Repository navigation
StrandsDecider
title: strands-decider type: tool created: 2026-10-08 last_updated: 2026-10-08 related: ["jev-ultrafast", "jev-review", "SREGym Typed-Judgment Diagnosis"] sources: ["https://github.com/strands-labs/strands-decider"] radar_quadrant: Tools radar_ring: Assess radar_position: inner
Agent workflows often call a full language model for small decisions such as which tool to run next or whether a result looks right. Each call is slow and returns text rather than a number that says how sure the model is. strands-decider is a small model built for exactly these decisions: it picks one option from a list, answers a yes/no question, or rates something on a scale, and attaches a confidence to every answer.
The model answers three kinds of question. A choice question picks one of several options. A noul question is a yes/no check scored from 0 to 1. A score question rates something against an ordered rubric. The option lists come with each request, so the same model can serve different workflows without retraining.
The project takes a pretrained 2-billion-parameter language model, removes its text-generation step, and adds a small scoring head that rates every option in one pass. It installs with pip, runs from a command line, and can serve an HTTP endpoint for other programs to call. The model is published on Hugging Face. GPU support covers CUDA and Apple's MPS, and the Apple MLX option is described as arriving with the next release.
The README reports 176 correct answers out of 231 tasks on the public JevBench set and a median response time of about 115 milliseconds on one consumer GPU. It also states that answers with confidence of 0.9 or higher are right about 95% of the time on short classification tasks. These figures come from the authors, rest on a single benchmark, and the authors themselves say differences under about 10 tasks are unresolved.
strands-decider is placed in Assess at the inner position. It addresses a real gap, a fast and calibrated decision step between plain code and a full language model, and it is the open counterpart to the hosted Jev service used by [jev-ultrafast] and [jev-review]. The repository is Apache-2.0 licensed with about 521 stars and 36 open issues when checked, but it was created on 2026-09-29, the performance numbers are self-reported, and there is no first-person use.
