System One

Clairic positions itself as a "System One extraction model" that returns typed values with confidence scores in a single forward pass.

System One is defined by Clairic as an entirely new model category focused on extracting information from data.

The commit message states that it implements a decision scorer powered by Qwen3.5, including non-causal attention, publisher readout, image processing, and System One prompt encoding.

The authors validated on six datasets, covering three types of tasks—content moderation, safety filtering, and document reranking—and the comparison targets include a general System One model, task-specific baselines, and a fine-tuned System One model with locally open-sourced weights.

The framework proposed in the study contains two key mechanisms: conservative fallback and risk-budget selection. Conservative fallback means defaulting to a safe but utility-conservative action when the model is uncertain; risk-budget selection means imposing a constraint on the false omission risk over the subset of inputs assigned to the utility-maximizing action.

The paper's core argument is: relying on probabilistic interfaces to drive deployment decisions is insufficient, and in scenarios lacking a human fallback mechanism, what truly needs to be controlled is the false omission risk on the subset of inputs for which the utility-maximizing action has been taken, rather than classification accuracy or AUC.

The performance differences among the general System One model, the task baselines, and the open-source fine-tuned model further point to one conclusion: improvements in model capability cannot replace risk control design at the decision layer.

In settings where the test risk is definable, the naive threshold-selection baseline exceeds the target risk level in about half of the cases, whereas the Learn Then Test (LTT) method exceeds it in only 7.8% of the cases.

Jev 1.14's System One model adds experimental image input capability.

The derivation gap appeared in all seven judge models the researchers evaluated, including the System One series' CLM and Jev.

System One scoring typically refers to a fast score given in a single forward pass, without involving multi-step reasoning, self-reflection or chain-of-thought.

The concepts of System One and System Two come from Daniel Kahneman's Thinking, Fast and Slow; TypeSafe borrows the concept to distinguish two model invocation paths within an agent.

The new version requires that a model must explicitly possess the decision capability before it can be scheduled to handle System One requests.

In the previous implementation, Ollama determined the scheduling of System One requests by matching the Qwen architecture and its renderer metadata.

eu/jev claims to be the first Jev-class (System One) model fully hosted within the EU and compliant with GDPR requirements.

Document processing can adopt a division-of-labor collaborative mode: using lightweight, fast System One models for high-frequency basic judgments, and reserving heavy inference resources for complex tasks

The System One model excels at four types of document tasks that require rapid decision-making: orientation detection, language detection, classification, and splitting

The System One model borrows from the dual-system theory in psychology, representing an intuitive, fast decision-making mode, in contrast to the deliberate System Two

Jev is an open-source System One lightweight model designed to handle rapid document decision-making tasks

System One models generally refer to lightweight models with relatively few parameters, shorter reasoning chains, and extremely low response latency, in contrast to slow-thinking reasoning models such as o1 and DeepSeek-R1.