TypeSafe AI Unveils Jev, a Structured Decision Model

On September 15, 2026, TypeSafe AI announced System One Models, a new family of foundation models, and Jev, the first publicly released model. It is available as an early release. Founder Diogo Almeida previously worked at OpenAI on developing instruction-following methods, leading research that led to ChatGPT. He says he has spent four years asking why automation has not advanced even though conversational ability has surpassed humans. According to reporting by albelfio on Hacker News (Best), this concern is presented as the starting point for the new architecture.

After about two years of private research, the company has built a technology stack specialized for automation. At its core are a new model architecture and parallel generation mechanism, along with RLCD, described as reinforcement learning for calibrated decisions. Jev is said to match existing large language models in intelligence on System One tasks while achieving a two-orders-of-magnitude improvement in speed and efficiency. Instead of generating strings, it is optimized for structured outputs.

unstructured state in, typed probabilistic decisions out.

The above is the phrase the company uses to describe how Jev works. It takes unstructured state as input and outputs typed probabilistic decisions. It is positioned as a function call with frontier-level intelligence.

A New Architecture Aiming for Automation

Beyond Chat

Existing large language models have aimed for outputs that match human preferences or can be measured with verifiable rewards. RLHF and RLVR are representative examples. Inputs are unstructured materials centered on text, and outputs are strings. Strings are a flexible format that can include conversational responses, code, erroneous generations, refusal responses, and structured values. To use them in business infrastructure, separate parsing and verification are required, leaving a risk of unexpected behavior.

System One and Jev place the optimization target on calibrated decisions. They share the handling of unstructured materials as input, but emphasize structured operational states such as planning state. Outputs are limited to predefined, type-safe structured values. There are no type errors, and every answer comes with a calibrated probability and confidence. The company describes it as designed not to hallucinate.

Regarding developers’ working environments, confusion over font handling has drawn attention. As Intel One Mono Revival Retracts Discontinuation in 2 Days shows, small changes in development infrastructure directly affect the field. Stability of output formats is a similarly important issue. Jev’s type safety is positioned as an attempt to reduce uncertainty in this layer.

Structural Shift from Sequential Generation

to Parallel Inference

Inference in existing methods is sequential. They generate one token at a time, conditioning the next decision on previous results. This approach suits generating long texts, but latency accumulates. It also constrains utilization of computing resources.

Jev’s inference is parallel. It generates all outputs in a single query. Through hardware-aware optimization, it claims a two-orders-of-magnitude speed improvement. The pricing table in the public materials lists input tokens at $0.20 to $10 per million tokens, and output tokens at about 5x the input price. Previously, the cost of output tokens pushed up the total. It aims to transform the cost structure through parallelization and structuring.

Interest in low-level efficiency is growing across operational infrastructure. Linux 7.2 Kernel Speeds Up /proc/filesystems Reads by up to 444% is an example of how fine-grained improvements at the foundation ripple through overall performance. Model inference mechanisms are moving in the same direction. The idea is to avoid token-by-token sequential processing and handle decisions in bulk.

How Type-Safe Outputs and Calibrated

Probabilities Work

Jev’s outputs predefine possible values and structures. Business processes receive values directly according to the definitions. Parsing and syntax repair for retries become unnecessary. Type consistency is guaranteed by the model architecture.

The other pillar is calibrated probabilities. Each answer is assigned a probability and confidence. This means epistemic honesty, showing low confidence when it does not know. In automation, routing uncertain cases matters more than getting every case right. If probabilities are calibrated, the boundary between automated processing and human review can be designed numerically.

The company describes RLCD as a training method for calibrated decisions. Rather than human preferences or verifiable correct answers, it optimizes the correctness of probabilities on System One tasks. Detailed training materials and evaluation criteria are limited within the scope of public materials. Additional information is needed for third-party reproduction.

Impact on Developers and Early Access Form

Jev is used as a component that returns decisions rather than strings. It receives unstructured state and returns typed decisions. From a business process perspective, it is close to a function call. The step of parsing dialogue text to create branching conditions becomes unnecessary. Contamination by erroneous generations and exception handling for refusal texts are also reduced.

In early access, use cases are expected to be narrowed down. Targeted tasks are those where the output space can be defined, such as approvals, classification, routing, and resource allocation. It is not suited for open-ended text creation where the output space cannot be defined. This is the cost of abandoning string generation. Appropriate, selective use is assumed.

Pricing and offering design will also affect adoption. As Meta Pauses Conversation Focus Billing shows, changes in billing structures directly affect user decisions. Jev appeals to cost advantages based on efficiency. Actual total costs will vary with input scale and call frequency. Measurement in pilot deployments is essential.

Remaining Challenges and Issues to Verify

Verification of the claims is still to come. It matters on which task groups the two-orders-of-magnitude speedup and equivalent intelligence level were measured. The definition of System One tasks, the names of comparison models, and measurement conditions cannot be sufficiently identified from public information alone. Independent evaluation is awaited.

The claim that it does not hallucinate also needs interpretation. Even if type correctness is guaranteed, factual correctness is a separate issue. The possibility of making a wrong choice among defined options remains. Whether calibrated probabilities can absorb errors is key in practice. Disclosure of calibration curves and operational materials is needed.

On the operational side, auditing and allocation of responsibility need to be organized. When a business process adopts a decision with probabilities, responsibility for misjudgments is distributed between designers and operators. It is necessary to define the basis for setting thresholds and how records are stored. When combined with existing language models, the boundary design between the string-generation part and the decision part also becomes an issue.

Editorial Opinion

In the short term, we expect competition in inference infrastructure specialized for structured outputs to intensify. Designs premised on string generation and downstream verification will be forced to reconsider. In 3 to 6 months, pilot deployments of business automation will progress centered on function-call-type use.

In the long term, we expect decision components that return calibrated probabilities to be embedded in business infrastructure. Methods for managing automation failure conditions with probabilities will spread. For users and society, developing accountability and audit frameworks will become the challenge.

The remaining issue is the scope of verification for the claim that no hallucinations occur. Type correctness and factual correctness are separate issues, and error rates in operation cannot be judged from public materials alone. The question is whether probability calibration is at a level that can withstand practical decision-making.

References

• “Introducing System One Models and Jev”, by albelfio — Hacker News (Best), 2026-09-15T19:25:03.000Z (ARR)
• Source URL: https://typesafe.ai/blog/introducing-system-one-models-and-jev