Multi-layer QA: catching script-adherence and metadata errors

Quality is where most data programs fail — quietly, and only visible once a model regresses. A single review pass isn't enough. We run a multi-layer QA process that catches different classes of error at different stages.

Native linguists in the loop

Native-linguist involvement is the backbone of our QA. They catch orthographic, dialectal and pragmatic errors that automated checks and non-native reviewers miss entirely.

Script adherence and guideline compliance

Each record is checked against the project guidelines — orthography, diarization, event tags and formatting — so the delivery is consistent, not just individually plausible.

Metadata verification

Metadata errors are easy to miss and expensive later. We verify language and dialect, speaker demographics, device type, environment category and location, so the dataset can be balanced and audited.

Errors feed back into guidelines

Errors are categorised and routed back into the guidelines and SOP workflows, so recurring issues get designed out. That's the difference between QA as a gate and QA as a system.