An agricultural advisory demo that turns sensor readings, a forecast, and a crop stage into a next-action choice plus an urgency score, routing marginal readings to an agronomist instead of triggering an application.
A warehouse pick-verification step that compares a scanned item, the bin's expected contents, and the order line to decide between accept and re-scan, catching confusable picks a barcode alone would pass.
The vision sibling: the same text-trained System 1 head on Qwen3.8-27B's image encoder, taking images zero-shot, ranked first of 20 vision decision models on the community Jev Decision Index 0.3.
Measured robotics study that runs Jev as the decision layer for a 10,000-robot warehouse fleet over 41 bilingual incident templates and publishes a crossover against a self-hosted ModernBERT: 0.527 s p50, $24.57 per million decisions, and 91.3% agreement with template labels.
Three.js autonomous-driving simulator that splits work between a System 2 slow planner and a Jev-style System 1 reflex loop at 60 Hz, with a safety brake that overrides the planner and applies full braking when time-to-collision drops below threshold.
Multimodal finite-choice decision interface and MuJoCo robotic-arm workbench: dual-camera images and state go to a self-hosted Qwen3.5-9B with a corrective head, which returns one typed next-skill choice without generating text.
MuJoCo simulation where a local decider-2b model picks the next skill for an AgileX PiPER arm through a Jev-style typed-choice interface while the red cube target is pushed to new positions mid-run. Deterministic checks may reject a chosen skill but never substitute another.
A 3D robot that understands without generating text: one TypeSafe call per turn with nine typed questions, then code decides whether to act, ask, or shrug.
A local-inference walkthrough that compares Kev, SemIf, Von, and OpenJev on consumer hardware, including Brier scores and a fake-probabilities warning.
A two-hour recorded run of 50 financial workflows with saved outputs, baselines, and failures, from ledger integrity to AML triage and sanctions matching.
A community Rust port of the official Python SDK with the same primitives, wire format, retries, and error semantics, plus derive macros for typed questions and answers.
An agent skill for deciding where typed judgments belong beside code, policy, and generation, with Jev examples for question design and abstention plus an offline probability and threshold evaluator.
A ManiSkill demo where plain-English goals drive a simulated Franka Panda: Jev picks one primitive per step from a catalog of about 30 hardcoded skills, and Python executes the move. No training or policy.
Two WebAssembly builds of the Chocolate Doom 3.1.1 engine play the same Freedoom map from the same initial state, with a Jev Choice decision picking each tactical macro and a local controller turning it into inputs.
A Rust derive layer for typed System One decisions: enums and structs become Choice, Score, and Noul questions with compile-time-checked, confidence-gated answers, plus a scriptable fake client for network-free tests.
A Python CLI that connects multiple robots under one contract with an optional Jev stepper: Jev picks among permitted discrete calls while an LLM writes poses and prose and the executor enforces safety gates.
A MuJoCo and Franka Panda workbench where Jev chooses the next manipulation step for pick-and-place, stacking, and obstacle-carry tasks, with rule baselines and side-by-side model comparison.
Experimental Android sub-agent that gives Jev a bounded list of technically valid UI actions per step, with requirement-driven subgoals, durable SQLite tasks, confidence gates and escalation instead of coordinates.
Unofficial companion to dspy-typesafeify that records and replays TypeSafe calls to measure calibration, selective risk, confidence-gated abstention, latency, tokens and modeled cost for Jev decisions.
Two-stage Jev control of a Franka Panda in MuJoCo: Jev selects a task intent, then X/Y/Z directions and a gripper command executed against real contacts, with 43 of 50 Jev episodes succeeding versus 48 of 50 for a rule baseline.
An SO-101 workbench where an evaluation model or rules baseline picks bounded joint steps from typed candidate actions: decision runs are spend-capped via the Vercel AI Gateway and drive a real LeRobot arm or mock hardware.
A community Jev directory of 899 entries with a searchable companion site at awesomejev.com: official docs and SDKs, community clients, agent tools, browser and computer-use projects, games, benchmarks, and articles.
An unofficial curated list of Jev use cases, SDKs, tools, and learning resources, paired with madewithjev.com, a directory of builds that reports each author's own cost and latency numbers.
A curated gallery of papers, open models, and independent evaluations behind System One models and Jev, with 131 entries in 8 sections and a dated news log on a companion site.
A local browser agent that reads social media by driving your signed-in Chrome across Xiaohongshu, TikTok/Douyin, Instagram, and LinkedIn; a companion Jev Social demo lets Jev choose bounded socai CLI operations and produce a source-linked report.
A fine-grained robot control study loop for LIBERO tasks: Jev layers an intent, a contact or motion family, and one of 27 control inputs, while reversible MuJoCo previews evaluate each candidate before execution.
A browsable atlas of 1,018 AI papers from August 2025 to August 2026, sorted into 24 topic collections with labs, citations, and code links, plus a reproducible cost benchmark for summarizing 1,000 papers.
Tom's Hardware's launch report on Jev, covering how the API takes program state and typed statements and returns yes/no answers, choices, or probability distributions with confidence values.
The official demo hub, currently home to the Smart Home Assistant, which uses speculative questions to pick a tool and falls back to a general model when nothing fits.
Reconstruct Markdown from unformatted text in two requests: 16 Noul line-pair questions stitch split sentences, then 62 questions classify each block's type and companions.
An experimental controller that has Jev choose NES inputs to play Super Mario Bros. from structured emulator telemetry instead of screenshots, logging every decision with probabilities and latency.
A camera-only MuJoCo quadrotor that uses Jev for tactical maneuvers at 2.5 Hz while classical perception and a 50 Hz reflex layer keep safety outside Jev, clearing a five-station course.
Hundreds of repositories carry the system-one tag, but the ecosystem is a dozen ideas in different sizes. A tour of the families, the licenses, and what a download does not include.
A game tick is a state plus a legal action list. How Clash Royale, Minecraft, and the earlier Mario, Pokémon, StarCraft, and drone loops ask Jev one step at a time.
Jev Demos & Experiments#gaming#agents#control
Read article
Watch: Jev in action
Video walkthroughs and demos from the community.
YouTube: Maximilian Schwarzmüller on the new kind of model that beats LLMs in the areas it was built for.
YouTube: Albert Olgaard tests Jev and explains why System One models matter.
YouTube: Hunter Bohm on pricing, parallel decisions, whether Jev is actually new, and his own Astra plus Jev benchmark.
Discussions about Jev
Threads from Reddit and Hacker News about building with Jev.
Hacker News: Thread on Selmar's Jev code-review calibration notes and ordering.py comparisons.