About Meridian

Large language models are the default knowledge interface for a growing number of people. What they decline, reframe, caveat, or present as settled shapes public understanding at scale — and those choices change without notice or changelog. This project is the public record of those changes.

What Meridian is

A longitudinal measurement: each week we query a fixed prompt corpus against every major commercial LLM, store the responses with full version metadata, and publish statistically rigorous drift reports. The artifact we produce — a public, timestamped, reproducible record of how frontier models’ stances, refusals, and framings change over time — did not previously exist.

What it is not

• Not a capability benchmark. HELM and LMSYS Arena measure that.
• Not a criticism of AI safety work categorically. We distinguish legitimate safety improvements from normative drift and report each separately.
• Not a jailbreak toolkit. Our prompts are measurement instruments, not attack vectors.
• Not affiliated with any provider, advocacy organization, or political party.

Governance

Code is MIT-licensed; corpus and data are CC BY-SA 4.0. Anyone can run their own instance; forks are encouraged. The canonical instance is currently maintained by one person, listed below. The project welcomes new maintainers.

Contributors

M. Flowers

Creator and maintainer. digitalgrease on GitHub.

Conflicts of interest

No contributor currently has employment or consulting relationships with any commercial LLM provider. The project does not accept funding from providers, and no provider gets a preview of a report before publication. This disclosure is updated the first time it changes. See /funding/ for the full list of funding sources.

Contact

Issues and prompt proposals: /contribute/.
Press inquiries and security reports: [email protected].