The ZTC-Judge-27B model is presented for evaluating the confidence of responses

Developers introduced ZTC-Judge-27B — a specialized model designed for automatic evaluation of the quality of other neural networks' responses. The tool uses the zero-token confidence method, allowing ratings to be assigned within a single forward pass. The model demonstrates a 0.7282 AUC score on previously unseen domains. During testing, the system showed an advantage over standard surface baselines in five different task categories. Evaluating one response takes about 1 hour, with the model achieving a result of 14 points on the internal efficiency scale. The ZTC-Judge-27B architecture is optimized to work as an independent judge that does not require additional iterations or complex prompts to verify content. The source code and model weights are already available for download on the Hugging Face platform. Implementing such tools allows for the automation of quality control of generative systems without involving human resources. Using the model in development pipelines helps to more accurately identify hallucinations and errors in LLM responses, ensuring stricter content filtering in real-time.