Jev: The Decision-Only AI Model That Sparked 14 Projects in 48 Hours

Background: Why a Decision-Only Model?

Diogo Almeida, who contributed to ChatGPT at OpenAI, questioned why AGI hasn't arrived despite chat models surpassing humans. He found that RLHF makes models good at pleasing humans but introduces overconfidence, mode collapse, and hallucinations — tolerable for chat but unacceptable for automation. After two years of stealth development, his company TypeSafe AI released Jev, a model that does not generate text, code, or chat. It only performs three structured decisions with calibrated probabilities: Noul (binary true/false), Choice (multi-class selection), and Score (ordinal rating). Almeida claims 20-200x speed improvement, 40-400x cost reduction, and free output tokens, calling it "the shortest path to the AI economic revolution." Within 48 hours, the community produced 14 open-source projects, including browser agents, rocket landing control, and Vercel integrating Jev routing into its eve framework.

What Jev Is: A Super Classifier for Natural Language

Jev is positioned as a "super classifier that understands natural language." Input is unstructured text or JSON; output is a typed decision with probabilities. The output schema is locked at API definition time, guaranteeing zero type errors. Almeida argues that LLM tool-call hallucinations are deal-breakers for automation; Jev's schema matching is mathematically guaranteed at 0% error. TypeSafe calls it a System One Model , referencing Kahneman's fast, intuitive System 1 versus the slow, token-by-token System 2 of LLMs. The name "Jev" honors economist William Stanley Jevons, whose paradox states that efficiency gains increase total consumption — Almeida predicts AI will follow the same path.

Three Core Operations

Noul : Judges a condition true/false with probability. Example: "Is this ticket urgent?" → {yes: 0.92, no: 0.08}.

Choice : Picks one from a set, returning full probability distribution. Example: "Which department?" → {billing: 0.71, technical: 0.18, sales: 0.11}. Supports up to 255 candidates per call.

Score : Rates on an ordered scale. Example: "Urgency 1-5" → {score: 4.2} plus per-level distribution.

Almeida summarizes Jev as a "frontier-intelligence function call": unstructured state in, typed probabilistic decision out.

Why It's Fast: Parallel Constrained Decoding

The speed comes from inference architecture, not model training. Standard LLMs generate JSON token-by-token via autoregressive decoding (150-500 sequential steps), risking invalid JSON. Jev uses a different approach:

Context and output schema are prefilled together to produce a KV-cache.

Parallel branches are created for each field, reusing that KV-cache.

Each branch attends only to its candidate tokens' logits; all others are masked.

Softmax normalizes to a probability distribution; the highest is selected.

No tokens are emitted; the model reads context once and answers all questions in parallel, turning generation into parallel classification. Harsha Gundala reproduced this with Qwen-2.5-1B on a 28-field enterprise ticket classification: autoregressive baseline 1,900 ms vs. parallel constrained decoding 270 ms (7x speedup, 100% schema compliance). Sam Snelling open-sourced System One Lite using MLX and local models. The technique is architecture-agnostic — any open LLM can adopt it by changing the inference engine. Jev's moat is TypeSafe's proprietary RLCD (Reinforcement Learning for Calibrated Decisions) training, which makes a small model more accurate and better calibrated on judgment tasks than general LLMs.

RLCD: Calibration Over Preference

RLHF optimizes for human preference; RLVR optimizes for verifiable rewards. RLCD optimizes for calibration : the model's output probabilities align with actual accuracy. High confidence means high accuracy; low confidence signals need for human intervention. Almeida notes that before generative AI, calibration was a basic ML requirement — uncalibrated models cannot be trusted in systems.

Speed, Cost, and "Free Output"

TypeSafe ran four automated workflow benchmarks, wrapping LLMs with a System One adapter for format compatibility, using GPT-6 Astra and Fable 5.1 averages as reference. Results: Jev is up to 193.6x faster than Claude Sonnet 5 and 444.6x cheaper than Claude Opus 5 . Pricing is public: $0.042 per million input tokens; output tokens are free (not a subsidy, long-term trend is further reduction). Logic: Jev generates zero tokens. End-to-end latency 70-500 ms. Almeida argues that latency has kept SaaS stuck in 2019; two orders of magnitude lower latency lets AI enter code inner loops.

Official Demos

Doom : Jev controls the character at ~10 Hz, making parallel judgments: fire? (confidence 0.82), top priority target? ("restore health" 0.57), dodge direction? ("dodge_left" 0.43). All decisions within ~100 ms, costing ~$7/hour. A hackathon is planned.

