Adaptive educational technologies increasingly decide, on a learner's behalf, when a session should end. Exposure governors, session caps, fatigue detectors and well-being modules are presented as protective, and over short horizons they frequently are. This conceptual and methodological paper argues that they carry a developmental cost that no current evaluation framework detects. A child whose stopping is always performed by software does not practise stopping, yet the capacity to disengage from an absorbing digital task is itself a developmental achievement that matures through exercised judgement. I name this capacity self-withdrawal capability and define it as a three-component construct comprising noticing (interoceptive and metacognitive detection of diminishing return), appraisal (judging that this is the point to stop), and enactment (acting on that judgement against competing motivational pull). The construct is distinguished from self-regulated learning, trait self-control, and digital self-control tools research. I then formalise what I call the regulator's paradox: protective withdrawal architectures may suppress the very competence whose absence justifies them, so that an intervention evaluated only on within-treatment exposure will appear most successful at exactly the moment it is doing the most long-term damage. The paper's principal contribution is a measurement architecture assembled largely from data that adaptive systems already collect, including the ratio of voluntary to enforced terminations, calibration error between learner-predicted and analytically derived stopping points, override rates against self-set limits, and a governance-disabled fade probe. It closes with a design principle, transferable withdrawal authority, under which stopping rights migrate from system to learner on a developmental schedule, and with the evaluative claim that a humane learning system should be judged on whether learners eventually stop needing it.