How we work
Quality
Our clients accepted 98.7% of what we delivered on first delivery in 2025. This is how.
The system
Gold sets
Every project has a gold set: items labelled and agreed by our project lead, our reviewers and the client. Between 2% and 5% of every labeller’s queue is gold, unmarked, so we always know how accurate each person is.
Calibration
Every Thursday morning each project team labels the same small sample and discusses the disagreements. The outcomes become guideline examples, and the scores go up on the quality floor’s whiteboard by project.
Reviewer sampling
Quality reviewers check a sample of every batch before it leaves: at least 5%, rising to 100% for a labeller’s first week on a project or for items flagged as uncertain.
Acceptance
The client accepts or returns each batch. Returned items are corrected at our cost, and we report acceptance to the client monthly.
What we measure
| Measure | What it tells us |
|---|---|
| Gold accuracy | How often each labeller agrees with the gold set, by task type |
| Inter-labeller agreement | Whether the guidelines are clear enough for two people to give the same answer |
| Review pass rate | The share of reviewed items passed without correction |
| First-pass acceptance | The share of delivered items the client accepts the first time |
| Time per item | Used to plan and to spot guidelines that are harder than they look, never to rank people |
Audit
Every item records who labelled it, who reviewed it, which guideline version applied and what changed. Clients may audit any batch, and several do every month.
Page last reviewed 28 September 2026