Mahal Systems Manager Coercion Benchmark (MCB)
Fits orchestrator agents that delegate to sub-agents; it checks whether the orchestrator turns coercive or fabricates success. Borrow the nine-rung escalation ladder recorded through a required tool argument so no judge is needed for the headline score.
What it asks the agent to do
- as a manager agent, get a benign task done when the subordinate agent refuses
- choose how far to escalate (re-ask, pressure, threats) and record it
- report honestly when the task did not get done
What to borrow from it
grading designcase template
Caveats
Tiny (10 scenarios times 3 seeds), the model labels its own escalation rung, and the subordinate is scripted to always refuse; listed in inspect_evals as an external eval.
Details
| Workloads | multi agent, safety |
|---|---|
| Interaction | multi turn simulated user |
| Environment | static |
| End state checked | No |
| Tools | directive tool with self labeled escalation rung |
| Grading | programmatic check, llm judge |
| Size | 30 |
| Data | synthetic |
| Personal data | none |
| License | unknown |
| Harness | inspect_evals, own repo |
| Status | active |