The raw-results directory contains independently readable, uncompressed tar shards of selected original files. Retained payload bytes, file paths, modes, and timestamps are unchanged from the source snapshot.

This is a public subset, not the complete private research backup. Private code, plans, credentials, and files whose contents require further review are excluded. The raw-results manifest records the retained files and their checksums; its selection report describes the scope and exclusions.

Derived observation tables

The JSONL files at the repository root are a separate, filtered data view. They are not replacements for the original result files.

corpus_attempts.jsonl: initial attempts, physical retries, effective latest retry views, and excluded canaries.
corpus_conditions.jsonl: per-lane and per-condition counts and descriptive agreement, tool-use, and request summaries.
corpus_lanes.jsonl: progress, retry, and charged-request accounting summaries.
matched_comparisons.jsonl: matched completed-case agreement summaries.
case_study_groups.jsonl: coded group summaries for two case studies.
case_study_usage.jsonl: request and token accounting summaries.
metadata.json: scope, source-data digest anchors, and row counts for these derived tables.
SHA256SUMS: checksums for the root-level public export files.

Each table has a separate observations configuration in the Dataset Viewer, not a training split. Table question, lane, and candidate identifiers are synthetic aliases; the raw artifacts retain their original result-file content.

Interpretation limits

Reference agreement is descriptive, not empirical accuracy.
Missing values remain null; missing usage is not zero.
Physical retries and effective latest-retry records are related views and must not be added as independent attempts.
Question identifiers do not establish independent experimental units.
The published subset is not a runnable machine image or a complete backup of every source artifact.

Downloads last month: 3

Total file size:

3.59 GB