This thesis develops a threshold-based semantics for the Turkish indirect evidential marker that predicts unexpected discrepancies in its distribution. The marker’s behavior in interrogatives, so-called interrogative flip, is shown in turn to follow from the structure of discourse, as formulated by models that incorporate speaker commitment. I first establish that the indirect evidential marks information for which a speaker’s evidence is at best second-best, given general knowledge about the world. I then formalize this generalization in modal semantic terms and show that it explains the marker’s canonical absence in reports of well-known historical fact, as well as its optional presence in evaluative and mirative expressions. After examining the account’s predictions on the level of discourse, I discuss two corollaries: that evidential content in Turkish is not propositional, and that an at best second-best account brings to light anaphoric parallels between indirect evidentiality and the present perfect relative tense.