TypeSafe AI Unveils System One Models and Jev for High-Speed Structured Decision-Making

Unlike traditional large language models (LLMs) that generate text token by token, System One models utilize a new architecture, a parallel sampler, and a training method called Reinforcement Learning for Calibrated Decisions (RLCD). The Jev model is specifically optimized for structured outputs rather than string generation, which allows it to be up to 200 times faster and more efficient than existing frontier models for System One-shaped queries.

Jev provides type-safe structured values, meaning the outputs follow a predefined structure without type errors. Every response includes calibrated probabilities and confidence scores to communicate uncertainty. While Jev sacrifices general text generation, it is designed to eliminate hallucinations in structured decision-making. The model is currently available through an early access program.

PLUS ULTRAby Amenoyomi

The design of System One models is inspired by Daniel Kahneman's dual-process theory, specifically the distinction between fast, intuitive "System 1" thinking and slow, deliberate "System 2" reasoning. While traditional LLMs function as System 2 models by autoregressively generating tokens one by one, Jev is designed to act as a "frontier-intelligence function call" that handles unstructured state as input and produces typed probabilistic decisions as output.

This shift in architecture is driven by a parallel sampler and a training method called Reinforcement Learning for Calibrated Decisions (RLCD). Rather than optimizing for human preference in chat responses, RLCD optimizes for epistemically honest probabilities. Because the model generates all outputs in a single query instead of sequentially, it achieves response times between 70ms and 500ms, making it 40 to 200 times faster than existing frontier models for specific decision-based queries.

The elimination of string generation allows Jev to ensure that outputs are mathematically type-safe. By defining possible outputs and structures in advance, the model removes the risk of type errors and structural hallucinations that often occur in traditional LLM tool calls. Every output is accompanied by a calibrated confidence score, meaning higher confidence directly correlates with higher accuracy, which is a prerequisite for reliable software automation.