Retrieval-Grounded Reasoning. A reasoning path where governed evidence is retrieved before the model generates, the model is constrained to reason over that evidence, and the output preserves the sources that support its claims.
Retrieval-grounded reasoning is the positive shape of Law V. The model stays inside the parametric memory boundary: it does not answer from parametric memory or free association. It receives the question, the admissible evidence, and the citation rules together, then produces a derivation that can be checked against its sources.
The failure modes are ungrounded confidence and citation theater. A system is not retrieval-grounded merely because it shows links near an answer; the retrieval must precede reasoning, constrain the answer, and preserve the derivation.