When the Curve Breaks

When an epidemic model predicts a smooth decline in infections, its clean curve can look as diaphanous as tracing paper, revealing every assumed contact beneath it. Yet transparency of appearance is not proof of reliability. At the Northbridge laboratory, researchers refused to vaunt their early forecast merely because it fit last winter's data. Instead, they printed each equation beside the evidence it represented and marked several cryptic parameters whose values could not be estimated confidently.

The first stress test attacked the model's most doctrinaire assumption: that people would reduce their contacts at the same rate once officials issued a warning. Mobility records showed otherwise, especially in neighborhoods where work could not move online. A penurious funding committee had allowed only a small household survey, so the team widened uncertainty ranges instead of pretending the sample was representative. Their report was intentionally concise; a prolix defense, they feared, would obscure the weakness and invite captious reviewers to quarrel over minor phrasing rather than the evidence.

Next, the team simulated a virus that remained refractory to control: masks reduced transmission less than expected, and infected people stayed contagious longer. A deft programmer then randomized school schedules, vaccination delays, and reporting gaps across thousands of trials, changing several conditions within minutes. Some combinations confounded senior epidemiologists because nearly identical starting points produced sharply different peaks. That surprise mattered. It showed that one confident line on a graph concealed many plausible futures, particularly when several uncertain assumptions interacted.

The most meretricious result appeared during a public demonstration: a bright dashboard displayed precise county forecasts, although its elegance rested on stale population estimates. When new census figures were substituted, the demonstration became a debacle; projected hospital admissions doubled, and officials withdrew the slides before the briefing ended. The lab director grew saturnine, answering questions with short murmurs and staring at the failed display. By morning, however, embarrassment had become a useful test result rather than a reason to hide the error.

At the final review, one adviser offered sententious warnings about scientists needing humility, as if a grand maxim could replace specific revisions. Another became vituperative, accusing the programmers of incompetence until the chair redirected discussion toward the code. The most serious resistance came from a hidebound agency office that still demanded a single forecast because its forms had always required one number. The team ultimately submitted ranges, failure conditions, and alternative scenarios, arguing that a model earns trust by exposing where it breaks.