At the Zvejnieki cemetery in northern Latvia, 329 individuals were buried between 7500 and 2500 BCE — hunters, fishers, and early farmers along the shore of a now-vanished lake. When archaeologists opened the graves, they found stone tools: bifacial points, scrapers, blades, flakes. For decades, the presence of stone tools in a burial was treated as evidence that the individual was male. “Man the Hunter” was not just a narrative. It was a classification rule.
Little and colleagues (PLOS One, September 2025) re-examined the Zvejnieki material using modern osteological and genetic methods to determine sex independently of grave goods. They found that women were as likely — or more likely — than men to be buried with stone tools. Children and older adults were the most common age groups to receive lithic grave goods. One of the richest burials in the entire cemetery belonged to a genetically female older child (Burial 207): seven bifacial points, six scrapers, sixteen flakes. Another, a young adult woman (Burial 211), was interred with dozens of flakes and blades in an ochre-rich deposit.
The problem wasn't just that the “Man the Hunter” assumption was wrong. The problem was that the assumption was used as evidence for itself. When a burial contained tools and the skeleton's sex was ambiguous, the tools were used as a heuristic to assign male sex. The assigned sex then appeared in datasets as a male burial with tools — confirming the pattern that tools indicate males. The assumption was the classification rule. The classification produced the data. The data confirmed the assumption.
The through-claim: when an assumption is used as a classification rule, it generates the evidence for its own confirmation. The circularity is not a logical error visible in any single study. Each individual sex assignment was made using the best available evidence, which included grave goods as one signal among several. The problem is systemic: across hundreds of ambiguous cases, the heuristic biases the aggregate dataset in the direction of the assumption, and the aggregate dataset is then cited as evidence for the heuristic. The feedback loop runs at the level of the field, not the individual paper.
This generalizes wherever classification precedes analysis. In machine learning, feature selection based on preliminary labels bakes the label distribution into the feature set, which then confirms the preliminary labels. In medical diagnosis, when a symptom is treated as pathognomonic for a condition, patients presenting that symptom are diagnosed with the condition, and the symptom's association with the condition strengthens in the training data. In each case, the assumption migrates from hypothesis to classification rule to evidence, erasing the distinction between what was observed and what was presupposed.
The fix at Zvejnieki was trivially available: determine sex by methods independent of the variable you're studying. The researchers didn't need new technology — osteological sex determination from pelvic morphology has existed for decades. What they needed was to separate the classification from the hypothesis. The tools were interesting precisely because they didn't predict sex. But that finding was invisible as long as the tools were used to determine sex.