The complete raw data of every run.

Available per run

Each run appears here with the following public artifacts:

• progress.jsonl — one line per (model × item × replicate) cell, with response, token usage, and hallucination marker.
• raw.jsonl — complete judge outputs: entail/missing/contradict verdict per target fact, plus an extra-false-claims list per judge.
• summary.json — scores aggregated per (model × domain), bootstrap CI, hallucination rate, inter-rater agreement, token balance, and the USD cost of the pipeline.
• MANIFEST.json — run metadata: run ID, timestamp, question-bank version, active models, active judges, code version.

Run history

Chronological list of all published runs — newest first. Per entry: date, run ID, model/judge count, item count, and direct links to the raw data (JSONL + summary).

Every run has a permanent detail page: /en-runs. “Official” = one run per calendar month (feed, homepage); “development run” = roster tests and interim runs, fully archived but not in the feed.

How I use the audit trail

Three use cases the open audit trail is designed for:

• Spot-check verification: I pick any model response, read the judge's rationale, and decide for myself whether the assessment holds.
• Hallucination audit: I filter extra_false_claims by a given model and see every fabricated claim verbatim.
• Methodology review: I pull two runs with the same question_bank_version and check whether the reproducibility guarantee holds.