JEV.sk – articleshttps://jev.sk/en/articlesNews, how-tos and research on the Jev decision model – explained, with sources.en Jev as a dispatcher for multiple agents: fewer tokens, similar accuracyhttps://jev.sk/en/jev-as-dispatcher-multi-agent-preprinthttps://jev.sk/en/jev-as-dispatcher-multi-agent-preprintSun, 18 Oct 2026 13:38:07 +0200The Jev-LiteMAS preprint let Jev route agents and screen intermediate results. Generative-model tokens fell by 12.7 to 54.6%, while accuracy stayed close. Jev as a RAG answer checker: similar accuracy, more errors let throughhttps://jev.sk/en/jev-rag-claim-verification-preprinthttps://jev.sk/en/jev-rag-claim-verification-preprintSat, 17 Oct 2026 13:40:09 +0200A preprint compared Jev 1.13 with GPT-6 Astra on RAG claim verification. Accuracy was similar and Jev far cheaper, but it passed unsupported claims more often. One extra option shifts Jev's probabilities: what a new preprint foundhttps://jev.sk/en/jev-option-list-shifts-probabilities-preprinthttps://jev.sk/en/jev-option-list-shifts-probabilities-preprintFri, 16 Oct 2026 14:08:32 +0200A preprint from 12 October 2026 reports that jev-1.13.0 probabilities depend on which other options a Choice question lists. What it means for online shops. Jev 1.14 in preview: images and the new Box question, not yet for productionhttps://jev.sk/en/jev-1-14-preview-images-and-boxhttps://jev.sk/en/jev-1-14-preview-images-and-boxThu, 15 Oct 2026 13:37:47 +0200On 14 October 2026 TypeSafe released jev-1.14.0 as a preview with Box, a question type for finding things in images. What changed, the limits, and why not production yet. Injected text can flip Jev's decisions: what a new preprint foundhttps://jev.sk/en/jev-decision-hijacking-preprinthttps://jev.sk/en/jev-decision-hijacking-preprintWed, 14 Oct 2026 13:38:14 +0200A preprint from 13 October 2026 found that injected text flipped Jev's decisions in 81.5% of cases across 228 apps. Common defences helped only a little. Jev as a citation checker for financial calculations: what a preprint showedhttps://jev.sk/en/jev-citation-check-financial-calculations-preprinthttps://jev.sk/en/jev-citation-check-financial-calculations-preprintTue, 13 Oct 2026 13:37:51 +0200A preprint from 6 October 2026 tried Jev as a check on whether AI calculations cite the right figure. It added one record over a rule and shifted with table format. Jev in AI agents: faster decisions do not automatically mean better resultshttps://jev.sk/en/jev-in-ai-agents-faster-decisions-preprinthttps://jev.sk/en/jev-in-ai-agents-faster-decisions-preprintMon, 12 Oct 2026 13:36:56 +0200A preprint from 9 October 2026 put Jev into five types of AI agents. Single decisions got cheaper, yet task success dropped in places. What it means for developers. Jev against 19 language models: a preprint shows strengths and weak spotshttps://jev.sk/en/jev-vs-19-llms-preprint-knowledge-and-mathshttps://jev.sk/en/jev-vs-19-llms-preprint-knowledge-and-mathsSun, 11 Oct 2026 13:35:04 +0200A new preprint compared Jev with 19 language models on 13 tests. It keeps pace on knowledge tasks and trails on maths. What this means for a Slovak online shop. Option labels can override definitions: what a new preprint showshttps://jev.sk/en/option-labels-override-definitions-preprinthttps://jev.sk/en/option-labels-override-definitions-preprintFri, 09 Oct 2026 08:23:54 +0200A USC preprint reports that typed decision models follow option labels more than written definitions. What this means for Choice questions in an online shop. Vercel adds a fallback model for uncertain Jev decisions: how it workshttps://jev.sk/en/vercel-ai-gateway-fallback-for-uncertain-jevhttps://jev.sk/en/vercel-ai-gateway-fallback-for-uncertain-jevWed, 07 Oct 2026 12:43:47 +0200Since 6 October 2026, Vercel AI Gateway can rerun a Jev decision with another model when confidence is low. Conditions, costs and traps explained. Jev for moderating hateful posts: what a new preprint foundhttps://jev.sk/en/jev-hate-speech-moderation-preprinthttps://jev.sk/en/jev-hate-speech-moderation-preprintMon, 05 Oct 2026 18:19:28 +0200A University of Cyprus preprint compared Jev 1.13 with GPT-6 at classifying hateful posts. Accuracy, cost, latency and what it means for online shops. Claude and Jev together: add focused decisions to your AI workflowhttps://jev.sk/en/claude-and-jev-decision-helperhttps://jev.sk/en/claude-and-jev-decision-helperMon, 28 Sep 2026 21:34:29 +0200Claude can write and plan; Jev can answer narrow decision questions. A practical Claude Code setup, safe API key handling, and limits for automatic permissions. How to try Jev without an invitation: a simple first testhttps://jev.sk/en/try-jev-without-an-invitationhttps://jev.sk/en/try-jev-without-an-invitationMon, 28 Sep 2026 21:21:41 +0200You do not need to install Jev or join a waitlist. Use OpenRouter to create a limited API key and run your first yes-or-no decision in Windows PowerShell. Jev as a low-cost first judge: what a preprint on a cascade with GPT-6 showedhttps://jev.sk/en/jev-as-first-judge-cascadehttps://jev.sk/en/jev-as-first-judge-cascadeThu, 24 Sep 2026 00:28:36 +0200A new preprint compared Jev 1.13 with 16 AI judges. It was enough for confident verdicts; uncertain ones went to a stronger model. The study’s numbers, weaknesses and limits.