Predictive and Causal Access dissociate across levels of Biological Organization

Which level of biological organization gives the best forecast of evolutionary change? Which gives the most reliable handle for intervention? We argue these questions have different answers. Using a simulation study spanning six evolutionary regimes—random genetic drift, directional selection, Moran dynamics, frequency-dependent selection, eco-evolutionary feedback, and group-structured selection—we evaluate four biologically motivated description levels (gene composition, lineage composition, organism-level fitness, and an organism-plus-ecology summary) as predictors of four evolutionary targets, and four level-targeted interventions for their causal effects on the same targets. We identify 9 dissociation cells in which the level achieving the highest cross-validated predictive score differs from the level whose targeted perturbation produces the largest causal effect on the same outcome. The clearest mechanistic case is eco-evolutionary feedback on the resource target: the organism-plus-ecology summary achieves R² = 0.483 [0.473, 0.491] for future resource state, yet a gene-level intervention is the only perturbation that reliably alters resource state, and does so negatively (gap = −0.008 [−0.009, −0.008]) because the favored allele carries the highest per-capita consumption rate. Across all 9 dissociation cells, gene-level interventions dominate causally while higher-level summaries dominate predictively. We conclude that identifying the best-predicting level and identifying the best-intervening level are distinct empirical questions, and both are necessary for a complete account of multilevel evolutionary dynamics.