ZTC Technology Takes First Place on HF Leaderboard, Surpassing TypeSafe JEV

South Korea's ZTC (Zero-Token Classification) technology has taken first place on the official Hugging Face Typed Decisions leaderboard, surpassing TypeSafe JEV. This technology produces scores by processing the backbone model's final hidden states through a readout head without generating tokens.

Key Achievements and Performance

The Darwin-27B-ZTC model ranked first with 0.743 accuracy, KL 0.204, and Brier 0.097 on the LocalLLaMA/typed-decisions benchmark designated for October 2026, with zero errors. Accuracy by format is 0.845 for noul, 0.732 for choice, and 0.675 for score. This is higher than competing models Liquid AI d1 (0.742) and TypeSafe Jev 1.13 (0.727).

Additionally, on the S1MB (System One Mosaic Benchmark) consisting of 137 English benchmarks, Darwin-27B-ZTC-v2 ranked first with a Borda score of 89.58 and Task Avg of 66.46. It showed particular strength in the Score format, outperforming the second place by 5.38 points, and maintained an advantage across all three formats.

Technical Features and Implementation

ZTC completes decisions with a single forward pass per question, achieving a median latency of approximately 50ms on B200. It is structurally free of parsing failures and directly provides calibrated confidence (ECE, Brier, KL). The implementation is based on autojev (MIT), and inference and server code have been released.

Verification and Deployment

During the measurement process, metric definitions were verified by reproducing Uniform baselines, and fairness was ensured by increasing the limit (8,192 -> 32,768 tokens) instead of arbitrarily truncating when the input limit was exceeded. Model weights and code have been released via Hugging Face and public repositories, and have also been submitted to the Decision Index 0.3 leaderboard.