Calibration Trap

An epistemically self-aware agent should hold a view about how reliable its own judgment is, and should discount its judgments accordingly. But the agent’s assessment of its own reliability is itself one of its judgments, and so must be discounted by exactly the reliability it purports to establish. This paper formalizes the resulting self-referential structure. Modeling self-assessment as a self-trust function mapping candidate reliability estimates to the credence they warrant, I prove three things: (i) coherent self-assessments always exist — they are fixed points of the self-trust function, guaranteed by elementary topology; (ii) they need not be unique — natural self-trust functions admit multiple equilibria, including a “confident” and a “skeptical” fixed point; and (iii) distinct equilibria can be internally indistinguishable: two agents with identical evidence, identical reasoning, and identical experiences may occupy different self-trust equilibria, and no reflection either can perform will reveal which one it occupies. The trap, then, is not incoherence — Humean skepticism about reason turns out to be too strong — but radical underdetermination: self-calibration has solutions, but choosing among them is not an epistemic act the agent’s own faculties can underwrite. I distinguish the trap from nearby results on higher-order evidence and anti-expertise, examine four escape routes (track records, other minds, pragmatic vindication, and default entitlement), argue that each at best relocates the fixed point rather than verifying it, and draw consequences for epistemic humility and for the self-evaluation of artificial reasoners.