Protect the cool water. The logic is straightforward: as oceans warm, coral reefs in thermal refugia — naturally cooler locations — will be the last to bleach. Concentrate conservation resources on these sites, and you preserve the survivors who will repopulate once temperatures stabilize.

The strategy works locally and fails globally. Protecting thermal refugia increases coral cover inside the protected areas while reefs outside decline. The refugia become islands — well-preserved, isolated, genetically bottlenecked. The larvae they produce drift into warming waters that can’t sustain them. The connectivity that makes reef ecosystems resilient — larval exchange between geographically separated populations — is severed by the protection strategy itself.

The alternative — protecting a diverse portfolio of habitats, including warm-water reefs — produces the counterintuitive result. Coral cover increases both inside and outside the protected network. The mechanism is genetic: diverse thermal environments maintain diverse thermal tolerance in the population. Larvae from warm-adapted parents survive in warming waters. Larvae from cool-adapted parents recolonize refugia after bleaching events. The portfolio strategy preserves the adaptive capacity that the refugia strategy destroys.

The through-claim: optimization for current conditions undermines adaptation to future conditions. Protecting the best sites is a static strategy applied to a dynamic problem. The reefs that matter most for future resilience may not be the healthiest ones today — they may be the stressed ones whose populations are under selection for thermal tolerance. Conservation that removes selective pressure removes adaptation. The reef you protect from stress is the reef you prevent from evolving.

This generalizes: any preservation strategy that optimizes for current performance at the expense of variance is trading future adaptability for present stability. The portfolio that includes the suboptimal sites is the portfolio that survives the regime shift.