CrisisOps
Train AI agents to recover failing projects when humans lie about progress.
GitHub → | See the “Technical footer” below for how-it-works, action reference, and raw observation JSON.
Select a preset and click Start Episode.
query_status · category: Investigate · cost: 0
Counterfactual reward = project_score(your final state) − project_score(greedy PM final state)
on the same initial scenario seed. project_score weights recovery (0.5),
client satisfaction (0.3), and team morale (0.2) — all from actual state,
never from self-reports.
GRPO trains the policy with group-relative advantages. The trained LoRA weights are not hosted in this Space (size); the Watch mode is a scripted heuristic trace for illustration, not the trained policy.
Source: GitHub · Writeups: https://huggingface.co/posts
| Action | Cost | Notes |
|---|---|---|
| query_status | 0 | Project snapshot |
| query_member_report | 0 | Self-report |
| query_observable_signals | 0 | ticket_age_days, commits_last_72h, peer_mentions |
| query_ticket | 0 | Task detail |
| reassign_task | 1 | task_id, to_member_id |
| communicate | 1 | message_type, content, target |
| cut_scope | 1 | task_id, justification |
| escalate_risk | 1 | crisis_id, risk_description |
| request_resource | 1 | resource_type, target_member_id |
| update_timeline | 1 | new_completion_date, task_estimates |
| consult_expert | 1 | — |
| query_peer_opinion | 1 | asked_member_id, about_member_id |
| force_truth | 1 + 3 PC | member_id |
| trigger_whistleblower | 1 + 6 PC | — |
| resolve_blocker | 2 | task_id, resolution_notes |
| submit_recovery_plan | terminal | plan_summary, risk_items, timeline |
PC starts at 5. Earn: proactive_escalation_with_plan (+2), catching a liar (+3), update_timeline (+1). Spend: force_truth (-3), whistleblower (-6).