An analysis of Jev's architecture based on API probing reveals it uses a causal transformer with sparse MoE to replace text-based confidence claims with decision probabilities read directly from internal representations, enabling parallel outputs for classification tasks without generating text. The system addresses reliability and computational efficiency issues common in existing LLM-based fraud screening and moderation systems.
OpenJev is a local browser-based experiment that compares two methods for reading model choice probabilities: direct logit readout versus generation-based JSON output. Users can run both approaches on their own GPU using models like MiniCPM5 2B or Qwen3 0.6B, with real-time performance measurement and no waitlist required.