Campaign use cases
Operating an ML-Assisted Cloaking Flow
Machine learning can combine weak traffic signals, but operators still need to own policy, exceptions, and the final decision boundary.
State the decision the model may influence
Name the flow, destinations, traffic sources, and risk band where learned scoring is allowed to affect routing. Keep hard business requirements in explicit rules so a model update cannot quietly redefine geography or a trusted testing path.
Preserve enough request context
Source labels, network traits, device details, browser signals, language, referrer, and repeat behavior give the score meaning. Missing fields should remain visible as missing rather than being treated as reassuring evidence.
Begin with observed decisions
Run the score beside the current policy before it controls a response. Review disagreements and boundary cases, then activate only a narrow slice with an owner who can restore the previous rule path.
Expect drift and confident mistakes
Traffic composition changes after a source, creative, or placement update, while legitimate privacy tools can resemble unwanted sessions. A high score is therefore a reviewable signal, not proof about a visitor's identity or intent.
Verify with labeled fixtures and samples
Replay known cases after threshold changes and inspect random live events from each major source. A healthy workflow can explain the chosen destination, the decisive evidence, and which manual override would take precedence.