github.com/gitchw/LCT" rel="noopener noreferrer">https://github.com/gitchw/LCT</a></p> machinelearning opensource python ai