Wikiracing : From "Baseball" to "Sun", choosing among hundreds of links per step. Jev beats GPT-5.6 Terra, Claude Haiku, and Claude Sonnet (non-reasoning) in speed and step count. High-cardinality selection advantages compound. Over 255 candidates uses two-stage: independent scoring then refinement.

14 Community Projects in 48 Hours

Curated by 0xLogicrw. Categories:

Model Routing (earliest validated use case)

jev-codex-router

: Jev rates programming task difficulty per turn, routes to cheap or strong model. 237 real turn replays, ~60% cost reduction. jev-router (Pratyush Garg): Four tiers (Fast, Balanced, Strong, Long) for Claude Code and Codex; routes once per turn, reuses within tool loop. Author reports 43% token reduction. /jev-explain shows reasoning: complexity 0.82, reasoning need 0.91, tool complexity 0.64.

Vercel's eve framework integrates auto (default evaluator typesafe-ai/jev) and evaluate function for custom tools, including tool approval. Model ID on Vercel AI Gateway is typesafe-ai/jev, no waitlist.

Riley Brown built an agent with eve showing Jev's choices in real time; Agent Native published full prompt for replication.

Browser Automation

jev-ultrafast

(Browser Use team): Jev decides action and element each step; calls small model only for text input. Google Flights search completes in 7.1 s. No screenshots in main loop; Jev consumes structured DOM state directly.

Coding Toolchain

Winnow

: Context garbage collection for Claude Code. Jev judges relevance of Read/Bash/Grep outputs before passing to model. Jev Review: Pre-screen code reviews for correctness, security, reliability; forwards only high-priority issues to heavy model. Blink: Semantic navigation in codebases. At each directory level, Jev picks files most relevant to the query, focusing search resources. SemDecide: Unix CLI tool wrapping Jev for scripts and CI.

Games & Physics Simulation

1v1

: FPS bot at ~9 Hz deciding move, aim, shoot, jump.

Utkarsh Agrawal's Falcon 9 simulation in MuJoCo: 42 m booster, 425 t, 9 engines, real atmosphere, Mach-dependent drag, wind, fuel consumption. Jev controls launch to landing with no autopilot scripts, preset trajectories, or safety vetoes. Decision frequency ~1 Hz, output latency ~0.4 s; rocket becomes unstable every 0.8 s. After 12 experiments, booster landed successfully. 245 API calls, $0.04. Jev chose 5-engine ignition for landing segment.

Quantitative Trading

Prism

: Jev judges toxic flow, market pressure, mean-reversion probability in liquidity pools. Runs in shadow/advisory mode, not directly executing trades.

Agent Integration & Desktop Automation

typesafe-mcp

(Go) and jev-mcp (Node.js) connect Jev to Claude Code, Claude Desktop, Codex. Latter adds fact verification, prompt injection detection, semantic ranking. jev-desktop: Reads Accessibility Tree; Jev picks UI element to act on. neo4jev: Probabilistic pathfinding on knowledge graphs.

Controversy and Response

Niels Rogge tweeted: "12M views for a JSON classifier? We're in a bubble." Technically, Jev uses pretrained encoders for parallel classification on fixed schema — similar to BERT-era token classification. Counterargument: autoregressive LLMs perform poorly here, giving miscalibrated confidence at high cost, yet are used because they're "just good enough." Jev's value is making this fast, cheap, and reliable enough to be a dependable software function. Almeida's FAQ acknowledges Jev is not a small LLM; it uses new architecture, sampler, and training. Goal is not benchmark scores but a software-dependable interface.

A Pattern Emerges

The 14 projects share a pattern: Jev repeatedly makes rapid judgments — "whether, which, how much, what next." Generation (code, prose, complex reasoning) stays with traditional LLMs. Previously, agent decision layers forced LLMs to output JSON for parsing — using a writing tool for structured judgment. Jev separates decision from generation. Free output tokens signal TypeSafe's bet: decision call volume will far exceed generation call volume. If Almeida is right, AI automation needs not a stronger chat model but a software-dependable interface. As Jevons paradox predicts, each efficiency order-of-magnitude explodes total usage. A $0.042/M input, free output, 70 ms judgment primitive may be just the beginning. Jev is in early access with fast waitlist clearance; immediate access via Vercel AI Gateway with model ID typesafe-ai/jev.