ArXiv ID: 2603.06333
Summary: As recursive self-improvement moves into practice, it risks alignment drift. SAHOO introduces a framework with three safeguards: (i) the Goal Drift Index (GDI), (ii) constraint preservation checks for critical invariants, and (iii) regression-risk quantification. Evaluated across 189 tasks, it shows quality gains in code and reasoning while maintaining alignment, mapping the capability-alignment frontier in self-evolving systems.