Constellation is not a collection of tools that pass files to each other. It is a single distributed runtime, every simulation, every observation, every closed-loop run shares the same data plane, the same provenance, and the same compute budget.
Nova-Sim, Nova-Obs, and Nova-Loop are interfaces, not silos. Whatever one module writes, the others can read, with full lineage, no export, no copy.
Content-addressed, immutable, and global. A calibrated frame produced in Nova-Obs is the exact same object a Nova-Sim experiment graph consumes, referenced by hash, never duplicated.
All work, kernels, pipelines, hypothesis loops, competes in a single placement engine that understands locality, precision, and your remaining budget.
Every artifact records what produced it, with which inputs, under which code. Reproducing a four-year-old result is a query, not an archaeology project.
A grant is a finite thing. Constellation treats compute as a metered resource attached to a PI and a funding line, so the question "what did this result cost?" always has an exact answer.
A budget is bound to a PI and a grant. Quotas in node-hours, fp64-hours, and storage are set per funding line.
Every run debits the budget in real time. Kernels report exact device occupancy; storage bills against the content-addressed footprint.
The provenance graph ties each artifact to the budget that paid for it. Cost rolls up by experiment, by paper, by grant.
Export a signed cost-and-lineage report for a reviewer or a funding body. Reproducible down to the kernel.
The platform is the floor everything stands on. The next layer up is the three modules, and how each one becomes a face of the same runtime.