Recursive Self-Improvement Stability under Endogenous Yardstick Drift

Abstract

This preprint develops an interface theory for recursive self-improvement under endogenous yardstick drift, where a system changes its own evaluator, benchmark, memory, and verification process. It formalizes replayable conditions for distinguishing claimed improvement from stable improvement under delayed audit, evaluator drift, verification backlog, and governance safety constraints.

Keywords

• recursive self-improvement
• endogenous yardstick drift
• evaluator drift
• self-modifying systems
• replayable interfaces
• stability
• delayed audit
• delayed challenge
• shadow certification
• stable gain
• admissibility
• AI safety
• AI governance
• governance safety
• error debt
• contradiction preservation
• semantic retention
• semantic volume
• proof-carrying
• verification backlog
• no-meta
• benchmark decay
• autonomous agents
• AI
• AGI

Identifiers and source records

• DOI
• Works entry
• ORCID