Safeguards | Research for reliable AI agents

Selected work

Detecting and correcting factual errors in financial question answering.

ICML 2025 · Workshop on Assessing World Models

Interpretable signals for detecting unsupported answers.

NeurIPS 2025 · Workshop on Mechanistic Interpretability

Policy hooks, approvals, and decision records for AI agents.

Open source · Apache-2.0

We study how AI agents use evidence, follow human intent, and respond to feedback.

Google Scholar

Flow & constraint