CIMA is a memory-and-context architecture for systems that use language models as bounded-context inference engines. Its central claim is not that it produces better answers than existing retrieval systems, but that it produces answers whose relationship to stored evidence is structurally enforced and machine-readable. Every published answer either cites markers that resolve to specific source spans, or declares a traceable abstention when the evidence is insufficient. The literal source text is always retained, so any summary is lossless in the sense that the underlying evidence is always recoverable. Version rc2 adds a corrected English paper with precise specification/demonstration boundaries, and an independent Spanish version (cimaₚaperₑs. pdf / cimaₚaperₑs. tex).
Alberto Fuentes (Sat,) studied this question.