DEJAN optimizes grounding snippets through an iterative Bayesian acquisition loop that seeds causal hypotheses, applies localized text edits, and updates posterior belief distributions based on rank improvements scored by an AI judge model.