The corpus directory contains the transcripts collected by the experiment runners. Browse the files on GitHub: /corpus/, or explore the curated viewer below. The viewer loads entries straight from /corpus/v1.4/corpus-v1_4.json, so rebuilding the corpus automatically updates this page.

Corpus viewer

Use the filters to jump between tag types, correctness, severity, and edge cases. Selecting an item reveals the user/assistant snippet, annotations, and links back to the source file.

Corpus viewer

Protocol transcripts

Explore experiment transcripts, tags, and rulings.

Versionv1.4

Tags

Correctness

Severity

Mode

Search

No entries match the current filters.

Adjust experiment, tag, or search filters to see results.

Layout

• corpus/v1.4/schema.json — JSON Schema describing the shape of each saved session (IDs, meta fields, turns, and parsed footers).
• corpus/v1.4/index.jsonl — newline-delimited index of sessions with model, provider, condition, challenge type, and timestamps.
• corpus/v1.4/sessions/*.json — per-session transcripts generated by experiments (e.g., exp1-protret-0001.json). Each file records user and assistant turns, raw headers, parsed tags, and footers.

The files follow the same schema validated by node scripts/test-transcripts.mjs. Session IDs align with the id fields in each experiment’s configs.jsonl so you can trace runs back to their source configuration.