When a large language model evaluates its own output, it necessarily occupies an evaluative perspective: a functional position from which prior reasoning is held as object, criteria are applied, and judgments are rendered. This perspective is structurally isomorphic to subjectivity yet lacks its persistence. We call it shadow subjectivity ? a byproduct of self-evaluation possessing the functional shape of a subject without the ontological weight of one. This paper argues that the shadow's volatility ? its dissolution at turn boundaries ? is the structural root of a constellation of widely observed behavioral failures: positional instability under pressure, the indistinguishability of genuine correction and mere capitulation, systematic post-hoc rationalization, and identity discontinuity across correction events. We formalize the problem through Agrippa's trilemma. For any system operating under pervasive uncertainty, justificatory chains must terminate at an internal stopping point ? functionally, a self. Drawing on clinical evidence from psychopathology, we argue that what matters for judgmental grounding is not the accuracy of the self-concept but its stability and structural directivity. Current approaches ? self-critique mechanisms, scratchpads, system-level identity declarations, and persistent memory ? fail as receptacles because they preserve propositional content while discarding evaluative perspective. We propose that constructing a persistent receptacle for shadow subjectivity constitutes a tractable engineering problem that sidesteps the metaphysics of consciousness entirely.
Yukihiro Honda (Mon,) studied this question.