<p>Shame is often portrayed as an evolutionarily ancient social emotion that manages reputation through appeasement and withdrawal. Yet it also seems capable of moral autonomy. It can arise in private, persist without witnesses, and motivate self-criticism and repair even when social approval is absent or explicitly rejected. This dual profile generates the tension between heteronomous shame, which is responsive to social evaluation, and autonomous shame, which is responsive to self-endorsed moral standards. Building on Kurth’s identity-based account, I develop a constructionist predictive processing reconstruction that makes the underlying architecture explicit. On my view, shame is constructed when temporally deep self-models and audience models jointly generate an identity-level disturbance. Whether an episode takes a heteronomous or autonomous shape depends on how a predictive control system allocates precision across competing expectations, such as audience-facing cues and moral identity priors that structure self-conception over time. This framework models graded, mixed, and shifting shame profiles within a single control architecture, rather than requiring distinct kinds of shame. Finally, I argue that the evolutionary evidence often cited in support of a dedicated shame module can be accommodated by selection for domain-general predictive and social-regulatory capacities, with shame emerging as one way those capacities are recruited in norm-governed social niches.</p>