Span-01 vs. Jev

Span-01 represents a cutting-edge hyper-parallel reasoning classifier designed to detect complex patterns within large production traces in a single forward pass. Rather than generating tokens sequentially, it evaluates natural-language behavior definitions and outputs direct probabilities that indicate whether a specific behavior is present, absent, or unobservable. The model begins with a foundational training phase focused on general classification reasoning through Reinforcement Learning with Human Feedback (RLAIF), which is followed by fine-tuning aimed explicitly at behavior detection. This training methodology empowers Span-01 to effectively leverage previously unseen definitions for both new behaviors and diverse domains, thereby circumventing the constraints of a fixed taxonomy. During inference, its hyper-parallel definition branches are capable of assessing multiple behaviors at once against the same trace context. Additionally, the use of hybrid attention enhances its ability to reason over lengthy and complex inputs. Engineered for the scrutiny of production environments, Span-01 emphasizes safety, reliability, grounding, agent behavior, and response quality while tackling challenges like jailbreak and prompt injection detection alongside privacy and confidentiality issues. This innovative methodology not only makes Span-01 a noteworthy milestone in the domain but also paves the way for more efficient and thorough investigations into intricate behaviors, ultimately leading to improved decision-making processes. As a result, its deployment could significantly enhance operational standards across various industries.

Jev is a low-latency System One Model from TypeSafe AI built for structured decision-making rather than open-ended text generation. TypeSafe describes System One Models as a new class of frontier models intended to make fast decisions that can be consumed directly by application code. Jev takes unstructured information as input and returns predefined typed outputs with probabilities and confidence scores instead of generating unrestricted natural-language strings. Its parallel sampling architecture produces outputs in a single query rather than sequentially generating one token at a time, which is designed to reduce latency and inference cost. The model is trained with Reinforcement Learning for Calibrated Decisions, which focuses on producing well-calibrated probabilities for structured System One tasks. TypeSafe positions Jev as a software-native intelligence layer for tasks including classification, routing, scoring, extraction, decision branching, and other workflows that benefit from probabilistic logic. It can also be applied to model evaluation and safety workflows by scoring, judging, verifying, guardrailing, or detecting jailbreaks in prompts, reasoning traces, and generated outputs. The company reports service response times of roughly 70 to 500 milliseconds and says this can make Jev suitable for interactive and real-time applications where traditional frontier models may introduce too much latency. Jev is designed to return consistent schemas without type errors, allowing its results to be incorporated into code without the same parsing and validation steps commonly required for free-form LLM responses. TypeSafe also highlights map-reduce style processing over large datasets as a potential use case, where the model can turn large amounts of unstructured information into structured features and insights.