This work introduces diagnostic methods for land surface model (LSM) evaluation that enable developers to identify structural shortcomings in model parameterizations by evaluating model 'signatures' (characteristic temporal and spatial patterns of behavior) in feature, cost-function, and parameter spaces. The ensemble-based methods allow researchers to draw conclusions about hypotheses and model realism that are independent of parameter choice. I compare the performance and physical realism of three versions of Noah LSM (a benchmark standard version [STD], a dynamic-vegetation enhanced version [DV], and a groundwater-enabled one [GW]) in simulating high-frequency near-surface states and land-to-atmosphere fluxes in-situ and over a catchment at high-resolution in the U.S. Southern Great Plains, a transition zone between humid and arid climates. Only at more humid sites do the more conceptually realistic, hydrologically enhanced LSMs (DV and GW) ameliorate biases in the estimation of root-zone moisture change and evaporative fraction. Although the improved simulations support the hypothesis that groundwater and vegetation processes shape fluxes in transition zones, further assessment of the timing and partitioning of the energy and water cycles indicates improvements to the movement of water within the soil column are needed. Distributed STD and GW underestimate the contribution of baseflow and simulate too-flashy streamflow. This work challenges common practices and assumptions in LSM development and offers researchers more stringent model evaluation methods. I show that, because of equifinality, ad-hoc evaluation using single parameter sets provides insufficient information for choosing among competing parameterizations, for addressing hypotheses under uncertainty, or for guiding model development. Posterior distributions of physically meaningful parameters differ between models and sites, and relationships between parameters themselves change. 'Plug and play' of modules and partial calibration likely introduce error and should be re-examined. Even though LSMs are 'physically based,' model parameters are effective and scale-, site- and model-dependent. Parameters are not functions of soil or vegetation type alone: they likely depend in part on climate and cannot be assumed to be transferable between sites with similar physical characteristics. By helping bridge the gap between the model identification and model development, this research contributes to the continued improvement of our understanding and modeling of environmental processes. / text