Use this page when you have a scientific question and need to turn it into a computation whose assumptions, execution, evidence, and claim can be inspected.
The organizing chain is:
representation -> computation plan -> execution -> audit -> evidence -> claimReaders can enter through a research question, a concrete data pipeline, or an executable investigation. The workflow is not necessarily linear: an audit may send you back to revise the representation or computation plan.
Plan preprocessing, data partitions, and fixed-step training without leakage or hidden state. These pages describe a planned capability, not a current API.
Prepare quantities and atmosphere artifacts at explicit host/runtime boundaries while retaining released coordinates and provenance.
Connect the executed JAX program to derivative audits, branch behavior, and scientifically meaningful sensitivity claims.
Own random keys, runtime manifests, source cards, and evidence-to-claim boundaries explicitly.
Predict, compute, audit, and state the warranted claim with repository-owned examples and validation targets.
| Family | Status | Primary output |
|---|---|---|
| Scientific ML | Planned | Explicit preprocessing, split, batch, and audit contracts |
| Data pipelines | Current | Validated local artifacts and prepared runtime inputs |
| Differentiable research | Current guidance | Derivative meaning and independent audit plan |
| Reproducible research | Current guidance and APIs | Key lineage, manifests, evidence boundaries |
| Investigations | Current and executable | Metrics, audits, limitations, and bounded claims |
Start with Science patterns enabled by Jaxstro when the research question is primary, Query atmosphere spectra when an artifact is primary, or Executable research investigations when you want a complete executable example.