Really enjoyed this perspective. The point that an AI request can return a perfectly valid HTTP 200 while still being completely unusable is easy to overlook in real-world systems.

I especially liked the idea of evaluating the output at the application boundary and letting semantic failures feed into the circuit breaker. That feels much closer to how AI reliability actually works than simply counting timeouts and 5xx responses.

The discussion around fallbacks, telemetry, and keeping circuit state distributed also makes this feel practical beyond just a code sample. “Reliability begins after HTTP 200” is a great way to summarize the whole idea. Great read!