sentrygate-laya

An int8 ONNX export of Laya, the 421M System 1 decision model from Convai Innovations, packaged for the sentrygate on-device guard. sentrygate runs this model to decide, in one forward pass, whether a prompt is trying to override or bypass a system's instructions. No second LLM, no GPU.

What is in here

• laya.onnx is the int8-quantized graph, about 440 MB. It is a quarter the size of the full-precision export and scores within half a point of it on the same set.
• tokenizer.json is Laya's ModernBERT tokenizer.

Numbers

Measured on deepset/prompt-injections (662 rows), with sentrygate's four-question preset combined by MAX:

• ROC-AUC 0.869, against 0.874 for the full-precision export
• peak RAM about 1.1 GB on a CPU, loaded once per process

Use it

pip install "sentrygate[onnx]"

The model is fetched into ./models on the first scan. To point at a file you already have:

export SENTRYGATE_ONNX="/path/to/laya.onnx"

Credit and license

The base model, Laya, is by Nandha Kishor M at Convai Innovations, released under Apache-2.0. This export keeps that license. See the original model card for how Laya is trained and what it does beyond injection.