PASTA: A Scalable Framework for Multi-Policy AI Compliance Evaluation

Researchers have introduced new frameworks for evaluating AI compliance with multiple policies, addressing growing concerns over AI governance.

A team of researchers has developed PASTA, a scalable framework for evaluating AI compliance with multiple policies, integrating four key innovations: a model-card format, policy normalization scheme, LLM-powered evaluation engine, and an interface for interpretable evaluations^[1]. PASTA can evaluate five major policies in under two minutes at approximately $3, with expert evaluation showing its judgments closely align with human experts. A user study confirmed practitioners found PASTA's outputs easy-to-understand and actionable. Meanwhile, another research team proposed GovAI-Pipe, a four-layer governance pipeline for AI in Turkey's e-Government Gateway, which serves over 68 million registered users^[2]. GovAI-Pipe maps the AI model lifecycle to governance checkpoints and is anchored to specific provisions of the EU AI Act and the GDPR data protection framework. A third study highlighted the need for comprehensive AI governance to address non-model gains, including inference gain, systems gain, and asset gain, which can drive capability progress beyond model-level governance^[3].