Each session is kept step by step: the prompt, assistant turns, tool calls with their arguments and results, errors, and reasoning where the source kept it.

Files ship as compressed JSON Lines: full trajectories, chat-format messages for fine-tuning, per-tool-call training rows and failure-then-recovery episodes.

Models are labelled by capability tier, not by name. People and repositories get stable pseudonyms, and redaction swaps names, emails and keys for placeholders.

Our release scanner reads every file of a full release, offline browser included, and must find zero high-severity issues. Lesser residual risk is published.

We remove copies of a session across sources, even when each copy was anonymised differently. Re-runs of the same task stay as separate sessions.