Climate models are often discussed as scientific tools. But increasingly, they function as infrastructure: embedded systems that societies implicitly rely on when making long-horizon decisions.

This shifts the key question from “how accurate are the models?” to:

what does it mean to maintain predictive infrastructure for a complex, partially observable Earth system?

Infrastructure, not instruments

Unlike traditional scientific models, Earth system models are:

• continuously evolving
• distributed across institutions
• dependent on shared code ecosystems (e.g. CLM, FATES, CESM)
• deeply coupled to observational pipelines

They behave less like experiments and more like infrastructure stacks.

The hidden layer: model ecosystems

The real object of interest is not a single model, but an ecosystem:

• land surface models (CLM, JULES, ORCHIDEE)
• vegetation and disturbance modules
• atmospheric and ocean components
• data assimilation systems

Each component encodes assumptions about process hierarchy.

Maintenance is scientific work

A critical but underappreciated aspect is that predictive skill depends on:

• code maintenance
• structural refactoring
• parameter governance
• reproducibility across versions

These are not “technical tasks” separate from science — they are the mechanism by which scientific hypotheses are operationalised.

A reframing

We should think of Earth system modelling as:

the maintenance of a predictive infrastructure for a non-stationary planet

rather than:

the construction of a single best model

This reframing has consequences for how we evaluate progress, allocate effort, and interpret uncertainty.