The open agent stack for experimental science.
Start with SciPhys OS · Formats · Benchmarks · Recipes · Hosted product
SciPhys connects raw instrument files to deterministic analysis, reproducible figures, evidence-grounded writing, and expert review. XRD is the first working vertical slice. The contracts are designed to scale across spectra, images, time series, structures, and new community instrument plugins.
General scientific Skills help an agent know what workflow to attempt. SciPhys makes the workflow inspectable and measurable: original data hashes, vendor adapters, typed outputs, provenance, claim-to-evidence links, benchmark protocols, and review decisions travel together.
flowchart LR
A["Instrument files"] --> B["Adapters and scientific tools"]
B --> C["Evidence Bundle"]
C --> D["Agent Skills"]
C --> E["Figures and papers"]
C --> F["Evaluations and expert review"]
| Repository | Contract | What contributors can add |
|---|---|---|
| sciphys-os | Runtime, plugin manifest, Evidence Bundle, Skills | parsers, QC, scientific primitives, agent workflows |
| sciphys-formats | Versioned measurement and evidence schemas | mappings, conformance cases, ontology review |
| sciphys-bench | Evaluation manifests, data governance, scoring | licensed cases, expert labels, baselines, audits |
| sciphys-recipes | Executable workflow manifests | end-to-end, reproducible scientific recipes |
Open contracts keep the system portable. The hosted SciPhys product can provide private workspaces, permissions, managed compute, collaboration, organizational knowledge, certified support, and governed evaluations without making public science artifacts proprietary.
git clone https://github.com/sciphys-ai/sciphys-os.git
cd sciphys-os
corepack enable && pnpm install && pnpm build
node dist/scripts/sciphys.js xrd inspect examples/zno.xy ZnOInstall or inspect the cross-instrument Skills:
npx skills add sciphys-ai/sciphys-os --listCurrent Skills cover instrument-data analysis, reproducible scientific figures, evidence-grounded Results writing, and claim auditing. Instrument-specific science stays in deterministic plugins.
We separate code regression from scientific evaluation. Synthetic self-generated fixtures prevent regressions but do not justify real-world accuracy claims. Public and governed hidden benchmarks require lawful data provenance, immutable versions, independent expert labels, artifact/subgroup coverage, uncertainty, baselines, and reproducible result bundles.
The target metric is Verified Workflow Completion Rate: correct, reproducible, evidence-covered work with no disallowed unsupported claims and acceptance under a declared review protocol.
- Share one anonymized, licensed instrument-format fixture.
- Document a parser or analysis failure mode.
- Implement one adapter, QC check, or scientific primitive.
- Review a schema field or evidence claim in your domain.
- Curate an evaluation case with provenance and disagreement notes.
- Turn a real workflow into an executable recipe.
Start with the relevant repository's CONTRIBUTING.md. Never place private sample names, operator identities, unpublished results, institution paths, or proprietary reference data in a public issue.
- Evidence before eloquence.
- Deterministic calculation before generative interpretation.
- Interoperate with established scientific formats instead of replacing them.
- Make uncertainty, alternatives, failure, and abstention visible.
- Give data curators and expert reviewers first-class credit.
- Keep open artifacts portable outside the hosted product.
SciPhys is pre-1.0 and actively looking for instrument users, scientific software engineers, data stewards, and domain reviewers.