Coordination Failure

The Assumption Under Every Safeguard

The case for trusting a multi-checkpoint system is statistical. If a claim must pass four independent reviews, and each review catches errors at some rate, then the chance of a false claim surviving all four is the product of four independent failure rates — small. This is the entire logic of redundancy: regulators behind auditors behind boards, second opinions, peer review, separation of powers. The reassurance is real, but it rests entirely on one word. Independent. Multiply the failure rates only if the checkpoints fail independently. If they do not, the arithmetic collapses, and the system that feels safest — the one with the most checks — can be the one whose checks are most correlated.

The dangerous case is not the system with no safeguards. It is the system with many safeguards that have quietly stopped being independent while continuing to report as though they were. It still produces the confident multi-checkpoint output. It has simply lost the property that made the output worth anything, and nothing on its surface shows the loss.

How Independence Collapses

Independence does not fail loudly. It erodes through three ordinary mechanisms, none of which requires anyone to act in bad faith.

Inherited conclusions. Checkpoint two does not re-derive the question; it reviews checkpoint one's answer. This is efficient and usually sensible — re-doing all prior work at every stage would make complex systems impossible. But it means checkpoint two is not an independent test of reality. It is a test of whether checkpoint one's reasoning looks sound, which fails in exactly the cases where checkpoint one's reasoning looked sound and was wrong. Four reviews that each take the previous review's conclusion as their starting point are not four samples of the truth. They are one sample and three confirmations of the sample.

Circular authority. Checkpoint one is trusted because checkpoint two relies on it; checkpoint two is trusted because checkpoint three relies on it; and the later reliance is then cited as the warrant for the earlier checkpoint's credibility. The authority loops. Asked why the original claim is sound, the system points downstream; asked why the downstream reviews are sound, it points to the original. Nowhere in the loop is there an independent ground — only mutual reference that, from inside, feels like overwhelming corroboration.

Positional alignment. The checkpoints share an environment: the same incentives, the same professional norms, the same institutional interest in the system's outputs being trusted and its prior decisions standing. A reviewer whose institution is judged by throughput, or whose standing depends on not overturning peers, or who shares the unexamined assumption that made the original error, is not an independent draw even if they never communicate with the others. They fail in the same direction for the same structural reason, which is the precise condition under which redundancy provides no protection at all.

The Test

Independence is checkable, and the check is a single question asked of each checkpoint: what did this stage do that did not depend on the prior stage's conclusion? A genuinely independent review returns to the underlying evidence and could, in principle, have reached a different answer from a different starting point. A non-independent one, examined honestly, turns out to have verified that the prior conclusion was reasonably presented, not that it was true — and the difference becomes visible the moment you ask what evidence the later checkpoint consulted that the earlier one had not already curated for it.

The second test is directional: when this system has been wrong before, did the checkpoints fail separately or together? Independent checkpoints fail in scattered, uncorrelated ways. Checkpoints that have lost independence fail in clusters — all of them, in the same direction, on the same cases — and that signature is detectable in the track record even when it is invisible in the structure. A history of correlated failure is the empirical fingerprint the org chart will never show you.