CHOOSEAgents & automation
Jev: Non-Autoregressive Decision Model
Hemant@heman10x
Jev is a non-autoregressive decision model that processes application state and typed questions in a single forward pass, returning calibrated probabilities for efficient filtering, gating, and selection tasks.
TakeYou send it your application state plus typed questions (Choice, Score, Noul), and it returns calibrated probabilities in a single forward pass.
Pattern↑ Jev AI decision models
TypeSafe AI launched Jev this week, and my timeline is full of benchmarks. But most people are still testing it like a standard text model.
Jev is a non-autoregressive decision model. You send it your application state plus typed questions (Choice, Score, Noul), and it returns calibrated probabilities in a single forward pass. Input costs $0.042 per million tokens, and output tokens are free. Because all questions evaluate in parallel in 70 to 300 milliseconds, twenty questions on the same state cost roughly the same as one.
Here are 5 high-leverage use cases from the cookbooks and architecture patterns that people should try:
Instant tool-call compaction for coding agents Instead of running a slow summarization prompt when your agent hits context limits, pass the tool trace to Jev. It scores each tool call for task relevance and strips the dead output in one second. A 1M token session drops to 80K tokens without losing critical state.
RAG chunk filtering on read Standard vector search returns noisy, marginally relevant chunks. Run a fast Noul check over retrieved passages before they reach your context window: "Does this chunk contain facts required to answer the query?" Drop chunks scoring below 0.5. You get higher retrieval precision without paying LLM latency per chunk.
Hybrid browser control loops Frontier vision models are too slow for smooth browser navigation. Extract interactive DOM elements, pass them as candidates to Jev, and let it pick the click target in 70ms. Wake up a large model only when the task requires unstructured typing or visual interpretation.
Sub-50ms shell safety filters Run an evaluation check on bash commands and file edits before auto-mode executes them. If Jev returns confidence above 0.85, the command runs immediately. If confidence falls below the threshold, execution halts and asks the user for confirmation.
High-throughput log triage Point Jev at exported event logs or customer messages to evaluate multiple attributes concurrently (severity score, target service, and intent classification). You can process thousands of events per minute for pennies because output tokens carry no cost.
The practical pattern: reserve frontier language models for creative writing and open-ended planning, and route the intermediate filtering, gating, and selection steps to fast decision engines.