A booking form on a site I'd built stopped producing bookings. Not all of them. Some came through, some didn't, and nothing anywhere said anything was wrong.
The form worked. The server worked. There was no error screen, no failed request, nothing in the logs that looked like a problem. I only noticed because a real booking was made and no notification ever arrived.
What actually happened
The form had a hidden field meant to catch spam bots. A person never sees it, so a person never fills it in. If it arrives filled in, the server assumes a bot and quietly throws the submission away.
The field was named company. Chrome's autofill saw a field called company and helpfully filled it in for real people. The server did exactly what it was built to do and discarded their bookings as spam.
Every part was working as designed. The browser helped. The form submitted. The server applied its rule. The booking still vanished.
That changed the question
If I'd only tested whether someone could make a booking, the system passed. If I'd only tested whether the server could receive one, that passed too. The failure existed in neither component. It existed in the transition between them.
So instead of asking "does A work?" and "does B work?", I needed to ask: can I show that every valid booking leaving the form arrives as a booking on the server? Those are different tests. The second one forces the boundary into view, and boundaries are where assumptions collect. One side assumed a hidden field stays empty. The other assumed a filled field meant a bot. Neither assumption was written down anywhere.
The space between the boxes
Architecture diagrams draw systems as boxes connected by arrows. We spend most of our time designing the boxes. I've become more interested in the arrows.
An arrow can hide a transformation, a delay, a retry, a permission boundary, a naming mismatch, a duplicate, a missing acknowledgement, a change in ownership, or an assumption nobody realized was an assumption. Writing "form → server" makes all of that look instantaneous. It rarely is.
What I changed
The fix was small: rename the trap field to something no browser would recognize. The bigger change was that nothing gets dropped silently anymore. A submission that trips the trap now sends me a flagged email instead of disappearing, so if the trap is ever wrong again, I'll see it.
The more useful result was the change in how I test. If an important state crosses a boundary, I want evidence for the transition, not just proof that both ends are alive: what left, what arrived, whether it arrived once and intact, how long it took, and who finds out when it doesn't.
A system can be broken while every box looks green.