Directories for each evaluation's test artifacts. Each eval creates files in eval/{eval_id}/ (e.g., eval/multistep/, eval/needle/). Agents create test files here during evaluation execution. Ephemeral - cleared after evaluation completes.