the strings you cannot see are the ones missing

A product that ships in two languages has a defect class no reviewer catches by reading. A missing translation looks exactly like a finished screen, in the language the reviewer happens to speak. Multiply that by dozens of files and several workstreams and the only reliable answer is mechanical.

The gate walks the interface source and flags multi-word literals that sit outside the translation function and are not a declared pair in both languages. The interesting engineering is in the exemptions, because a naive version would flag identifiers, asset names and single words until everyone ignores it. The check classifies candidates by language and by shape, and it fails the build rather than reporting.

The gate is then held to the same standard as the code. It must pass on clean source, and it must fire on a deliberately planted single-language string — a check that cannot fail is not a check, and the planted case is what keeps the exemption list honest as the surface grows. A runtime language toggle against a served build stays as the visual backstop, because a mechanical gate proves strings are wrapped, not that both sides read well.

Cost: every legitimate exception has to be argued, and the exemption list becomes a reviewable artefact of its own. The gate also slows a change that is otherwise a one-line copy edit, which is the point of a gate.

What it proves: the team turns a judgement reviewers cannot reliably make into a test, and then tests the test.