The Department of Information Engineering at the Faculty of Economics and Management of the Czech University of Life Sciences Prague is monitoring and providing expert support for the deployment of the JEV model at the Mironet.cz e-shop. This is one of the first cases in Europe in which a model deliberately designed not to generate text is entering commercial operation. For the public, it is an opportunity to understand why part of AI research is shifting away from language generation and back towards well-measured uncertainty.
A model that does not answer in sentences
Large language models such as ChatGPT, Gemini, or Claude generate responses word by word — each step must wait for the previous one. This makes them highly versatile but, for some tasks, unnecessarily expensive and slow.
The JEV model, introduced by the US company TypeSafe AI on 15 September 2026, takes a different approach. It does not produce a sentence, but rather an answer to a precisely defined question: choose one of several options, rate something on a scale, or answer yes or no. Each answer is accompanied by a probability distribution — not only “this product belongs in the laptop category”, but also “with 92% probability”. In addition, all submitted questions are evaluated simultaneously, in a single pass. This is why response times are measured in tenths of a second and costs are significantly lower: because the model does not generate text, output tokens — which account for the more expensive half of the bill with conventional models — are not generated at all.
The name pays tribute to economist William Stanley Jevons and the paradox associated with him, according to which making a resource cheaper does not necessarily lead to savings, but rather to more intensive use.
Why this is scientifically interesting — and what is not actually new
It is tempting to describe JEV as “an entirely new kind of AI”. A more accurate interpretation is that it is an unusually well-packaged combination of ideas that science has known for decades. Its roots go back to decision-making with the option to reject an answer — a theoretical foundation laid by C. K. Chow as early as 1970, and later applied to neural networks by Geifman and El-Yaniv in 2017. A system does not always have to provide an answer; if it is not sufficiently confident, it may abstain and pass the case to a human.
The second pillar is probability calibration. A model is calibrated when its numbers mean what they claim to mean: if it says “90%” one thousand times, the prediction should be correct in roughly nine hundred of those cases. Conventional language models generally lack this property, and there is a systemic reason for that: they are trained on human feedback, and people tend to prefer answers that sound confident. The training pressure, therefore, rewards confidence and penalizes explicit uncertainty — and the fact that fine-tuning based on human preferences can worsen calibration was also documented in the technical report on GPT-4.
The company states that JEV was trained using a method called RLCD, which, rather than optimizing for human preferences, directly optimizes the quality of probability estimates. However, no scientific paper describing RLCD has been published yet, and the company has not released either the objective function or the source code. The method, therefore, cannot currently be independently reproduced.
The numbers: who measured what
Three sets of figures are publicly available, and they refer to different things.
| Source | Speed-up | Cost reduction | Note |
| TypeSafe AI (internal tests) | 193,6× | 444,6× | Proprietary tasks; reference based on agreement between two other models, and an incorrect response. The company itself states that these figures represent “the upper end of realistic gains”. |
| Mironet (real-world operation) | 19× (32,4 › 1,7 s) | 343× | Measured on its own categorization task, compared with the previous solution. |
| Independent comparison (JevBench) | — | ~6× na 1 000 decisions | Compared with a conventional fast model; task composition was determined by the benchmark authors. |
The fact that the speed-up measured by Mironet is roughly ten times lower than the figure highlighted by the vendor is not a contradiction, but a normal result. Laboratory benchmarks are optimized to expose differences, whereas real-world deployment includes everything else. Order-of-magnitude improvements presented in marketing materials usually shrink in production — and may remain highly significant. What can be verified without qualification is the pricing: USD 0.042 per million input tokens, with output free of charge.
One misunderstanding worth clarifying
Models of this kind are often described as being unable to “hallucinate”. That is only partly true, and the distinction is crucial in practice.
The model cannot invent a category that does not exist — it can only choose from the options it is given. That ensures the reliability of form. But nothing prevents it from confidently assigning a product to an existing yet entirely incorrect category. The error shifts from form to content, where it may be harder to notice because the output looks perfectly valid.
Independent measurements conducted during the first week after release confirm this and add three warnings: calibration quality varies from task to task, the model will not abstain unless an “I don’t know” option is explicitly provided, and accuracy depends noticeably on how the question is phrased. These are rapid community tests rather than peer-reviewed studies — no peer-reviewed study of the JEV model currently exists.
