Accounting for the unexpected: fault diagnosis out of the ivory tower

Published approaches to fault diagnosis almost always assume the existence of an ideal environment: a sterile room where faults always behave according to their textbook definition, test patterns always propagate errors the way they are supposed to, and tester measurements are never wrong. Such an environment is clearly a fantasy. The starting point for diagnosis is that something has gone wrong; the behavior of a real silicon defect does not necessarily follow the strictures of any idealized fault model. In order to make fault diagnosis more successful, do we need to change the way circuits are designed? I say no-design-for-test techniques meet the needs of diagnosis. Instead of design-for-debug, we need to replace our assumption of ideal behavior with realistic expectations and robust techniques