Stress-testing the alerts.

A seeded synthetic population where the suspicious behaviour is put there on purpose, so every alert can be scored and every miss is visible. Turn a scenario's intensity down until a rule stops catching it, and you have measured that rule's blind spot. It is a stress test, not evidence of real-world performance, and the studio labels every rate that way.

A flat amount threshold isn't the lever: no setting of it is good, only differently bad. Moving it anyway would mean deciding what a missed case is worth, which nobody has stated, and tuning a live threshold to a population whose truth we planted. The real fix is a rule relative to each customer's own baseline, tested as the challenger in week 6.

The recall a rule reports and the recall a team delivers are different numbers once the queue is counted. The owner kept the line and named the gap as a staffing ask (about 1,390 reviews a day at bank scale, on stated assumptions) rather than quietly trading coverage for a queue that fits. Until that is met, the drowned queue is the known state, and the log says so.

Compared at the same workload and scored on a population the lines were not fitted to, splitting retail from everyone else catches 27 more planted customers than one line. Four lines add five more, but refitting on fresh populations shows the thin lines swing by tens of thousands of pounds, and the one-case correspondent line switches itself off. Two lines keep nearly all the gain with none too thin to defend. The decision is recorded but not applied: the live threshold stays £75,000 until the governed apply step in week 9, which first has to teach the rule registry per-segment lines.

The challenger catches 63 planted customers the champion misses, and its negative control (the same labels with the behaviour removed) shows no lift, so its catches come from the behaviour. It is advanced to validation, not adopted, for three reasons on the record: planted truth favours it by construction, because the planted spike is relative to each customer's own scale; it alerts on 2.7× as many distinct customers; and 49 planted customers spike before they have enough history for a baseline, so it would need a fallback to run live. The champion stays at £75,000.

A threshold rule works on the few payments above its line. A drift that barely moves the bulk of payments can move that tail a lot, so the standard distribution checks read stable while the flat lines' workload more than doubles. The decision holds £75,000, records the flat lines as unstable under drift, and makes each rule's own monthly alert volume the monitoring trigger, with the flat lines recalibrated at least annually. The drift itself is a stated scenario, not an observation.

Ranked and worked to the team's budget, the blended score finds nobody neither rule already finds, so the gain does not justify governing a scoring model. The cap is different. It flattens every score above it to one number, and something has to order the tie. While the team can work every tied case the cap is free; once the tie is larger than capacity, recall falls to the plain £75k rule's level. The decision holds the rules as they run and makes the size of the tie at any cap, against review capacity, a monitored figure. The score also ranks well out of sample but is not calibrated, so it must never be read as a probability. The live estate has too few decided cases to fit its weights; the studio previews what a weight change does through the real scorer, not whether it is right.

The first draft of this decision carried the customer-baseline rule, with a fallback, forward as the candidate. An independent reviewer found that it had been compared only with the £75k rule, and that most of its extra catches came from the simulation's start. Both findings were checked by measurement, and the decision was taken again. Scored against every rule the log already holds, and again with the first three months removed, only two rules pass both stated conditions: the £75k rule and the two-line split. The split is now the candidate for the capstone, and the customer-baseline rule is the challenger to beat. Nothing is live. The governed apply step refuses any change whose effects the decision did not predict, and the tuning paper is generated from the records rather than written by hand.

The capstone ran the whole programme on a new rule: cash deposits over £10,000 in 30 days, with planted structuring, meaning deposits kept under a reporting line in fixed amounts rather than scaled to the customer. The owner stated the terms first: at least 20% fewer alerts, and no more than 5 points of recall lost. At the team's fixed capacity, what gets caught depends on the order the queue is worked in, not on where the line sits, and every plan tested lands in the same place, so the terms cannot be met by tuning. The staffing gap found in week 4 now covers a second rule. The decision proposes the rule rather than applying it: £15,000 is the target once capacity exists, and the structuring signature only orders the queue. Like D7, it was taken again after an independent reviewer showed the first draft had asserted the queue order rather than scoring it. The first version of the synthetic scenario was also discarded before any decision, because it handed one rule a perfect score by construction.