Performance of Open-Weight On-Premise Backbones

Quantify the performance of the ARCCS regulatory compliance checking system when deployed with open-weight, fully on-premise language-model backbones instead of the GPT-5-family models accessed through a commercial API.

Background

ARCCS is described as model-agnostic, but the reported experiments use only GPT-5-family models through the OpenAI API. The paper notes that the system can operate with self-hosted open-weight models for air-gapped deployments, yet no empirical evaluation of that configuration is provided. Establishing performance with such backbones would assess whether the reported compliance-checking results transfer to locally hosted models and would be relevant for deployments involving confidential regulatory or procurement documents.