Scheming & Deceptive Alignment
How the field evaluates whether frontier models scheme - pursue misaligned goals covertly, deceive overseers, sandbag or fake alignment - mapped across its research questions, the behaviors measured, the pressure applied, and the safety-case claims the evidence supports. This map reads the scheming / deceptive-alignment evaluation literature as a question ledger: the ROOT axis RQ carries the field's research questions about whether models scheme and how we can tell. The other axes are read per paper. BEHAVIOR is the specific scheming behavior a study elicits or measures (strategic deception, sandbagging, goal-guarding, oversight subversion, self-exfiltration, alignment faking, sabotage/backdooring, collusion, situational awareness, reward gaming). PRESSURE is how much elicitation the behavior needed, from fully natural to trained-in. SETTING is how realistic the eval environment is, from a toy prompt to a real deployment. EVIDENCE is how the behavior is read out (behavioral output, chain-of-thought, white-box probe, black-box interrogation, or an aggregate rate). GOAL is what the study is for (demonstrate a capability, measure a propensity, build a detector, test a mitigation, check measurement validity, or argue a safety case). SAFETYCASE is which safety-case claim the evidence bears on (inability, control, monitoring, trustworthiness). 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 deliberately complementary to the vcrc map (measurement validity) and refusal_geometry (jailbreaks / refusal directions).