Phillip Linstrum presents Project Shadow, a system for AI and civic QA that applies corrective and preventive action (CAPA) principles to address the gap between literal compliance and actual safety in language models. The author argues that current AI governance often fails because it focuses on surface-level adherence rather than functional outcomes, leading to "literal compliance" where systems obey instructions without addressing root causes.
- Project Shadow separates roles such as observation, classification, authority, execution, and review to prevent the concentration of power in a single model pass.
- The architecture treats UNKNOWN as a classification failure requiring pause or escalation, rather than allowing models to fill evidence gaps with plausible continuations.
- Actions must be named by their downstream human effect to ensure accurate harm analysis and authority checks.
- Completion of a task is distinguished from effectiveness, requiring later verification that controls reduce false positives or harm.
- The system includes a persistent Record to preserve the chain of claims, evidence, and revisions beyond the immediate news cycle.
The architecture aims to prevent AI systems from reinforcing operator ego or causing unintended harm by enforcing rigorous, long-term accountability mechanisms similar to those in regulated healthcare operations.