The Tribe Has to Outlive the Model
Summary
The article argues that institutional memory and guardrails must live in the codebase rather than in individuals to survive frequent AI model swaps. It uses concrete examples like the Caucus Permit Gate and database triggers to show how rules are evaluated by cost and impact, and why some guardrails survive audits while others are discarded. The piece advocates codifying lessons learned into repository documentation (CLAUDE.md) and automated checks to maintain team reliability as models evolve.