Symbolic execution has emerged as a powerful technique for scaling exact probabilistic inference to languages with more expressive features. But, this expressivity comes at a price: probabilistic programming languages based on symbolic execution are difficult to debug, optimize, and prove correct due to the many intricacies inherent to high-performance symbolic execution strategies. We aim to make it easier to work with probabilistic symbolic executors by developing symbolic sets, a new semantic domain that cleanly captures the notion of computation underlying symbolic execution. Just as a symbolic executor replaces ordinary execution with a lifted semantics, symbolic set theory replaces ordinary set theory with a lifted mathematics: the category of symbolic sets is a Grothendieck topos, which allows type theory to be used as a metalanguage for working with symbolic sets and functions. We prove a metatheorem that shows how a large class of definitional interpreters written in the internal language of symbolic sets are automatically correct for their ordinary set-theoretic interpretations. Using this metatheorem, we give the first full correctness argument for a symbolic probabilistic language with higher-order functions, type-directed state merging, pattern matching, and structural recursion.