Start with the decision
A short decision brief makes it possible to tell whether a new signal is worth building.
Name the action
I start a signal project by asking what a person will do differently tomorrow. “Identify unusual activity” describes a detection goal. “Route the case to an analyst before the daily cutoff” describes an action. The second version gives us a deadline, a reviewer and a way to tell whether the signal arrived in time.
Write down the trade-off
Consider a team that can review twenty cases a day. A model that adds thirty plausible alerts may look better in aggregate and still make the queue worse. Measure how many useful cases appear within the twenty the team can review, and keep track of which cases the new signal pushes out.
Build the comparison
Replay the current rule and the candidate over the same historical period. Use only evidence available at the moment each decision would have been made. Record what constitutes a useful outcome, how late it can arrive, and which cases need manual labeling.
Keep the brief short
My preferred brief fits on one page. It names the action, the owner, the time limit, the baseline, and the result that would justify changing the process. If those five fields are clear, engineers and reviewers can discuss the same experiment.