Scientific representations sit between mathematical methods and research workflows. They specify what an array means before a numerical method acts on it and before a workflow interprets the result. A representation fixes conventions such as units, coordinates, sampling semantics, static metadata, and the boundary of a scientific claim.
Choosing an array dtype and shape is not the same as choosing a scientific representation. Two arrays can have identical values and shapes while representing different physical dimensions, coordinate frames, spectral densities, or parameter constraints. Jaxstro makes those distinctions explicit where the runtime supports them.
Choose physical constants, named unit systems, exact dimensional quantities, and explicit equivalencies without hiding conversions inside numerical kernels.
Represent vectors, rigid transforms, sky frames, parallaxes, and proper motions with their orientation, angular, unit, and singular-domain conventions visible.
Keep spectral coordinate, sampling, density semantic, source provenance, local artifact identity, and evidence-gated interpolation policy attached to the data.
Separate structured physical state from unconstrained optimization coordinates, and record what an executed method consumed without confusing runtime provenance with scientific validation.
Planned Jaxstro capability. Distinguish probability, uncertain state, noise, numerical error, and model discrepancy before comparing linearized, sigma-point, and ensemble propagation. No runtime module or schedule is promised.
Deferred abstraction. Separate field values, domains, coordinates, topology, operators, units, and boundaries. A common runtime awaits evidence from two real consumers; these guides promise no module or schedule.
| Material | Status | Runtime boundary |
|---|---|---|
| Units, coordinates, spectra, atmospheres, parameters, and state | Current | Implemented Jaxstro owners with explicit contracts |
| Uncertainty propagation | Planned | No jaxstro.uncertainty runtime module |
| Fields and discretized domains | Deferred | Common runtime awaits evidence from two consumers |
How to read this section¶
Each implemented capability page begins with a seven-row representation contract. Read that table before the examples: it names the mathematical object, convention, runtime owner, shape and unit policy, JAX transform boundary, evidence, and downstream interpretation boundary. Planned and deferred guides instead begin with an exact status and keep their derivation, failure modes, evidence gate, and claim boundary visible.
Status labels separate implemented Jaxstro capabilities from planned propagation work and deferred field abstractions. Numerical algorithms live under Numerical methods, research workflows remain at their current published locations until their dedicated migration, and validation anchors live in Validation.