Systems that introduce self-correction as a first-class capability in language models — enabling the model to detect and revise its own reasoning errors without external supervision. Functionally, this reintroduces the "middle state" (processing/judgment) that standard inference pipelines skip
