An AI university is really an annual audit for rule systems

如果给 AI 办一所大学

The author audited a half-year-old agent rule system called 'Old Duck Soup' and found many obsolete rules—like a bug workaround from early 2026 that should have been removed but nobody remembered why it existed. The root cause: rules are only added, never retired, and their discovery context gets lost, turning them into superstitions agents follow blindly. The fix is periodic bare runs: send an agent into the real environment without any legacy rules. If it fails, refresh the rule with fresh context; if it succeeds, delete the dead rule. This merges testing, textbook revision, and re-education into one pipeline—an annual audit line against knowledge decay.

Why it matters: A substantive first-person postmortem. The author audits their own 6-month agent rule system and surfaces 'rule degradation' — how rules lose causal context and become superstition, illustrated with a real Claude Code workaround case. Not theory, but scar tissue. Score stays a...