Are the benchmark numbers real?

They reuse the same Jev judgments for greedy and tree policies, isolating harness use: Game24 100/100 versus 7/100 and MiniGrid 93/100 versus 70/100, with max mass error 1.33e-15. The small Opus slice is secondary, includes a 180-second transport timeout, and uses more provider tokens.

When should I not use JevTree?

It fits structured states with enumerable legal actions, executable or predictable transitions, and a verifier. For webpages or tools use a receding-horizon controller with replanning, and gate irreversible actions with a human or policy.

Which runtime matches cyclic or budgeted search?

Use enumerate_probability_tree for small complete trees, build_probability_graph when equivalent states and cycles should merge, and build_adaptive_probability_graph under a hard query budget with success, failure, and unresolved mass reported.

What does setup actually require?

Python 3.10+, pip install with the dev extra, and pytest. The heuristic provider runs a keyless smoke command; live trees need TYPESAFE_API_KEY in an ignored .env file, and the demo keeps the key server-side.

What does the runtime actually do, step by step?

It expands candidate futures, asks Jev for batched local P(action|state) distributions, composes joint path mass and merges equivalent states, aggregates success, failure and unresolved mass, then selects from the Pareto frontier of local probability, downstream quality and risk rather than taking the one-step greedy choice.

What is the benchmark evidence based on?

The primary comparison reuses the same Jev probability judgments for both policies, isolating how the harness uses one probability space: Game24 100/100 against local greedy 7/100 and MiniGrid 93/100 against 70/100, across 200 tasks with no provider errors and a maximum mass-conservation error of 1.33e-15.

What must my adapter implement?

The typed FiniteDecisionProblem contract: initial_state, common_state, decision_key, make_query, apply, is_terminal, terminal_outcome and state_payload. An example adapter ships, and the adapter stays responsible for legal actions, transitions, terminal verification and safety gates.

CommunityJev Integrations & Adapters

Jev Tree — Chuf-H

Typed Python planner runtime that turns batched Jev action distributions into executable probability trees and graphs for finite decision tasks.

See highlights

Open github.com

Stars: 41
Language: Python
License: APACHE-2.0
Published by: Community
Author: Chuf-H
Last verified: Oct 2026

Chuf-H/jev-tree

User repository on GitHub

View on GitHub

Jev-native probability tree and graph runtime for verifiable multi-step decision making.

Forks: 6
Last pushed: Updated Sep 2026

Repo stats from the GitHub API, cached Oct 2026.

About this project

Turn fast local Jev judgments into an executable multi-step probability tree. JevTree is a small, typed Python runtime for finite multi-step decision problems. Your adapter defines states, legal actions, transitions, and terminal verification. JevTree asks Jev for local action distributions, composes them into path probabilities, merges equivalent states, tracks unresolved mass, and selects actions using downstream outcome mass instead of one-step greediness.

Quoted from the project's README. Read the full README.

Tagscommunity jevbest catalog-import

Works forJev for Classification

Used inJev for Software & DevTools

Languages: Python

Highlights

Adapters define states, legal actions, transitions, and terminal verification while Jev supplies local P(action|state) distributions.
Composes joint path mass, merges equivalent states, and reports success, failure, and explicit unresolved mass.
Selects from the Pareto frontier of local probability, downstream quality, and risk instead of one-step greedy choice.
Ships an offline heuristic backend for keyless smoke runs plus a TypeSafe backend with replayable token and latency traces.

Watch out

Designed for finite deterministic tasks with enumerable actions and a terminal verifier; adaptive pruning stays explicit unresolved mass, not calibrated real-world odds. Copy reviewed against the repository README on 2026-10-04, not from running it.

Questions about this entry

Short answers drawn from the linked project or page.

Other listings of this project

Forks and copies of this project in the directory. This entry is the canonical one.

jev-tree

A TypeScript library that walks a taxonomy and calls Jev once per level, so a choice can cover thousands of leaves despite Jev's 255-option limit per question.

Jev Repos & SDKs#community#typescript#sdk

Community

reachjalil/jev-treeTypeScript

More Jev resources like this

loki

Agent runtime with a terminal UI and a messaging gateway that keeps memory, self-built skills, session search, scheduling, and subagents across Telegram, Discord, Slack, WhatsApp, Signal, and CLI.