Oceanographers who prioritize direct observation warn that large carbon-cycle models can hide uncertain assumptions behind elaborate computation. Modelers reply that observations are too sparse to estimate future ocean uptake without simulation. Both claims are correct as far as they go. Models are indispensable for linking processes across scales, but agreement among models may reflect shared simplifications rather than independent confirmation. Confidence should grow when models reproduce multiple observed relationships, survive tests outside the data used to tune them, and respond plausibly to process-level experiments. An ensemble is evidence only to the extent that its members earn rather than merely repeat their agreement.