The Short Version

Every Formula Has the Same Error

The Rogue Wave proved that any model treating coupled components as independent misprices compound events by a factor of dⁿ. It measured the error across twenty-nine datasets in twelve domains. The error ranged from 100× to 12,000,000×.

This book asks the next question: what should the formula say instead?

Seven domains have governing equations—formulas that institutions trust, regulators enforce, and practitioners use daily. Every one contains an independence assumption. Every one is wrong by dⁿ when the components are coupled. The correction is always the same: find the independence assumption, measure d, replace p with dp.

When Diversification Works in Reverse

More suppliers. More backups. More redundancy. The standard principle: spreading exposure across more components reduces risk.

True when the components are independent. When they’re coupled, there’s a crossover point: past it, the model becomes exponentially more wrong with every component you add. The crossover is at:

At d = 2, a system with more than 7 components has a pricing error exceeding 100×. At d = 3, more than 4 components exceeds 100×.

The risk manager who adds suppliers to a coupled system is being rewarded for making the model worse. That’s not a failure of risk management. It’s a failure of the formula.

The Chain

The standard procurement formula: the probability that the supply chain survives is 1 − qⁿ, where q is each supplier’s disruption rate and n is the number of suppliers. Six suppliers at 5% disruption rate: the model says total failure is a one-in-sixty-million event.

But the suppliers share geography (a large share of leading-edge chip fabrication is in Taiwan), raw materials (roughly half of the world's semiconductor-grade neon came from two Ukrainian plants), logistics (the Strait of Malacca), and just-in-time inventory philosophy. They are independent companies. Their disruptions are not independent events.

At d = 1.8, the model is 34× too optimistic. The automobile industry lost $210 billion in 2021. Toyota, widely reported to have held larger chip stockpiles since the 2011 earthquake, cushioned the early shortage, though it too cut output that September.

The Update

The SLA availability formula: A = 1 − fⁿ, where f is a single component’s failure rate and n is the number of redundant layers. Two backups at 0.1% failure each: six nines of availability.

On July 19, 2024, CrowdStrike pushed a faulty update to its Falcon agent on Windows machines worldwide. The update operated at the kernel level. It did not matter how many backup systems you had. If they all ran CrowdStrike, they all crashed. 8.5 million devices. Airlines. Hospitals. Banks. Emergency services.

At d = 5 (shared infrastructure), two backups give four and a half nines, not six. When dp ≥ 1—total coupling—redundancy provides zero benefit. CrowdStrike was a dp ≥ 1 event: every layer shared the failure mode.

The Grid

Grid reliability standards assess generator outages independently: the probability that generators i and j are both offline is the product of their individual outage rates. The N−1 criterion tests one failure at a time.

February 15, 2021. Winter Storm Uri. The gas plants needed electricity to run compressors that pushed gas to the gas plants. A feedback loop. Cold fronts are big. When Dallas is freezing, Houston is freezing. 246 dead. $195 billion.

At d = 2.5, ten generators failing together is 9,500× more likely than the standard says. At the corrected probability, a Uri-scale loss has an expected annual cost of about $2 billion, so winterization costing $5–15 billion pays for itself within three to eight years.

The Dose

Drug safety is assessed per drug, with pairwise interaction checks. The probability of a multi-drug adverse event is modeled as: 1 − ∏(1 − r_i), where each r_i is an individual adverse-event rate.

The model does not account for drugs sharing a metabolic pathway. Five drugs that are all metabolized by the CYP3A4 enzyme compete for the same enzyme capacity. Inhibit it with one drug, and the blood levels of the other four rise. The adverse events are not independent.

At d = 2, a five-drug regimen’s adverse-event rate nearly doubles—from 9.6% to 18.5%. Opioid–benzodiazepine co-prescription: about 3.9× the overdose-death risk in a 2015 study of veterans. The correction calls for multi-drug coupling assessment from observational data, not just pairwise screening.

The Building

Structural inspection assumes elements degrade independently. The probability that m of n elements fail simultaneously: C(n,m) · p^m.

