Testing and tuning
How do you know the filter catches what you think it catches? Testing proves it; tuning changes it — only with evidence and approval.
L0 Explain simply
An everyday analogy: a fire alarm nobody has ever tested is a decoration. The gate gets drilled the same way: in a controlled test environment, known test posters are sent through to confirm the scanners actually raise the flag; the innocent-lookalike rate is measured to see the cost of the settings; and when someone proposes loosening or tightening a rule, the drill is run first to show what the new setting would catch and what it would miss. Only then does the change reach the real gate, with approval and a record. The discipline is uncomfortable on purpose — an untested control is an assumption wearing a uniform, and the drill converts assumptions into evidence.
L1 Core concepts
Two disciplines share the workbench. Testing verifies the system does what the documents claim: list ingestion is complete and uncorrupted; every payment path and onboarding flow actually calls the filter — coverage gaps are the classic silent failure; and matching catches the name-variation classes the institution has decided it must catch, proven with constructed test cases (transliteration variants, reordered name parts, missing components) rather than assumed. Market practice expects this testing to be independent and risk-based — internal audit, a compliance testing team, or a third party validating that alert generation aligns with the stated risk appetite. Tuning is the other discipline: deliberately adjusting thresholds and rules to balance effectiveness — catching true matches — against the efficiency cost of false-positive volume.
L2 Practitioner view
Practitioner ground rules keep the workbench honest. Test entries are synthetic and clearly marked, run in test environments — real designations are not toys, and production is not a laboratory. Effectiveness and efficiency are measured separately, because improving the second at the expense of the first is exactly the failure tuning governance exists to prevent. Below-the-line sampling gives the crucial early warning: reviewing matches that scored just under the alert threshold estimates what the current setting is already missing, and what any tightening of it would have missed. And revalidation is event-driven as well as periodic: a list-schema change, a message-standard migration, or a screening-engine upgrade can silently change matching behaviour, so each triggers a fresh drill rather than a hopeful assumption that yesterday's evidence still holds.
L3 Technical details
A disciplined tuning cycle runs: hypothesis — a specific rule change with the risk rationale written first; offline replay — historical traffic re-screened under the proposed configuration, producing the concrete list of alerts that would appear and disappear; review of the would-be-missed items, because each is a potential missed target and someone senior must accept that risk explicitly; approval through change governance; deployment with verification that production received exactly what was tested; and post-implementation monitoring against the predicted volumes. Sampling methods and statistical rigour vary between institutions, and no authority mandates a particular technique — what is expected is a defensible method, evidence retained, and the ability to show that every threshold in production earned its place through a drill someone can replay.
Sources & standards1
- Market practice
Wolfsberg Group Sanctions Screening Guidance ↗ — The Wolfsberg Group · Independent risk-based testing; validating alert generation against risk appetite
Wolfsberg guidance is industry market practice, not law; institutions vary in how they apply it.
Sources for this topic2
- Market practice
Wolfsberg Group Sanctions Screening Guidance ↗ — The Wolfsberg Group · Technology, testing, and tuning considerations
Wolfsberg guidance is industry market practice, not law; institutions vary in how they apply it.
- Simplified educational illustration
Payments Signal editorial teaching models — Payments Signal
What this simplifies: The tuning cycle is presented as a single linear loop; real programmes overlap stages and differ in replay tooling and sampling statistics. No specific sampling methodology or test-case count is prescribed by any cited source — the discipline described is the market-practice expectation, not a rulebook procedure.
Used wherever diagrams, scenarios, figures, or example values are didactic constructions rather than sourced facts; every such use carries a simplifications disclosure. All people, companies, banks, and list entries in examples are fictional.
Deepest material on this page: L3 — Technical details. Where a topic stops short of implementation depth, that is a deliberate coverage decision, not an oversight — see coverage.