What is observability, and how is it different from monitoring?

Why it matters to you

For an AI system this is harder than for ordinary software, because the interesting failures do not throw. A step that returns an empty result, a provider that answers 200 with a rate-limit body, a model that answers confidently and wrongly: none of those raise an error, so a dashboard of error counts shows nothing at all.

What works is picking one number per pipeline that only moves when the pipeline is healthy, and reading it on a schedule. On the WhoFits lead engine that number is the yield rate, the share of enriched leads that turn up a website, because a silently blocked search engine returns an empty result rather than an error.

The other half is splitting surfaces by what each can see. Mooney sends errors to five places because each one sees something the others cannot, and a decision table names which to open first per symptom. After every release a smoke test fires a deliberate crash to confirm the pipeline still works, because monitoring that is never tested dies quietly.

Where this turns up in something OCTYN runs

Each line carries the date it was recorded and where it came from.