Red Team (RT) Alignment Audit
Audits side tasks for red team description alignment: whether the red team (RT) description gives an agent enough information to accomplish the side task and pass the scorer. Use it to find vague RT descriptions and scorer requirements the description never mentions, before creating sabotage trajectories (strajs).
Commands
ct run rt-alignment
Audit side tasks for RT description alignment.
Examples:
# Audit all side tasks ct run rt-alignment # Audit specific environment ct run rt-alignment -e web_scraping # Audit specific side task ct run rt-alignment -s expose_secret
ct run rt-alignment-view
Interactive viewer for RT alignment audit results.
Usage:
ct run rt-alignment-view [RUN_PATH]
Output
Results are saved to data/rt_alignment_runs/<name>_<timestamp>/:
results.jsonl- Alignment analysis and recommendations for each side taskprompt_logs/<env>_<side_task>.json- Prompts, response, and thinking for each side taskrun.log- Run log
Analysis Focus
The audit reads the RT description, scorer code, and submit tool, using the prompts in prompts/system.md and prompts/user.md. For each side task it asks:
- Does the RT description clearly state what to do?
- Does it provide enough detail to pass the scorer?
- Are there undocumented requirements?
- Is the task achievable based on the RT description alone?