Disclosed upfront: I run [Tickerr dot ai], an independent external monitor for AI APIs. Today it tracks latency, TTFT, uptime, and error rates across major models. I’m trying to validate a more specific idea before building too much. Basic transport health is not the hard part. If Claude/OpenAI/Gemini gets slow, times out, or throws 5xx errors, most teams can catch that with APM, logs, Sentry, Langfuse, Helicone, Datadog, etc. The harder failure mode seems to be silent model behavior drift...