Every moderation pipeline eventually meets a human. The quality of that human's work, and their willingness to keep doing it, depends on the interface you ask them to use. Reviewer tooling is the area where engineering teams most consistently underinvest, and the cost is paid in attrition.
Cognitive load is the metric
The wrong way to measure reviewer tooling is throughput. Cases per hour is a metric that improves when you make decisions worse and burn out faster. The right metric is sustained cases per shift, measured over weeks, against a fixed reviewer cohort.
Sustained throughput goes up when cognitive load goes down. Cognitive load goes down when the interface tells the reviewer exactly what they need to know, and nothing else, and never asks them to switch contexts mid-case.
The patterns that survive
Batch by category. A reviewer who is in the headspace for image cases performs better in image cases. Context switching between policies costs accuracy and energy.
Surface history before the case. The single highest-value piece of context is what this account or piece of content has done before. Put it above the fold; do not make the reviewer go find it.
Make appeals visible. Reviewers who never see the appeals of their decisions calibrate slowly. The feedback loop is the training signal for human reviewers as much as for models.