GATE.INBOX

Inbox: the gate at work

40 of 50 held-out synthetic tickets through SmolLM2 135M (onnx-community/SmolLM2-135M-ONNX), one by one. Gold labels are synthetic (data/items.json).

Reference model: optional audit, off by default.

Adapter

Model (thresholds are per model id, never shared across models)

Checking download size…

Custom Hugging Face model idExperimental

In-browser ONNX only: any causal LM with ONNX weights the transformers runtime supports, a preset, a custom id, or a dropped model folder. Architecture, size, or memory failures surface as a plain error, never a hang. Local servers and pre-scored files are the two other adapters above.

Local model folderExperimental

Drop a model folder here, or choose it

Needs config.json, tokenizer files and single-file onnx weights. Nothing uploads anywhere.

Export the gate

Take the thresholds with you

Confidence ≥ 0.5, entropy ≤ 6 bits (committed evidence). Every receipt below re-verifies on /verify.

Inbox is ready

Click “Run inbox” above to stream the 40 held-out tickets through the on-device gate.

Every counter above is computed from the streamed rows. Reproduce the numbers: pnpm capture && pnpm calibrate.

Measured accuracy Byte Verifier