Run data (JSONL)

Download

Download runs.jsonl →

The export includes the selected public targets and allowed models; a review sample includes only its selected targets. It includes earlier observations as well as current ones, so counting passing lines does not reproduce the site’s current repo counts.

Identify the observation

run_id identifies one run; repo names its target. date is the recorded observation time. commit and, where present, tested_commit pin the source. model, platform and agent_execution describe the tested setup; test_version identifies the check. purpose distinguishes initial trials, reruns and fix verification. Where present, setup_identity retains source, OS/architecture, agent/runtime, operation and evidence versions. Missing identity stays unknown. The repo/model rollup is one selected observation; its additive setup_observations preserves history and is not a compatibility verdict.

Read the result

outcome, works and pass describe the saved run, not a promise about every workflow. checks contains checker observations. Where present, claims, example_checks and cycle_checks record what was tried and the evidence. evidence_strength describes the scope; safety records limited observations, not a safety certification.

Follow the evidence

Available transcript, report and error links point to redacted evidence. Safety summaries retain attributed findings; safety.evidence_url opens the full observations on request, and safety.evidence_file downloads their gzip JSON. Missing or null fields mean the information was not recorded or exported. Private paths and secrets are redacted. For current per-setup answers and older-checker labels, start with the scorecards; read about Coldpass for the method and limits.