The evidence, nine attempts in full

Nine attempts from the 250,000-token runs, each in full: the prompt, every command with the node it went to, what the agent wrote between commands, its last message, and every check with its output. The rig did not record what each command printed, so no page shows that. Server addresses, lease names and passwords are replaced; nothing else is cut.

They are the cases the article's calls rest on: one reported success the server does not bear out under any reading, one example of each disputed reading, the three runs of pg_resetwal, and one attempt our own rig corrupted. If you read one of them differently from us, write to eugene@docka.ai.