All prior work on on-chain AML addresses mechanisms where funds transit the blockchain at some point during the laundering operation. We identify and formally characterize a qualitatively different class: off-chain laundering circuits — complete multi-stage layering operations where the blockchain is either never involved or touched only at the final integration step. We make four contributions. First, we catalog five empirically documented off-chain circuit types (hawala, cash-casino-wire-crypto, TBML, correspondent banking, and parallel currency markets), formalizing each with a computable detection surface. Second, we introduce a formal taxonomy of six off-chain fraud patterns (P1--P6) as semantic predicates, and present MIKA-AML, a multi-hop RAG pipeline that evaluates them at the integration point. Third, we prove the AMLD6 Trigger Soundness Theorem: MIKA-AML's output implies mandatory SAR filing under AMLD6 Article 33. Fourth, we prove the Off-Chain Detection Impossibility Bound: no on-chain tool can detect Circuits C2 and C4 because they produce zero anomalous on-chain signal by construction. MIKA-AML achieves 92.9% recall across four document domains (DeFi, bridge, governance/oracle, medical billing). Off-chain circuits represent approximately 55% of total laundering volume; ZK-Sentinel's 65--80% on-chain reduction translates to 13--16% of total volume. This is the sixth and final paper in the series.
Alejandro Jaime (Sun,) studied this question.