Damage lost while learning
Protocol, evidence and limits
Every qualifying task changes which response works. Admission requires at least 12 HP of loss from keeping any previously optimal choice, in both directions, plus an observable reward gap. All 32 candidates per seed are retained, including failures. This is a selected challenge population.
Recovery means three consecutive pre-choice distributions losing no more than 35% of the available best–worst damage range. Cases already within that criterion are separated; cases that never reach it stay in the denominator. Context tests use frozen memory, so they measure learned situational choices rather than recovery.
Reset memory is an externally timed experimental control. It does not demonstrate automatic detection of a changed player strategy. The squad test uses a joint shared-reward learner; its independent-role controls receive the same reward.