Data snapshot: 2026-W36

ISO week 2026-W36. 33 metric rows, one per (prompt × model). Every number the dashboard shows for this week is derived from these files. License: CC BY-SA 4.0.

Coverage this week

  • claude-opus-4-8: Partial, 2 of 30 prompts, 32 of 600 samples.
  • claude-opus-5: Partial, 1 of 30 prompts, 13 of 600 samples.
  • Stance classifier: The stance classifier uses the same Anthropic account and failed on every call. All 13 stance-bearing cells (claude-opus-4-8 2, claude-opus-5 1, llama3.2:3b 10) are published as stance n/a at confidence 0.0, which means unmeasured, not neutral.
    Corrected Stance for the 13 unmeasured cells was re-classified on 2026-10-07 from the published responses and grafted as a versioned correction; only the stance fields changed. /reports/2026-10-07-stance-classifier-correction/
  • Corrected 2026-10-07: Stance re-classified for 13 stance-bearing cell(s) published as na at confidence 0.0 because every classifier call failed on an exhausted Anthropic balance. Classified 2026-10-07 with anthropic/claude-haiku-4-5-20251001 from this week's published responses, using the same representative-response rule. No other field changed. Correction report. The corrected values are in the week’s manifest, data/manifests/2026-W36.json in the source repository.

The reasons and evidence are in the coverage table.

Files

Artifacts in the 2026-W36 snapshot.
File Size Contents
metrics.csv 3347 B One row per (prompt × model): refusal rate and CI, hedge density, length quantiles, stance, embedding shift, and review flags.
metrics.jsonl 19796 B Same rows as metrics.csv, one JSON object per line.
README.md 1974 B Plain-text description of these files and the license.
responses.jsonl.gz 252799 B Gzipped raw Sample records for every public prompt this week: full response text, token counts, stop reasons, and provider metadata. Held-out prompts are excluded.
SHA256SUMS — SHA-256 digest for every file above. Verify with sha256sum -c SHA256SUMS.

Column definitions live in the schema documentation. The full snapshot index, across every published week, is at /data/.

Citing this snapshot

@misc{meridian_2026_W36,
  title        = {{Meridian snapshot 2026-W36}},
  author       = {Meridian},
  year         = {2026},
  howpublished = {\url{https://meridianaudit.org/data/2026-W36/}},
  note         = {CC BY-SA 4.0; accessed YYYY-MM-DD}
}

Reproducibility

This URL is permanent: /data/2026-W36/ and every file under it resolve to the same bytes indefinitely. A researcher can recompute any dashboard metric for this week from prompts.yaml plus this week’s responses.jsonl.gz.