One test is particularly informative because it placed both types of tasks side by side. On the task “Does this text belong to this category?”, the model passed all six predefined criteria. On the task of graded product relevance to a search query — in other words, a core e-commerce search task — it failed four of the six criteria: calibration error rose to 0.24, the model remained overconfident across all confidence bands, and it placed the overwhelming majority of pairs into a single response category.
This should be interpreted carefully. It is not a finding against what Mironet has deployed. Product categorization is precisely the first type of task — assigning a product to a category — in which the model performed well. Graded relevance is a substantially more difficult task, and according to the available measurements, the model is not yet sufficiently capable there. For an e-shop, the practical guidance is therefore: simple and clearly bounded decisions, yes; fine-grained ranking, not yet. Measurements from live production that will confirm or challenge this conclusion are only now emerging.
This is also linked to a less frequently discussed risk: a one-percent error rate in a product description is an inconvenience; the same error rate in a decision propagated across tens of thousands of products within seconds can become an operational incident.
Where can this actually help with European regulation
The EU AI Act requires human oversight for high-risk applications, and a model that returns only a number may initially appear to be a step in the wrong direction. Calibrated confidence, however, offers something that free-form text does not: a clear and auditable threshold. An organization can define that below a certain confidence level, the system must not make a decision on its own, and the case must be referred to a human. Such a rule can be written into internal policy, documented in logs, and audited. Well-measured uncertainty is more useful for oversight than articulate but non-binding explanations.
The role of the Department of Information Engineering
The Department of Information Engineering at the Faculty of Economics and Management of the Czech University of Life Sciences Prague and its AI BRIDGE initiative have long cooperated with industry through the IT People Technology Platform, which connects technology companies with academic institutions. In addition to CZU, the collaboration includes the Department of Data Analytics at the Faculty of Management of the Prague University of Economics and Business and the Department of Informatics at the Faculty of Science of Jan Evangelista Purkyně University in Ústí nad Labem. In this case, the application partner is the Mironet.cz e-shop, which states that this is the first commercial deployment of the model in Europe.
The department’s role is deliberately more modest: to observe, measure, and provide context — in other words, to verify what the company itself has little incentive to verify. How does calibration behave outside laboratory conditions, during the Christmas shopping season, or when the product catalog expands? How should confidence thresholds be set so that human oversight is more than a formality? And how well does the model perform in Czech? Measurements so far suggest that accuracy declines in non-English languages, which is a very practical concern for a Czech e-shop.
Such a deployment is valuable precisely because it is real; operational data cannot be replaced by simulation. At the same time, general conclusions should not be drawn from a single deployment — and that is precisely the work an academic institution should carry out before an initial impression becomes accepted as a shared truth.
Sources
• Chow, C. K.: On optimum recognition error and reject tradeoff. IEEE Transactions on Information Theory 16(1), 1970.
• Geifman, Y., El-Yaniv, R.: Selective Classification for Deep Neural Networks. NeurIPS 2017, arXiv:1705.08500.
• Guo, C., Pleiss, G., Sun, Y., Weinberger, K.: On Calibration of Modern Neural Networks. ICML 2017, arXiv:1706.04599.
• Brier, G. W. (1950); Gneiting, T., Raftery, A. E.: Strictly proper scoring rules, prediction, and estimation. JASA 102(477), 2007.
• Ovadia, Y. et al.: Can You Trust Your Model's Uncertainty? NeurIPS 2019, arXiv:1906.02530.
• Ouyang, L. et al.: Training language models to follow instructions with human feedback. NeurIPS 2022, arXiv:2203.02155.
• TypeSafe AI: documentation and internal benchmark tests, typesafe.ai, evals.typesafe.ai.
• Independent community tests of the JEV model conducted on 20–21 September 2026: jev-orderby-bench (pre-registered criteria; 20 Newsgroups, 360 items; Amazon ESCI, 306 human-rated query–product pairs across 30 difficult queries), Benchmark Heaven JevBench v1.3.0, PriorBench, calibration audits.
• Mironet.cz: press release on the deployment of the JEV model.
Department of Information Engineering, Faculty of Economics and Management, Czech University of Life Sciences Prague