R²: Repairability-Preserving Recursive Repair for Multimodal Agentic Systems

Abstract

Current-incident success does not show whether a persistent repair preserves the evidence and rollback paths needed for the next failure. R² audits whether the repaired system can detect, localize, repair, and verify a different unseen fault within budget. It separates pre-authorization probes from sealed faults, confines proposals to registered actions, and assigns execution to deterministic code. In an authored study of four incident families, five policies, and 50 seeds (1,000 episodes), every policy fixed the visible incident. Across 400 sealed trials per policy, future-repair success was 4.0% for performance-only and regression-gated selection, 14.8% for static-quality and council selection, and 50.3% for R². False recovery was 33.8%, 18.5%, and 0%, respectively. The experiment replays four GPT-5.6 Thinking council records as fixed inputs and makes no live model calls. It validates the construct in a controlled simulator, not production prevalence or an agent-count effect.