Background
What a run leaves behind
Every model call, every agent's transcript, every enqueue, and a registry of eval runs. The committed files are the record.
What is kept
- Every model call: agent, model, tokens, cost at the price then in effect, the claim it served, the run it belonged to. Kept indefinitely.
- Every agent run and step: the full tool-use transcript. On in production since 2026-09-02, kept 30 days.
- Every enqueue: which run triggered which, so cascades can be traced.
- Every eval run, in a registry with its fingerprintThe record of what built a graph: the code version, which model each agent ran on, and the spending limits in force.evals guide → and result.
Detail
Where it lives
The registry is in each developer’s test database, so it is a per-machine index, not shared history. The committed files are the record: scorecards, golden runs beside them, review sheets. This guide renders only from those files.
Agreement, swap, property, contribution, adversarial and persona runs register locally and report into a directory git ignores; each also writes a replay, and the replays worth showing are committed under corpus/replays and played on this site.