A rule can be followed formally while failing to achieve its purpose. I therefore test safeguards, approval processes and acceptance criteria with this question: How could someone meet every requirement and still produce an unwanted outcome? This may result from deliberate circumvention, time pressure, unclear responsibilities or well-intended shortcuts.
For example, an order above a threshold requires additional approval. Splitting it into smaller orders keeps each below that threshold. Individual transactions follow the rule, but the intended control over the total commitment disappears. A robust safeguard must also account for related transactions.
The same applies to AI tools. Mandatory human approval offers little protection if the reviewer cannot see the sources or uncertainties. The approval step exists, but does not enable an informed decision.
I stress-test these rules against exception cases, possible workarounds and conflicting requirements. We close identified gaps and document necessary exceptions with clear accountability. The aim is a small number of effective safeguards that hold up in everyday work.
Asimov’s laws of roboticsAntipattern: The requirement is met, but its protective effect is absent. → Fix: Test workarounds and exceptions, close gaps and monitor effectiveness in use.