On June 24, 2021, Champlain Towers South in Surfside, Florida, collapsed in 12 seconds. 98 people died. The building had not been struck by a hurricane or an earthquake. A 2018 inspection had warned that the pool-deck waterproofing was not sloped and that water was severely damaging the slabs below. NIST’s 2026 investigation traced the collapse to the failure of two connections between garage columns and the pool deck, rooted in the original design and construction. The elements at risk shared a design, a construction crew, and a water pathway.

At d = 3, five coupled columns are 243× more likely to fail together than the independent model says. The building inspection schedule that assumes independent degradation underestimates how quickly a building transitions from “some spalling” to “structural collapse.”

The Redundancy

The flight safety case: n redundant channels, each with failure probability p. System failure requires all channels to fail: pⁿ. Triplex redundancy at p = 10⁻³ gives 10⁻⁹—the magic number in aviation safety.

The 737 MAX’s MCAS took its input from a single angle-of-attack sensor, with no cross-check against the second sensor on the airframe. When that sensor failed, MCAS commanded repeated nose-down trim that the crew could not overcome. The safety case leaned on the crew as the backup, and the crew shared the same confusing cockpit indications and the same time pressure as the automation. 346 people died on two flights, less than five months apart.

At d = 10 (shared software, shared power bus, shared environmental exposure), triplex redundancy gives 10⁻⁶, not 10⁻⁹—a thousand times worse than the safety case claims. The β-factor model already used in nuclear safety captures common-cause failures; aviation could adopt the same framework, with d ≈ β^1/n · p^−(n−1)/n.

The Ensemble

Forty CMIP6 models run the same experiment: how much warming for a doubling of CO₂? The standard analysis divides the spread by √40 to get a confidence interval. This is the standard error formula: σ/√n.

But these are not forty independent experiments. They share code lineages. They share calibration datasets. They share parameterization choices. Some are variants of the same model. They are forty guesses from perhaps seven independent perspectives.

With n_eff between 3 and 7 (illustrative correlations), the confidence interval is roughly 3× wider than the independent-model formula gives. The correction does not change the central estimate. It widens the tails of ensemble-based projections. (The IPCC’s own sensitivity range was not derived by averaging models, so this applies to ensemble means and their error bars, not to that range.)

A Dashboard, Not a Paper

The book’s third part turns the correction into practice. The key insight: d is not a constant. It changes with conditions. During calm periods, coupling is low and the independence model looks fine. During stress, coupling spikes and the model fails catastrophically.

A monitoring system computes d in a rolling window and displays four numbers: current d, lag structure, co-movement ratio, and trend. Three alert levels: standard, elevated, severe. Even a single measurement of d is better than assuming d = 1.

This is the cure: not a regulation, not a new model, but a number on a screen that updates weekly and won’t go away.

Coupling Is Rising

Supply-chain concentration: increasing. Software consolidation: increasing. Grid interconnection: increasing. Polypharmacy rates: increasing. Climate model code sharing: increasing.

If d compounds at a constant rate and n stays constant, the error grows exponentially. If both grow, the error grows faster than any fixed exponential. The correction must be recalibrated—a static d measured once is not enough.

Why Nobody Has Fixed It

Three barriers. The standard: regulators require methods from textbooks that assume independence; a corrected formula fails the audit because it doesn’t match the mandated method. The data: computing d requires joint time-series data that sits in silos (the fire actuary doesn’t talk to the flood actuary). The incentive: the firm that corrects its formula must hold more capital, more inventory, more reserves—more cost, less profit, until the disaster arrives.

But early adopters already practice the correction under other names. Toyota maintained semiconductor stockpiles. Netflix runs Chaos Monkey. The Bank of England’s 2024 system-wide exercise examined how the behavior of banks and non-bank institutions amplifies a shock. They measured d without calling it d.

The lowest-friction path is commercial: better risk pricing creates private value before it requires public mandate.

Key Numbers