Quick answer
A reproducible Claude Science workflow preserves more than the final figure or manuscript. Keep the question, sources, data, transformations, code, environment, model and skills, prompts, compute, outputs, failed paths, reviewer findings, and independent validation together.
Anthropic says Claude Science artifacts can include exact generating code and environment, a plain-language description, and full message history. That is strong provenance scaffolding. It does not guarantee valid data, statistics, citations, interpretation, or scientific conclusions.
A reviewable workflow
| Stage | Record | Independent check |
|---|---|---|
| Question | Scope, hypothesis, alternatives, decision rule, preregistration where appropriate | Domain reviewer |
| Literature | Query, databases, dates, inclusion and exclusion, source snapshots | Citation and evidence review |
| Data | Origin, consent, license, schema, exclusions, transformations, checksum | Data steward and quality tests |
| Analysis | Code, packages, environment, parameters, seeds, compute and logs | Code, methods and statistics review |
| Artifact | Figure or manuscript plus generating code and message history | Visual, numerical and narrative parity |
| Conclusion | Effect, uncertainty, limitations, conflicts, applicability | External expert and replication |
The Claude Science product guide explains the workbench and the access guide covers compute and data boundaries.
Literature and citation work
Record exact databases, connectors, skills, queries, filters, dates, languages, document versions, and excluded material. Check that citations exist, support the claim, match the named version, and have not been corrected or retracted.
A reviewer agent can catch some citation errors, but agents may share the same retrieval gaps. Manually inspect material sources and use a domain expert for contested or consequential conclusions.
Code, environment, and compute
Export runnable code and a locked environment. Capture operating system, hardware, container or modules, package versions, compiler, model, skill, connector, environment variables without secret values, scheduler script, job ID, and logs.
Do not rely on a long-lived conversational session as the only record. A future model or package update can change results even when the natural-language request looks identical.
Figures and manuscripts
Verify every plot against its generating table and code. Check units, axes, scales, labels, exclusions, uncertainty, color accessibility, image processing, and statistical representation. Ensure the manuscript distinguishes generated text from researcher judgment and does not turn exploratory results into confirmatory claims.
Forked sessions are useful for alternatives, but preserve the original branch and explain why one approach was selected.
Reviewer agents and external validation
Use internal reviewer agents for traceability, calculation, citation, and code-artifact consistency. Then use independent data, software, methods, reviewers, instruments, or laboratories appropriate to the claim.
For high-stakes biomedical, safety, regulatory, clinical, or engineering decisions, follow the governing protocol and qualified approval process. A workbench result is not medical advice, regulatory evidence, or experimental clearance by default.
Frequently asked questions
What does Claude Science include with a figure?
Anthropic says a generated figure includes the exact code and environment that produced it, a plain-language description, and full message history. Reviewers still need the source data, transformations, packages, parameters, and statistical checks.
Can Claude Science reviewer agents replace peer review?
No. Reviewer agents can flag citation, calculation, traceability, and figure-code issues, but they are internal workflow checks. Independent domain experts, methods review, replication, and applicable ethics or safety review remain necessary.
How should a Claude Science result be reproduced?
Freeze source and data versions, code, environment, packages, model and skill versions, prompts, parameters, seeds, compute, outputs, errors, and reviewer decisions; then rerun on an approved environment and compare against independent reference evidence.
Bottom line
Claude Science can make the path from question to artifact easier to inspect. Reproducibility still requires frozen inputs and environments, runnable code, source review, independent validation, and transparent limitations.
Official source
Source check: August 2, 2026. Verify current artifact, skill, connector, compute, data, and beta documentation.