Published
Updated
What Jev returns: System One in plain language
This topic was created 28 days ago, and the information it contains may have evolved or changed since then.
TypeSafe defines a System One model as parallel, typed, and calibrated decisions without text generation, distinguishing it from language models operating in JSON mode. On October 3, 2026 the confidence page prints the Choice and Score formulas. Noul still has no separate confidence field.
returns
Confidence on Choice and Score comes from the returned probabilities. Noul returns a probability between 0 and 1, without a separate confidence.
Rather than functioning like conversational chatbots, Jev operates more like a conditional branching statement that evaluates unstructured input.
TypeSafe distinguishes standard language models from what it calls System One models. Standard models generate sequential token streams, requiring client code to parse and validate output even in structured JSON modes. Jev requires callers to declare valid outputs in advance, restricting responses to predefined options, numeric scores, or boolean flags.
The system exposes three primary primitives:
| Primitive | What you ask | What you get back |
|---|---|---|
| Choice | Pick among named options (cardinality up to 255) | A distribution over those options |
| Score | A numeric judgment on a scale you define | A value plus confidence |
| Noul | A yes/no (the docs’ name for a boolean question) | A probability |
Multiple questions can be evaluated in a single request. TypeSafe states that sampling occurs in parallel rather than autoregressively, citing this architecture for claimed latencies of 70 to 500 ms compared to multi-second chat completions. Input is priced at $0.042 per million tokens, while output evaluations are provided without charge.
TypeSafe describes its training methodology as RLCD (Reinforcement Learning for Calibrated Decisions). While conventional RLHF optimizes text for reviewer preference, RLCD aims to calibrate reported confidence against empirical accuracy. This calibration allows applications to automate high-confidence actions while routing uncertain cases to review.
The company argues that schema conformity should be guaranteed rather than measured probabilistically. Predefining valid output shapes prevents issues like hallucinated tool calls, unexpected JSON keys, or unlisted options.
These operational constraints mean Jev does not produce prose, generate code, or provide explanatory reasoning chains. Almeida noted that latency and cost improvements stem directly from removing text generation. Applications requiring free-form text still rely on traditional language models, while Jev is intended for classification, routing, and decision checks.
The System One terminology draws on Daniel Kahneman’s cognitive framework dividing fast intuitive processing from deliberative analysis. TypeSafe positions Jev as a rapid reflex layer for structured application logic, reserving generative language models for tasks requiring compositional text.
On October 3 TypeSafe linked the confidence page. Choice confidence is (p_max - 1/n) / (1 - 1/n). With three options that is (3 * p_max - 1) / 2. The page’s pair of examples, (0.6, 0.3, 0.1) and (0.6, 0.2, 0.2), both come out at 0.4. The second-place mass does not enter the formula.
Score confidence is the spread around the chosen level, divided by the spread of a uniform distribution over the same levels, then subtracted from 1 and floored at 0. On three levels, (0, 0.5, 0.5) is 0.25 and (0.5, 0, 0.5) is 0. The page’s bug-severity example, (0, 0.57, 0.43), is about 0.35. The same page also names two simpler readings, the top probability and the ratio of the top two, and it says Noul has no confidence field. One optional reading of a Noul probability p is the absolute value of 2p - 1. A threshold sketch on the page, 0.5 as a floor and 0.9 for a transfer, is an illustration. It is not a published accuracy.
This site's reading
Editorial notes evaluating claims against primary sources, contextualizing findings alongside related implementations, and defining technical terms.
Verify
Choice, Score, and Noul align with TypeSafe's primitives documentation. Pricing at $0.042 per million input tokens with free output matches the models page. Parallel evaluation and 70 to 500 ms latency are company claims.
Confidence metrics on Choice and Score are calculated from the resulting probability distribution rather than measured external accuracy; Noul does not provide a separate confidence score.
Documentation states English is the primary training language, input is restricted to text, and weights are not fine-tuned per customer. We did not query the API.
TypeSafe, @typesafeai, October 3, 2026, 01:38 UTC, status 2106197282274324846, links the confidence page. The reply at 01:39 UTC is status 2106197453632659673.
docs.typesafe.ai/confidence, opened for this note, prints Choice confidence as (p_max - 1/n) / (1 - 1/n). For three options that is (3 * p_max - 1) / 2. The distributions (0.6, 0.3, 0.1) and (0.6, 0.2, 0.2) are both 0.4.
Score confidence is max(0, 1 - spread / MAD_unif), where the spread is the probability-weighted distance from the chosen level. For three levels, (0, 0.5, 0.5) is 0.25 and (0.5, 0, 0.5) is 0. The page's bug_severity example (0, 0.57, 0.43) is about 0.35.
Noul has no confidence field. The page offers the absolute value of 2p - 1 as an optional reading, and it also names the top probability and the top-to-second ratio. The 0.5 floor and the 0.9 transfer example are illustrations, not a measured eval.
Compare
JSON-mode language models generate tokens sequentially before parsing. Traditional classification models return labels and scores, but do not share this training architecture or multi-question interface.
LayerX CTO Yuki Matsumoto compared the System One label to Figure Helix's 80M System 1 transformer for robot reflexes. This comparison is conceptual (contrasting rapid reflexes with deliberative planning) rather than a claim of robotics capabilities, as Helix employs a visuomotor network while Jev operates as a text-in, typed-decision service.
The October 3 confidence page writes the Choice and Score formulas down. A three-way Choice at (0.6, 0.3, 0.1) and one at (0.6, 0.2, 0.2) both score 0.4, because the formula uses only the top probability and the option count. Noul still returns a probability and no second field.
Terms
- Choice
- Selecting one option from a user-defined set (cardinality up to 255), returning a probability distribution and confidence score.
- Score
- A numeric evaluation based on a custom rubric, returning an estimated position, distribution, and confidence level.
- Noul
- TypeSafe's term for a boolean question, returning a probability score between 0 and 1 without a separate confidence field.
- Calibration
- The degree to which a model's reported probability corresponds to empirical label accuracy, which TypeSafe targets via RLCD.
- Parallel sampling
- Evaluating multiple declared questions over a single input state simultaneously without sequential token generation.