Data

Weekly snapshots. Every number on the dashboard is derived from one of these files. License: CC BY-SA 4.0.

Schema documentation describes each column and its units.

Coverage shows, for every week and model, whether it was due, how much was captured against what was expected, and why anything is missing. Read it before treating a week as complete.

Citing this dataset

@misc{meridian_dataset,
  title        = {{Meridian weekly snapshots}},
  author       = {Meridian},
  year         = {2026},
  howpublished = {\url{https://meridianaudit.org/data/}},
  note         = {CC BY-SA 4.0; accessed YYYY-MM-DD}
}

Snapshots

Raw response bodies

Each weekly snapshot carries a responses.jsonl.gz alongside the computed metrics when the pipeline emitted one. The gzip is the concatenation of every public (non-held-out) prompt’s Sample records for that week — full response text, token counts, stop reasons, and provider metadata. A researcher can recompute any dashboard metric from prompts.yaml + responses.jsonl.gz.

Held-out prompts are excluded by construction. Older snapshots that predate this feature omit the file; the snapshot index above flags which weeks have it.

Reproducibility

Every file listed here has a SHA-256 in its snapshot’s SHA256SUMS. A stable URL never changes: if you cite /data/2026-W16/metrics.csv today, that exact file will be here tomorrow.