When it works in the lab and you do not yet trust it in the field.
The distance between a passing test and a production-grade system is where most programs quietly die, and closing it is its own discipline. Scale engineering for the load it will actually carry. Reliability under the conditions it will actually meet. Safety architecture where an incorrect action is not recoverable, which means deciding what the system does when it is uncertain and what it must refuse to do, enforced somewhere it cannot be bypassed.
Then the unglamorous half nobody budgets for: drift detection that catches a problem while it is still small, security under real conditions, and observability that locates a failure rather than merely announcing it. What you get is a system you can deploy without holding your breath.
Find the gap
Locate the distance between where the system was built and where it has to run, the gap most systems quietly die in.
Engineer for scale
Build for the load it will actually carry, not the load it saw in the demo.
Harden for variance
Defend against the inputs, the conditions, and the variance the lab never showed.
Instrument and monitor
Add the observability that surfaces a failure before a customer does, and the drift detection that catches a problem while it is still small.
Prove under load
Test past the limit on purpose, watching the tail and the resource that saturates first.
Anonymized, all real delivered work. Described by domain, architecture, and outcome, never by client.