decisionbridge
Adapter that turns existing LLMs into Jev-style decision functions with option scores, optional calibration and review thresholds.
Overview
DecisionBridge brings the decision-model approach to existing LLMs without fine-tuning. You define a question and its allowed answers, get a score per option, optionally calibrate on your labelled examples, and set a threshold below which the call abstains and the request goes to human review. The output is a selected label with scores, or an abstention.
It targets issue triage, document classification and request routing. OpenAI, Anthropic, OpenRouter and local MLX backends share the same Choice interface, though score sources differ by backend (for example verbalised JSON scores from API models). Each call answers one question, and your code composes calls into a workflow. It is inspired by Jev but does not use it; it is install-from-source with no PyPI release yet.
What the decision model decides
| Type | Question | Answers |
|---|---|---|
| Choice | What is the main intent? | bug · feature · howto · other |
How it works
- Define a Choice with option descriptions
- Backend LLM scores each option
- Optional calibration on labelled examples
- Return the label, or abstain below the threshold for human review
Strengths
- Works with existing LLM providers and local MLX
- Explicit abstention and review threshold
- Optional calibration step
Caveats
- Does not use Jev; scores may be verbalised rather than true probabilities
- No PyPI release yet
- One question per call
Getting started
git clone the repo, uv pip install -e '.[api]', set OPENAI_API_KEY and OPENAI_MODEL, then use APIDecisionModel.choose().At a glance
- Type
- Project
- Status
- prototype
- Language
- Python
- Runtime
- library
- Has code
- yes
- License
- MIT
Use cases and integrations
How curated lists describe it
A Jev-inspired decision interface for existing LLMs. Explicit choices, scores, calibration, and review thresholds.
hellogumbo/awesome-jev
Frequently asked questions
Does DecisionBridge use Jev?
No, it applies the Jev-style decision pattern to existing LLMs.
Which backends are supported?
OpenAI, Anthropic, OpenRouter and local MLX.
What happens on low confidence?
The decision abstains so your app can request human review.
Related projects
Decision-forensics library that explains a Jev answer by ablating parts of the input and measuring how far the probability moves.
ProjectbetaPythonDecision sidecar and Copilot CLI pre-turn hook that mimics Jev choice, noul and score using Azure Foundry Cohere Rerank.
HookprototypePythonExact-target audit finding Jev's Choice probabilities are a step function on stated base rates while per-option Noul tracks them closely.
ProjectresearchPythonResearch replicating Jev-style typed decisions with open LLMs via logprobs; Qwen3.5-9B reached 94% with good calibration in ~1 s.
ProjectresearchPythonRunnable yardstick that tests Jev's three headline claims on a small local model: typed output, no hallucination, calibration.
ProjectresearchPythonLocal reproduction of Jev-style inference that reads option-token logits from a small LLM in one pass, with an honest failure log.
ProjectresearchPython