Traditional monitoring rests on one quiet assumption that nobody ever writes down: the same input gives you the same output. Something breaks, you replay the request, you watch it break again, you fix it. Now send the same request to a model twice.
Source: [Dev.to](https://dev.to/swapnoneel123/ai-observability-explained-what-it-is-and-how-it-works-487)