Help improve the evidence behind OpenSubstance. You can submit a pull request with stronger sources, corrected data, or a clear account of where evidence is missing. Contributions can be researched by people or with an AI assistant.
Choose a small batch. Focus on 3 to 10 claims about one substance or topic. The research queue identifies claims with weak or missing sources.
Check the evidence. Find sources that support the exact claim, population, and measurement. Say whether you could read the full paper, an abstract, or only a secondary account. Missing evidence is a useful finding too.
Update the data and citations. Work on a branch in your fork or repository checkout. Update the claim’s source, confidence, and explanatory note. Add new references to the citation table and Sources page, and to the README when they are primary sources for a data category.
Record the research. Include a contribution record with the claims checked, sources consulted, and limitations. For AI-assisted work, record the actual model, tool or app used, and whether web research was available.
Validate and open a PR. Run the checks below, then submit your branch as a pull request to OpenSubstance. Summarize what changed, why, and which checks passed. If your environment cannot run the checks, say so. A human reviews the evidence before merging.
node scripts/research-queue.js
npm run validate
npm run build
With repository access, an assistant can follow the same instructions to research claims and prepare a PR. The interactive research workflow does not require separate API keys. Here is a starting prompt:
Read CLAUDE.md and docs/research-contributions/README.md. Pick a small batch of weakly sourced claims, check the available evidence, and prepare a pull request following the guide. Record sources, uncertainty, and the actual model and tools used. Run validation and the build if available, and report any checks you could not run.
AI-assisted edits still need human review. An assistant’s own rewrite does not count as an independent fact-check.
What makes a useful contribution?
Sources that directly support a claim, with relevant context and limitations.
Corrections to values or interaction explanations when the evidence contradicts them.
Honest uncertainty when a claim cannot be substantiated.
Plain-language explanations that describe the evidence without prescribing behavior.
Report an issue without a PR
Email the page or substance, the specific claim you think is wrong, and any supporting study or dataset. You do not need to edit code to flag a problem.