Part III: How Things Go Wrong

"More is lost by indecision than by wrong decision." — attributed to Cicero

If Part II was about how systems find answers, Part III is about how they lose them — and, more importantly, why they lose them in the same ways.

The failure modes of complex systems are not random. They are patterned, predictable, and — here is the crucial part — identical across domains that never talk to each other. A machine learning algorithm overfits its training data in exactly the same way that a historian overfits their favorite theory to the evidence. A hospital optimizing for measurable metrics degrades in exactly the same way a Soviet factory did when its output was measured by weight. A power grid cascading into blackout follows the same dynamics as a financial system cascading into crisis.

Part III catalogs eight universal failure patterns. Overfitting (Chapter 14) is the sin of seeing patterns that aren't there. Goodhart's Law (Chapter 15) explains why every metric that becomes a target ceases to be a good metric. Legibility and control (Chapter 16) shows what happens when authorities try to make complex systems readable. Redundancy vs. efficiency (Chapter 17) is the tradeoff that kills systems when resolved too aggressively toward efficiency. Cascading failures (Chapter 18) reveals how one small break can bring down everything. Iatrogenesis (Chapter 19) — when the cure is the disease — shows up in medicine, economics, foreign policy, and software. Legibility traps (Chapter 20) deepens the analysis of what happens when you make the complex simple. And the cobra effect (Chapter 21) demonstrates that incentives backfire in identical ways across every domain humans have tried to manage.

This is not a depressing catalog of doom. It's a diagnostic toolkit. Once you can name a failure pattern, you can see it coming — in your organization, your field, your decisions, your life.

Pattern Library checkpoint: Your failure-mode analysis begins here. Pick a real problem from your own field and diagnose it using at least three patterns from Part III, importing insights from at least two outside fields.

Chapters in This Part