Running the new system alongside the old one on real traffic, without acting on its output or showing it to anyone. It reveals how the new thing behaves on genuine inputs, which never look like your test set, with no risk to users.
Technical view
Mirrors production requests to the candidate system and logs results for offline comparison, isolating it from user-visible effects and side effects. It surfaces distribution differences between test and live traffic, latency and cost under real load, and error classes absent from curated evaluations. Care is needed that the shadow path cannot perform actions, and that duplicated requests do not double costs or trip rate limits.
Related
Part of AI Terminology Index, section: Measuring Performance.