Process

Quality isn't reviewed at the end. It's engineered in.

Most data programmes fail on ambiguous rubrics, not lazy annotators. Our process front-loads guideline design and calibration so production volume is trustworthy from week one.

Delivery phases

Six phases from brief to steady state

01

Scoping & feasibility

We map your task type, difficulty distribution, expected volume and acceptance criteria, then model realistic throughput and the tasker profile required.

Output · Delivery plan + capacity model

02

Rubric & guideline design

Our taxonomists co-author annotation guidelines with your researchers, resolving edge cases before they become inter-annotator disagreement.

Output · Versioned guideline doc

03

Pilot & calibration

A 500–2,000 task pilot measures inter-annotator agreement (Krippendorff's alpha), throughput and rubric ambiguity hotspots. Guidelines are revised before scale.

Output · Agreement report + revisions

04

Team assembly & training

Taskers are screened, assessment-tested on your rubric, then certified. Only certified taskers receive production access on the client platform.

Output · Certified production roster

05

Production delivery

Live queues run with gold-task injection, blind overlap sampling, reviewer tiers and daily throughput dashboards visible to your team.

Output · Daily throughput + QA metrics

06

Audit & iteration

Weekly quality councils review disagreement clusters, retrain where drift appears and feed rubric updates back into training material.

Output · Weekly quality council notes

Quality control

Four independent layers of verification

Each layer catches a different failure mode: inattention, ambiguity, systematic bias, and mismatch with the client's actual intent.

Layer 1 — Gold tasks

Known-answer items seeded at 4–7% of volume, scored automatically. A tasker falling below threshold is paused and re-trained, not silently degraded.

Layer 2 — Blind overlap

10–20% of tasks are double- or triple-assigned without the taskers knowing, producing live inter-annotator agreement statistics per rubric section.

Layer 3 — Reviewer tier

Senior reviewers adjudicate disagreements, sample high-risk categories at 100%, and write the disagreement notes that drive guideline updates.

Layer 4 — Client acceptance

A final sample is packaged for your researchers with rationale attached, so acceptance decisions are made on evidence rather than aggregate scores.

Reporting

What lands in your inbox every Monday

Inter-annotator agreement

Krippendorff's α, per rubric dimension

Gold-task accuracy

Rolling 7-day per tasker and per pod

Throughput

Accepted tasks/hour vs. capacity model

Rework rate

% returned by reviewer or client

Latency

Median and p95 queue-to-delivery time

Drift

Week-over-week rubric-section score movement