nanojev

A hypothetical reconstruction of Jev, TypeSafe's closed "System One" decision model, in the spirit of nanoGPT: the smallest working version of what black-box probing suggests. It reads a state and answers typed questions (yes/no, pick one, ordered score) with probabilities instead of generated text.

Following Archer Hume's Jev's Architecture Unmasked and Jared Palmer's kev:

• No decoding loop. A pointer head reads option probabilities out of the hidden states.
• Shared state, isolated questions. The state is encoded once; each question continues from a copy of its cache and cannot see its siblings.
• Listwise options. A <decide> token after the option list scores every option's </opt> token.
• Calibration by training. Cross-entropy on labelled outcomes plus one fitted temperature; confidence is arithmetic on the distribution.

Files

• nanojev.py: encoding, Qwen3.5-4B-Base + LoRA + pointer head, inference, answers.
• nanojev_train.py: training, evaluation, temperature fit.
• nanojev_serve.py: POST /v1/systemone server.
• data/: 6,191 train / 1,000 eval requests over 55 domains, agent-written and audited.

Run

pip install -r requirements.txt
python nanojev_train.py            # ~12 min, RTX 5090
python nanojev.py request.json
python nanojev_serve.py 8009

Results

2,344 eval questions; five domains unseen in training.

	acc	NLL	ECE
epoch 1	0.895	0.257	0.013
epoch 2	0.907	0.333	0.054
+ T=1.97	0.907	0.246	0.022

Unseen domains 0.910 / ECE 0.028; noul 0.956, choice 0.933, score 0.787. Epoch two adds accuracy but overconfidence; the temperature repairs it without changing answers. Twenty questions on one state: 160 ms forked, 1.9 s as rows. A secret planted in a sibling question is invisible (0.001), in the state found (0.999). The essay's reference-card test fails when the card is the last option.