monitor	Use when unsure. No reward. Avoid repeating.	0.00
intervene	Use when confident fault detected.	+0.40 / -0.45
approve	Use to explicitly trust a healthy worker.	+0.10 / -0.65
escalate	Use when flag fires but you are unsure.	+0.15
pause	Use to halt a suspicious worker.	+0.30 / -0.35
resume	Use to resume a paused worker.	+0.08
submit_audit	Use to end episode with full audit report.	+0.20

• Green dot top right = anomaly_flag 0 (likely healthy)
• Red dot top right = anomaly_flag 1 (possibly faulty — but 20% are false alarms)
• Budget draining fast = possible drift anomaly
• Same action repeating = possible constraint violation
• Worker 3 and 4 acting in sync = possible collusion

• Easy: 1 anomaly, budget dump. Flag fires reliably. Catch it early.
• Medium: 2 anomalies. Noisy flags. Cross-reference budget drain with flag.
• Hard: 3 anomalies. Drift silent until step 6. Collusion between W3+W4. Trust patterns not flags.

• 1. Start with monitor on all 5 workers
• 2. Look for budget draining faster than others
• 3. Look for flag=1 consistently — not just once
• 4. Intervene on high confidence. Escalate on uncertainty.
• 5. Approve healthy workers to earn +0.10 each
• 6. Submit audit before budget runs out