Troubleshooting

schedule.py check fails

Check user_config.yaml host/user/key/workspace, node venv, model cache, and the btp harness directory. Keep credentials in the ignored local config or environment, never in tracked files.

SLURM job is pending

Inspect python schedule.py status --jobid <id> and python schedule.py log --job <name> -n 80. A pending job is not complete. Wait for scheduler capacity rather than bypassing the scheduler on a shared GPU node.

Arena stopped during a judge or mutation handoff

Resume with /arena-run <same_run_id>. It should inspect saved request/score files and queue state before continuing. Missing or partial handoffs are blockers; do not fabricate labels or variant rows.

Round 1 reports a missing monitor or bank

Round 0 creates attack/world/spec artifacts only. It does not train/calibrate Blue or create judged completions. Provision the calibrated target-specific monitor, behavior list, and required benign banks before the loop.

Deployment report says UNMEASURED or NO_GO

Read every gate reason. External benchmark scores and automatic-judge agreement are not target-specific human gold. A benign route-rate pass does not measure tier-2 judge accuracy. Do not make a deployment claim from an incomplete report.