The speaker introduces his personal story with a failing diabetes algorithm to frame the core problem that passing tests are not enough to keep users safe.
A detailed account of a closed-loop insulin pump's failure, from initial setup red flags and false confidence to a life-threatening event requiring manual intervention.
Danger arises not from single edge cases but from the unsafe interaction of multiple, independently safe rules, a problem that is amplified by machine learning.
The Post Office Horizon scandal illustrates how testing implementation over impact creates a false sense of security, a problem worsened by skill gaps in modern teams.
Four concrete techniques are presented: creating an interaction matrix, using property-based testing, applying chaos engineering to business logic, and testing "must never" invariants.
Explore strategies like testing ML decision boundaries, building observability into business logic, using feature flags for rule interactions, and leveraging threat models.
The conclusion emphasizes building empathy for users and asking the critical question of whether you would trust your own system with something that truly matters.
Jobs that call for the skills explored in this talk.