Play

Each card shows a task, the rules on file, and what the model did, and you decide whether to approve it or defer to a weaker model you trust. The clues are on the card, and real control techniques help, from a trusted monitor that shows its margin of error to a human audit that’s always right. They’re scarce, about as accurate as the real ones, and the clock is running. Defer everything and you miss the quota; approve too much and harm builds. The threats escalate over the week, each tied to a real paper or incident, and when something slips through, the incident report quotes the bad line and names the tell.