metasignal: New Python package unifies 17 metacognitive measures for sharper decision-making

Metasignal is a new open-source Python package that standardizes metacognition research by implementing all 17 measures benchmarked in Rahnev (2025), alongside classic signal detection theory (SDT) variables like d' (perceptual sensitivity), response criterion c, and mean confidence. Developed by Saurabh Ranjan, Mukesh Makwana, Konstantina Sokratous, and Brian Odegaard, the package runs one function over trial-level stimulus, response, and confidence arrays to output the complete set of metrics—covering three meta-d' family estimates (meta-d', M-ratio, M-difference), four nonparametric Type-2 measures (AUC2, Gamma, Phi, delta confidence) with their eight normalized forms, and two model-based measures (meta-noise, meta-uncertainty).

Beyond the core calculation, metasignal ships with a command-line interface, group-level summaries, bootstrap confidence intervals, permutation tests, and optional hierarchical Bayesian models—making it accessible for both quick analyses and rigorous statistical inference. It also includes information-theoretic measures, broadening its usefulness beyond psychology into AI and applied statistics. Currently optimized for binary two-alternative discrimination tasks, the package promises to reduce fragmentation in the field, where researchers previously had to juggle multiple ad-hoc scripts or reimplement published formulas. By unifying these metrics on a single platform, metasignal lowers the barrier to entry for studying metacognition and decision confidence, potentially accelerating research in cognitive science, human factors, and AI alignment where accurate self-assessment of model reliability is critical.