ZTC scores AI action confidence from model internals with zero extra tokens

A new method called ZTC (Zero-Token Confidence) estimates the reliability of AI agent actions by reading the model's internal state instead of calling a second LLM judge. It reaches an AUC of 0.7289 versus 0.7350 for the JEV judge, while deciding in 0.06 s instead of 0.59 s and generating no extra tokens.

• ZTC AUC 0.7289 vs 0.7350 for the JEV LLM judge — under 1% gap
• Latency 0.06 s vs 0.59 s — nearly 10x faster per decision
• No extra tokens generated; runs on premise without external services
• Uses token probability distributions, representation geometry and attention patterns