MIT02 (mandatory checkpoints) is primary for 3 of 6 patterns. If you implement only one mitigation, this is the one. It doesn't require the agent to be smarter, doesn't require perfect training data, and doesn't depend on human judgment in the moment — it enforces review regardless of any of those factors.

Pattern 5 (dependency erosion) has only two primary controls — MIT04 and MIT09. These are both operational/training mitigations, not technical ones. No technology intervention directly addresses the gradual erosion of human independent judgment from long-term AI reliance. You can't buy your way out of Pattern 5; you have to operate your way out of it.

MIT08 (content provenance) is primary for both Pattern 2 and Pattern 4. Both involve attacker-controlled content reaching a human with AI credibility attached. Provenance labeling removes the credibility transfer in both cases — an agent that clearly labels "this claim came from an external email" cannot be used as a credibility relay.

MIT01 (uncertainty communication) is primary for Patterns 1 and 3. Both are fundamentally about the human receiving insufficient signal to know when independent scrutiny is warranted. Uncertainty communication is the signal that bridges that gap.