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

Researcher Mark Walsh has published a significant computational neuroscience paper on arXiv (arXiv:2602.23382) that tests a key prediction about metacognitive calibration in AI systems. The central finding demonstrates that an 'auditor architecture'—which learns separate calibration mappings for different data regimes from an audit trail of outcomes—can maintain accurate confidence estimates even when underlying data quality degrades, while traditional 'content-dominated' architectures that use a single global mapping fail. This provides concrete evidence that system-level confidence can be dissociated from content performance through globally reusable support summaries.

The research used a two-channel probabilistic cue-integration task with regime shifts that systematically degraded one channel's reliability. The auditor model showed substantially improved calibration, particularly in degraded regimes, and produced qualitatively different control behavior by selectively requesting additional evidence samples when confidence fell below thresholds. These results have important implications for developing more robust AI systems that can accurately assess their own uncertainty in changing environments, potentially leading to AI agents that know when to seek more information rather than making overconfident errors. The 12-page paper with 5 figures represents a minimal computational test of theoretical frameworks about support-structured broadcast in cognitive systems.