Model Fallback — References

Sources on retry, backoff, circuit breaking, and the error signals inference endpoints emit.

Retry and backoff

• Timeouts, retries, and backoff with jitter — Amazon Builders' Library; retry amplification, synchronised retry waves, and why jitter is not optional.
• RFC 6585 — Additional HTTP Status Codes — 429 Too Many Requests, the status that should drive a wait rather than an immediate retry.
• RFC 9110 — HTTP Semantics — Retry-After, and which methods and status codes are safe to retry at all.

Circuit breaking and bulkheads

• Release It! Design and Deploy Production-Ready Software, Michael T. Nygard — the origin of the circuit-breaker and bulkhead patterns as stated for production operations.
• resilience4j documentation — a concrete, well-specified implementation of retry, circuit breaker, bulkhead, and time limiter composed together.
• Circuit Breaker pattern, Azure Architecture Center — breaker states and the half-open probe described as a design pattern.

Provider error semantics

• OpenAI API reference — error object shapes, rate-limit headers, and finish reasons including truncation and content filtering.
• Anthropic Messages API — error types and the stop_reason values that distinguish a completed response from a truncated one.

No single source covers "HTTP 200 with an unusable body" as a resilience category; that classification remains the application's to define.