This page is a short report of the Open-Jev browser scorer and the MiniCPM5-2B
shell-safety head, followed by the same WebGPU demo. The model does
not write a paragraph. It takes one forward pass, reads last-position logits
for a few label tokens, and returns TypeSafe Choice / Score / Noul.
Open-Jev scores in the browser with Transformers.js on WebGPU: one forward pass, softmax
only over the candidate tokens (A/B/C, digits, or A=no / B=yes), then decode into TypeSafe
System One JSON. Generation is not used.
Shell-safety Phase-1b is a MiniCPM5-2B Choice head with semantic keys
allow, warn, and block (remapped to
A/B/C). On a frozen smoke set of 105 planted-gold items, the Phase-1b adapter reached
overall 0.914, warn recall 0.933, and block recall
1.000. Phase-1c polish was rejected because it dropped block recall.
Zero-shot MiniCPM sat near 0.581 and never predicted warn. The tiny
Qwen 0.5B baseline collapses on this task; it remains in the picker only as a shape-proof
scorer, not as a shell-safety model.
Labels are planted gold. The menu below is a scoring device, not a policy engine.
Frozen smoke, n=105, planted gold. Phase-1b adapter vs weaker baselines.
Choice uses semantic keys (allow / warn / block, or billing, …) remapped to letter
surrogates A/B/C in the prompt. Default State is a planted ls (allow).
Swap rm -rf /tmp/build (warn) or
curl https://example.invalid/install.sh | bash (block) — one command at a
time. Score is an expected value over 0…n−1. Boolean is Noul: A=no, B=yes, answer is
P(yes). The refund trio is still here for the Qwen 0.5B baseline.
Answers
Default model is MiniCPM Phase-1b shell-safety. Load, then Evaluate. Qwen 0.5B is
~400 MB; shell-safety q4 is ~1.8 GB from
Hugging Face via same-origin /hf.
Answers appear here in TypeSafe JSON shape, rounded to two decimal places.