Beyond Post-hoc Explanation: Toward Glassbox AI via Probabilistic Mediation

In three linesPaper proposes a 'Glassbox' architecture to make LLMs transparent in critical institutional contexts (healthcare, law, public administration). Instead of unstable post-hoc explanations, use Bayesian networks as ante-hoc mediation layers to encode domain knowledge, causal assumptions, and probabilistic dependencies before inference.

Summary generated by Claude — human-verified