The land models integrated in Earth System Models (ESMs) are critical components necessary to predict soil carbon dynamics and carbon-climate interactions under a changing climate. Yet, these models have been shown to have poor predictive power when compared with observations and ignore many processes known to the observational communities to influence above and belowground carbon dynamics. Here I will report work to tightly couple observations and perturbation experiment results with development of an ESM land model (ALM), focusing on nutrient constraints of the terrestrial C cycle. Using high-frequency flux tower observations and short-term nitrogen and phosphorus perturbation experiments, we show that conceptualizing plant and soil microbe interactions as a multi-substrate, multi-competitor kinetic network allows for accurate prediction of nutrient acquisition. Next, using multiple-year FACE and fertilization response observations at many forest sites, we show that capturing the observed responses requires representation of dynamic allocation to respond to the resulting stresses. Integrating the mechanisms implied by these observations into ALM leads to much lower observational bias and to very different predictions of long-term soil and aboveground C stocks and dynamics, and therefore C-climate feedbacks. I describe how these types of observational constraints are being integrated into the open-source International Land Model Benchmarking (ILAMB) package, and end with the argument that consolidating as many observations of all sorts for easy use by modelers is an important goal to improve C-climate feedback predictions.