Skip to main content
Research

Annotation Quality Frameworks

The frameworks that make labels trustworthy — multi-layer QA, native-linguist review, the seven transcription standards and inter-annotator review.

This page documents Cognegica's annotation quality frameworks: the multi-layer QA process, native-linguist review, the seven transcription standards, and inter-annotator review. Together they turn raw labelling into data a model can trust — with disagreements adjudicated rather than averaged, and quality reported per batch against an agreed bar.

Why frameworks, not vibes

Quality is a process, not a final inspection

Quality that's checked only at the end is quality you can't defend. Our frameworks build review into every stage — calibration before work starts, two passes during, native-linguist review against documented standards, and adjudication of disagreement. The result is labels with a recorded chain of decisions, not an opaque number.

The seven transcription standards

Transcription work follows a defined set of seven standards covering verbatim conventions, speaker labelling, timestamping, handling of disfluencies and non-speech, code-switching, and formatting — so two transcribers on the same audio produce comparable output.

Native-linguist review

A native linguist reviews against the guidelines for each language, catching the dialect, register and cultural nuances that a generic checker cannot. This is the layer that makes low-resource work trustworthy.

Inter-annotator review

We measure agreement between annotators and adjudicate disagreement, reporting the agreed quality bar per batch. The target bar is set with you per project — figures here are editable placeholders, not claims.

The QA layers

Multi-layer QA framework

Each layer has an owner and a purpose. Quantitative targets are set per project and shown here as editable placeholders.

LayerOwnerPurposeTarget (placeholder)
1. Guideline calibrationLead linguistAlign annotators on the spec before work startsCalibration sign-off
2. First-pass labellingAnnotatorProduce labels against the seven standards / guidelinesPer-task SOP adherence
3. Second-pass reviewReviewerIndependent review of first-pass output<set per project>
4. Native-linguist QANative linguistDialect, register and cultural-nuance check<set per project>
5. AdjudicationLead linguistResolve inter-annotator disagreementAgreed quality bar per batch

Representative QA layers. Inter-annotator agreement targets and pass rates are scoped per project and available under NDA — no fixed accuracy figures are claimed here.

Related work

Frameworks in practice

  • Annotation service

    Transcription, diarization and annotation with these QA layers built in.

    Explore
  • Collection methodology

    How the data reaches QA in the first place.

    Read more
  • Cultural evaluation

    Evaluating human and AI outputs for cultural correctness.

    Explore

About the frameworks

Questions about annotation quality

What is multi-layer QA?

Quality is built into every stage — calibration, two-pass labelling, native-linguist review and adjudication — rather than a single final inspection.

Do you report inter-annotator agreement?

Yes. We measure agreement and adjudicate disagreement, reporting the agreed quality bar per batch. The target bar is set with you per project.

What are the seven transcription standards?

A defined set covering verbatim conventions, speaker labelling, timestamping, disfluencies and non-speech, code-switching and formatting — so output is comparable across transcribers.

Know which labels you can trust.

Set the quality bar with us and we'll run your data through multi-layer QA.