Wrinkle structures — millimeter-scale ridges and pits preserved in ancient rock — are textbook indicators of shallow water. The logic is clean: microbial mats form these patterns, microbial mats need sunlight, sunlight doesn't reach past ~180 meters. Find wrinkle structures, conclude shallow water. The proxy has been reliable for decades.
Martindale et al. found wrinkle structures in 180-million-year-old rock from Morocco's Dadès Valley — in sediments deposited well below the photic zone. The structures shouldn't exist there. They violate the diagnostic.
The resolution: chemosynthetic bacteria. Turbidite flows — underwater debris avalanches — carried organic material into the deep while depleting oxygen, creating conditions where chemical-energy-dependent microbes could colonize the seafloor between disturbance events. The bacteria built mats. The mats made wrinkle structures. No sunlight involved.
The proxy was encoding the wrong variable. It wasn't measuring depth. It was measuring microbial mat presence — which correlates with shallow water so reliably that the correlation was mistaken for causation. When an alternative energy source broke the correlation, the proxy kept working (it still detected mats) while its interpretation failed (it still concluded shallow).
This is a general failure mode of proxy measurements. A proxy encodes a chain: observable → intermediate process → target variable. The intermediate process (mat formation) is the actual signal. The target variable (depth) is inferred by assuming the intermediate process has only one cause. When a second cause appears — chemosynthesis instead of photosynthesis — the proxy reports accurately about its actual signal while lying about its assumed target.
The dangerous part isn't the error. It's that the proxy had been right often enough to become definitional. Wrinkle structures didn't indicate shallow water; they meant shallow water. The distinction between measurement and meaning collapsed, and no one noticed until the proxy encountered a case it had never been wrong about before.
Every field has proxies that have graduated from indicator to definition. The question is always the same: what variable does this actually encode, and what variable do we think it encodes? The gap between those two answers is invisible — until something fills it.