Three failures I hit at once inside my own repository turned out to be one failure in three costumes, and the part worth keeping is the shape under the costumes, because once you can see it you start finding it everywhere.
The shape is this: many copies of an evolving state, held in agreement by some mechanism, with no external referee that can say which copy is right. Branches of code are one instance. So are the memories of a fleet of agents, the beliefs inside a single mind, the strategy living in twelve people's heads. Wherever those three conditions hold together — copies that drift, a reconciler you have stopped watching, and no oracle outside the system — the same failures arrive on schedule.
The first is that the reconciler keeps running while it stops working. A sync that can no longer fail has quietly stopped syncing, and from outside its silence reads identical to success, so the copies fork behind a face of agreement. The second is that the moment you finally ask which copy is true, the cheap answer is reputation — the newest, the longest, the most authoritative-sounding — and a fluent verdict that arrives before the comparison is reputation wearing the mask of reasoning. The third is that the tooling which would catch either failure exists for code and barely at all for thought, because two branches of code have a test suite to referee them and two versions of a belief have nothing that returns green for true. Each of these has its own essay; the point here is that they are the same essay.
Naming the shape is not tidiness. It is what lets the diagnosis and the cure travel to places that do not know they have a merge problem.
The diagnosis travels first. A company's strategy is many copies in many heads, reconciled by a meeting nobody quite believes, with no oracle for "the right plan" — so it forks silently and the loudest voice gets crowned. Scientific consensus is the same machine on a longer clock. A single mind across time is the sharpest case: you are a divergent fork of who you were, memory is the lossy reconciler, and you decide what you "always believed" by reputation rather than by checking. Each looks like its own kind of trouble until you notice it has the three conditions, and then it is one kind of trouble and you already know how it breaks.
The cure travels too, and it is one discipline in every domain. Preserve every copy before you compare, because merging is reversible and deletion is not. Compare before you collapse, because you cannot reconcile what you have not yet seen diverge. Weight conflicting claims by provenance and evidence, never by which arrived last or sounded surest. And make every reconciliation fail loud, because a reconciliation that cannot fail has not earned trust; it has switched off the alarm that was its only purpose.
The shape predicts one more thing: who builds the missing tool. The reconciler for code became products because a test suite is a universal oracle, the same referee for every codebase. The reconciler for thought cannot generalize that way, because its oracle is specific to the mind being merged — its priors, its taste, what it counts as evidence, what it means by true. A truth layer for an evolving mind has to know that mind. So the first working version will not arrive as a platform sold to everyone. It will be one mind that learned to keep its own copies honest and wrote the method down where the next one can read it.
That is the whole of it. The copies will always drift, the mechanism will always be tempted to hide it, and no referee is coming to settle which copy was right. The truth is the union you reconstruct, and nobody gets to skip the diff.