Open Problem Lab deals with neglected global problems. That does not make every contribution high-risk, but it does make sloppy certainty expensive.
| Level | Meaning | Required handling |
|---|---|---|
| Low | Framing, literature summaries, non-operational notes | One maintainer review |
| Medium | Data cleaning, ranking, recommendations, model outputs | Domain review and explicit limitations |
| High | Public-health actions, field deployment, resource allocation, forecasts used operationally | Domain review, red-team review, and replication |
Open a Safety Flag issue when a contribution could:
- Encourage unsafe medical, legal, or field action.
- Overstate a model or dataset.
- Hide uncertainty behind confident prose.
- Expose sensitive personal, health, or location data.
- Incentivize performative work instead of verifiable work.
- Do not merge unsupported operational advice.
- Do not convert issue-comment consensus into accepted knowledge.
- Do not treat citation count as evidence quality.
- Do not let agents create unbounded follow-up issues.
- Do not publish private or sensitive data without documented permission and minimization.
The burden sits with the submitter. If the reviewer cannot verify the method, source, or done condition, the correct status is needs-changes, needs-replication, or reject.