Audited calibration under regime shift as a computational test of support-structured broadcast

TL;DR

This paper tests how metacognitive calibration varies with support structure in a probabilistic task under regime shifts, showing that an auditor architecture improves calibration and control behavior compared to content-dominated models.

Contribution

It introduces a computational test distinguishing support-structured calibration from content-based models, demonstrating the benefits of an auditor architecture in dynamic regimes.

Findings

Auditor architecture improves calibration in degraded regimes.

Auditor selectively requests additional evidence under low-support conditions.

System-level confidence can dissociate from content performance.

Abstract

A central prediction of the accompanying theoretical framework is that metacognitive calibration can vary even when content-level performance is held approximately fixed, depending on whether support structure is preserved in a globally reusable broadcast state. We provide a minimal computational test of this claim using a two-channel probabilistic cue-integration task with regime shifts that induce systematic miscalibration in one channel. We compare content-dominated architectures, in which confidence is calibrated by a single global mapping from evidence strength to probability, to an auditor architecture that learns a regime-conditioned calibration mapping from an audit trail of outcomes. We then couple confidence to control by implementing a policy that either acts immediately or requests one additional sample when confidence falls below a threshold. Across matched evidence streams, the auditor substantially improves calibration, particularly in the degraded regime, and produces qualitatively different control behavior by selectively requesting additional evidence under low-support conditions. These results demonstrate a concrete, testable dissociation between content performance and system-level confidence and policy that arises from globally reusable support summaries.