Reliable AI Systems Research

PolicyStrata: Responsibility-Scoped Testing for Cross-Layer Policy Drift in LLM Data Agents

PolicyStrata tests whether policy obligations survive the transitions from model-visible capabilities to validation, SQL, database controls, and result release.

Evidence: Detected: 1,720/1,720 non-clean casesDetected: 1,720/1,720 non-clean cases | Clean controls: 80 with 0 false positives