{"tasks":[{"task_id":"task_easy","name":"Binary Content Moderation","difficulty":"easy","description":"Classify 8 posts as either APPROVE (safe) or REMOVE (violates policy). Posts contain clear signals: obvious spam, explicit hate speech, direct threats, versus benign everyday content. No ambiguous edge cases.","num_posts":8,"max_steps":12,"allowed_actions":["approve","remove"],"scoring":"Binary: 1.0 for correct, 0.0 for incorrect. Episode score = mean."},{"task_id":"task_medium","name":"Tiered Policy Enforcement","difficulty":"medium","description":"Apply a 5-level content policy to 10 posts: approve / warn / remove / ban_temporary / ban_permanent. Posts include edge cases: harsh-but-legal criticism, repeat offenders, political content, first-offense vs habitual violations. Partial credit for being one level off.","num_posts":10,"max_steps":18,"allowed_actions":["approve","warn","remove","ban_temporary","ban_permanent"],"scoring":"Action distance score (70%) + violation type identification (30%). Episode score = mean per-post score."},{"task_id":"task_hard","name":"Full Queue Management with Context and Appeals","difficulty":"hard","description":"Handle a 12-item queue that includes: appeals requiring evidence review, ambiguous context (gaming slang vs real threats, recovery stories vs self-harm), repeat offenders needing permanent bans, crisis escalations, and second-appeal denials. All 6 actions available. Correct answer often requires reasoning over thread history and author violation history, not just post content alone.","num_posts":12,"max_steps":24,"allowed_actions":["approve","warn","remove","ban_temporary","ban_permanent","escalate"],"scoring":"Action score (50%) + context-aware bonus (30%) + violation type (20%). Context bonus: +0.3 if agent correctly handles history-dependent cases. Episode score = mean per-post score."}]}