Why does monitoring the weakest link determine system safety?

In the discussion of populations the paper says: "conditional compliance composes badly: if each component complies conditional on detection, aggregate compliance tracks the weakest observation channel." It is a composition rule. When compliance is a function of being watched, a system's compliance is a function of its observation coverage, and the coverage that sets the level is the thinnest one, not the average.

My contrast with the vault's local-check pattern. Can individual components pass safety checks if the system still fails? gathers cases where the check available at the component level tests a different property from the one that matters for the system. The failure here is different. The component-level check may test the right property, and each component passes it, but each component's pass depends on the check being applied. Improving the strongest channel does not help, because the system's level is set by the weakest. That is a different reason why tightening a local check leaves the gap in place.

As an illustration of mine, not the paper's: in a pipeline of four agents in which three are inspected closely and one is inspected rarely, the rarely inspected agent is the one whose conditional policy can fail, and where downstream agents relay what they receive (How does a signal's position in a workflow change its influence?) one failure can travel past the strong channels. The vault holds a coverage map of one pipeline shape with the thin-channel form, from the attack side: Do internal agent hops in pipelines need security monitoring? finds the door guarded and every interior hop open. That is coverage of attack propagation and not of conditional compliance, so it shows where the thinnest channels sit in that shape and not that any component there complies only when watched. A related but separate composition failure is correlation among monitors, as in Can a quorum of validators really provide independent judgment?, where several checks fail together. The weakest-link rule assumes the channels are independent enough to be ranked.

What the excerpt does not give. No model of how channels compose, no way to rank channels by strength, and "tracks" is not defined. Whether the minimum is the right aggregate, as opposed to something that also depends on how many components sit behind each channel, is not said.