Your Model Is Only as Good
as Your Annotators.
One person with 50 accounts is poisoning your training data right now. Sybil attacks on labeling pipelines are invisible — until your model ships broken.
Sybil attacks on labeling pipelines
Your inter-annotator agreement score can look healthy while one person drives 40% of labels.
How it works
One DID per device. Multiple devices by one person form a detectable cluster.
Your labeling platform embeds the APTOGON widget before granting access to annotation tasks.
A unique gesture confirms a live human. Device signals prevent automation scripts from passing.
Each physical device gets one DID. One person with 10 phones = 10 flagged accounts, not 10 annotators.
Your backend can see when multiple DIDs share the same trust graph — exposing sybil clusters before they contaminate labels.
Why alternatives fall short
Nothing else ties an annotator to a physical device.
What this means for your model
Ready to trust your training data?
Start free with 1,000 verifications/month. Works with any labeling platform via API.