Golden runs

A golden run is a real payload paired with the output it should produce. Collect a handful that cover the shapes you actually see — the clean lead, the one with a missing phone number, the one where the company field is an empty string, the duplicate — and you have a regression test for the whole flow. Replay them after every change and you find out whether you broke something before the client does.

Build it as its own flow: a trigger you fire manually, a loop over the stored golden payloads, and a comparison step that checks each result against its expected output field by field. Compare the fields that matter rather than the whole blob — a timestamp or a generated id will differ on every run and would fail every comparison for no reason. Aggregate to a pass count you can read at a glance, and keep the failures visible with enough detail to see which field diverged.

Where it breaks: a golden set built from imagination. Payloads you wrote yourself all have every field populated and spelled the way you expected, so the set passes forever while production keeps breaking on the records you never thought of. Harvest them from real traffic, and every time something breaks in production, add that payload to the set so the same failure can never return unnoticed.