---
license: mit
language:
  - en
tags:
  - system-one
  - decision-model
  - jev
  - noul
  - calibration
  - transformer
pipeline_tag: text-classification
datasets:
  - fancyzhx/ag_news
  - google/boolq
  - stanfordnlp/sst2
---

# Build a "Jev" From Scratch — toy System One model

**⚠️ IMPORTANT — honest framing.** This is a **toy-scale reconstruction of the System One
model *interface*** that TypeSafe AI's **Jev** demonstrated (announced Sep 15, 2026). The
real Jev's internals are **proprietary and unpublished**. **This model is NOT Jev and does
not claim to be.** The architecture, heads, losses, and calibration here are our own design
that reproduces Jev's *proven interface* (state + typed questions → calibrated parallel
probabilities). See `ARTICLE.md` for the full honest story with sources.

## What this model does
Given one `state` (text) and several **typed questions**, it answers them **in parallel**
(one encoder pass over the state), no text generation:
- **noul** → probability a yes/no statement is true
- **choice** → probability distribution over options + confidence
- **score** → a value in a range

## Real results (trained on CPU, 8-core, 15 GB RAM)
| Type | Dataset | Acc | Brier | ECE |
|---|---|---|---|---|
| noul | BoolQ + SST-2 | 59.7% | 0.236 | **0.027** |
| choice | AG News | 75.9% | 0.331 | — |

Training loss 1.55 → 1.00 (3 epochs). ~3.1M params. Low accuracy is expected (toy, tiny
data slice, CPU-only); the **calibration (ECE ≈ 0.027)** is the architecturally meaningful
result. Full details: `RESULTS.md`.

## Files
- `ARTICLE.md` — the full "let's build a Jev from scratch" article (simple English)
- `RESULTS.md` — real training + eval transcript
- `README.md` — setup + usage
- `jev_toy/` — model, data, train, eval, serve source (PyTorch)
- `checkpoints/model.pt` — the trained checkpoint (cfg + state_dict + vocab)

## Usage (inference)
```python
import torch
from jev_toy.model import SystemOneConfig, SystemOneModel
from huggingface_hub import hf_hub_download
import pickle

# load checkpoint
p = hf_hub_download("azharmo/build-jev-from-scratch", "checkpoints/model.pt")
ck = torch.load(p, map_location="cpu")
cfg = SystemOneConfig(**ck["config"])
model = SystemOneModel(cfg); model.load_state_dict(ck["state_dict"]); model.eval()
```
See `jev_toy/serve.py` for a full Jev-shaped serving example.

## Reproduce
```bash
python -m jev_toy.train --epochs 3 --agnews 1500 --boolq 1500 --sst2 1500
python -m jev_toy.eval --ckpt checkpoints/model.pt
python -m jev_toy.serve --ckpt checkpoints/model.pt
```

## Sources
- Jev / TypeSafe: <https://typesafe.ai/blog/introducing-system-one-models-and-jev>
- Needle / Cactus (fully open): <https://github.com/cactus-compute/needle>

*Educational reconstruction. Not affiliated with TypeSafe AI or Cactus Compute.*