SiliconLabAI / OpenJev

# OpenJev
Open source System One–style decision playground: **state + typed questions → structured answers**.
Three backends:
| Mode | What it is |
|------|------------|
| **parallel** | LLM micro-scorers (one tiny `{p}` call per option, then softmax) |
| **oneshot** | Single structured JSON call to any OpenAI-compatible model |
| **decider** | Mapika/decider — real System One weights (calibration-aware RL on v10) |
## Tutorial
https://youtu.be/xtXq279B4Go
## Quick start (LLM backends)
 [code] 
Open http://localhost:3001 — mode **parallel** or **oneshot**.
## Integrate Mapika/decider (RLCD / System One)
decider is an open System One model family (Qwen3.5 fine-tunes). 
`decider-2b` **v10** includes calibration-aware RL. It speaks TypeSafe’s wire format: `POST /v1/systemone`.
### 1. Serve the model (needs a CUDA GPU, ~4 GB for 2B)
 [code] 
Smoke test:
 [code] 
### 2. Point OpenJev at it
In the UI: set **mode = decider**, **decider url = http://localhost:8000**
Or via API:
 [code] 
Or in `.env`:
 [code] 
### 3. Python (no OpenJev) — direct
 [code] 
### How OpenJev wires it
 [code] 
Same question types as TypeSafe Jev: **choice**, **score**, **noul**.
## Environment
 [code] 
## Layout
 [code]