Model Failover Is a Policy Decision, Not a Retry Strategy

Mon, 31 Aug 2026 00:00:00 +0000

A model failure does not automatically justify another model call.

The next action could be a bounded retry, a switch to an approved backup, a degraded deterministic path, or a safe stop. Choosing among them is a policy decision because every path changes the system’s risk, capability, cost, and evidence.

Treating failover as “catch exception, call model B” hides those changes behind resilience code.

Start with the failure, not the provider

Different failures need different responses.

Resilience on Andreas Nissen

Model Failover Is a Policy Decision, Not a Retry Strategy

Start with the failure, not the provider