How does this compare to calling a chat model for classification?

It follows the model-and-session shape familiar from Foundation Models and AnyLanguageModel. A DecisionSession holds one immutable text or JSON state, questions stay independent, and choice answers decode to native enum cases.

Does the local backend generate text or just score answers?

The MLX backend runs one forward pass and reads next-token probabilities for allowed labels without generating text. Choice uses A-Z labels and scores use 0-9 labels, with explicit token-mapping checks before inference.

Can I swap the local model for hosted Jev later?

JevDecisionModel speaks the System One API or any compatible server, so the same session code can target cloud Jev or a local endpoint. Core types and the Jev backend also build on Linux.

What is the catch with trusting the returned probabilities?

Raw model probabilities are often overconfident, so apply calibration with temperature and bias, consider rotation debiasing for position bias, and read allowed-answer mass in diagnostics rather than treating it as correctness.

What do I need to build it?

Swift 6.2 or later on macOS 26 or iOS 26. The MLX backend needs Apple silicon and the Metal toolchain via xcodebuild -downloadComponent MetalToolchain; Core AI needs Xcode 27 to build and macOS 27 or iOS 27 to run. The core library and Jev backend also build on Linux.

How do I install just the parts I need?

Add the package from version 0.4.0 and enable the MLX or CoreAI trait you want. Without those traits the package has no third-party runtime dependencies and ships the core types plus the Jev backend, so a cloud-only setup stays lean.

How does asking several questions together work?

Pass an array of binary, choice, and score questions to decide; answers come back in question order with modelID, usage, and local diagnostics. The MLX backend batches prompts up to maximumBatchSize, 16 by default, for models whose layers use plain key-value caches.

CommunityJev Practices & Patterns

AnyDecisionModel

Swift package for typed yes-or-no, choice, and score decisions with MLX, Core AI, and Jev backends behind one session API.

See highlights

Open github.com

Stars: 15
Language: Swift
License: APACHE-2.0
Published by: Community
Author: mattt
Last verified: Oct 2026

mattt/AnyDecisionModel

User repository on GitHub

View on GitHub

A Swift package for typed decisions from language models (probabilities, choices, and scores), with support for local MLX models and the TypeSafe Jev API.

Forks: 2
Last pushed: Updated Oct 2026

Repo stats from the GitHub API, cached Oct 2026.

About this project

AnyDecisionModel is a Swift package for typed decisions: yes-or-no probabilities, choices among options, and scores on ordinal scales. It has three backends. MLXDecisionModel runs a small language model on Apple silicon and reads each decision from next-token probabilities in one forward pass, with no text generation. CoreAIDecisionModel does the same with a Core AI model on macOS 27 and iOS 27. JevDecisionModel calls TypeSafe's Jev through the System One API, or any server that implements the same API.

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

Tagscommunity awesomejev catalog-import

Works forJev for Classification Jev for Ranking

Used inJev for Software & DevTools

Languages: Swift

Highlights

Single DecisionSession API over text or JSON state with prewarm and batched decide calls.
MLX backend with local Hub caching, prefix-cache option, and batched prompt evaluation.
Jev backend over System One API with provider traits keeping core dependency-free.
Typed Choice, Score, and binary answers with distributions, confidence, and diagnostics.

Watch out

Choice is capped at 26 options and scores at 10 levels by label design, and MLX trait versions must stay at compatible releases. 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.

More Jev resources like this

jev-rag-benchmark

Vendor-neutral harness that tests whether Jev reranking improves a small RAG stack, using frozen hybrid BM25 plus BGE-M3 candidates and isolating retrieval, rerank, and generation stages.

Jev Practices & Patterns#community#jevbest#catalog-import

Community

erendikmenn/jev-rag-benchmarkPython

local-jev

Local offline System One server speaking the hosted /v1/systemone wire format, so existing SDK code runs unchanged against small on-machine models.