AI Control
How the field controls an untrusted, possibly-scheming model during deployment: the protocols, red/blue control games, monitors and side-task evaluations that keep safety despite intentional subversion, mapped across the field's research questions. This map reads the AI-control literature as a question ledger. AI control asks how monitoring, trusted/untrusted-model composition, editing, resampling and auditing can keep a deployment safe even when the untrusted model is adversarial and trying to subvert oversight. The ROOT axis RQ carries the field's research questions. The other axes are read per paper: PROTOCOL is the control technique studied (trusted monitoring, untrusted monitoring, defer-to-trusted, trusted editing, resampling, scalable oversight, or none); THREAT is the side task the untrusted model attempts (backdoor, exfiltration, sabotage, covert side-objective, collusion, research sabotage); SETTING is how realistic the environment is, from a toy APPS backdoor to a real codebase; MONITOR is the oversight channel (trusted action monitor, chain-of-thought monitor, white-box probe, human audit, or an ensemble); and GOAL is what the study is for (protocol design, attack capability, detection method, benchmark, measurement, or safety case). A cell is one full coordinate point; the papers there answer it, authored empty points are research gaps, and papers are weighted by recency and citations so the current front and the canon read heaviest. This map is complementary to the scheming map (a model's propensity to scheme) - control assumes adversariality and studies the protocol.