Semantic quality is part of reliability even though it is harder to measure than HTTP status codes.
FOUNDATIONAI SRE
Treat Quality as a Reliability Signal
For AI systems, a technically successful response can still be a user-visible failure when the answer is irrelevant, unsafe or ungrounded.
A RAG assistant stays available during a bad ingestion rollout but begins retrieving outdated procedures. Availability remains 100%; task reliability does not.
Which automated evals or sampled human checks represent unacceptable quality degradation?
REMEMBERFor AI, '200 OK' can still be an outage.
No uploads · No company data · No account required