Help improve Bearing
You don't need to write code or use GitHub. Bearing depends on people spotting incorrect information, testing recommendations, challenging assumptions and contributing better evidence.
Choose your route
I don't want to code
- Correct a model fact. Found an outdated price, missing capability, wrong context length or retired model? Use the website form.
- Question a recommendation. Explain the job, what Bearing suggested and why it seemed wrong. Use a synthetic or redacted example, not a private prompt.
- Challenge the methodology. Think a scoring factor, weighting or evidence source is misleading? Read the challenge guide and share your reasoning.
- Improve accessibility or clarity. Report what was confusing through the website form.
A short explanation is sufficient. Links to dated provider information or reproducible examples are helpful, but you don't need to know the data schema. You can submit the website form without a GitHub account. Reports are moderated and cannot directly change model facts. Technical contributors can still use GitHub.
I want to contribute technically
- Code, tests or documentation: follow CONTRIBUTING.md.
- Model evidence: raise an evidence issue first. The checked-in registry is generated from canonical data; editing it directly is not a reliable correction route.
- Database or schema changes: use the safe database contribution guide and schema-change guide.
- Scoring methodology: see how models are rated and challenge the methodology.
You can run code checks without production keys or Neon. Migration tests use disposable PostgreSQL rather than a shared database.
The planned local contributor workbench is in PR #122, and reviewed, auditable model corrections are in PR #132. Neither is part of the default branch yet.
What makes good evidence?
Tell us which model or model variant, what is incorrect, what you expected, and how you know. Provide a source, date and any uncertainty. Where rankings are concerned, say what task and constraints you used.
Please don't submit private prompts, user records, production data, credentials or material you cannot share. A clearly described disagreement is valuable even if we do not immediately change a score.
Maintainers review contributions before modifying canonical data or production systems. Read the public data model to understand what is published and under which constraints.