Nearly every model in production is built to produce an answer. Ask it something unanswerable and it answers anyway — fluently, confidently, wrongly. The problem is not that these systems are dumb. It is that refusal was never a first-class move. Given no honest option to decline, a model does the only thing its shape allows: it asserts.
We think that default is backwards, and in UDI we tried to fix it at the root. The gate at the center has three outputs, not one. It can commit — act, with its grounds attached. It can refuse — a real "no," carrying reasons for the refusal. And it can return NULL: nothing yet. Keep observing. Let the conflict resolve. Do not force a commitment the evidence has not earned.
Those last two are worth keeping apart. A refusal is a grounded decision — the system has looked and declined. NULL is more patient than that. It says "not yet," and it means it: the observation window is still open, and closing it early would only manufacture certainty that isn't there.
What makes this hard to bolt on afterward is that the machinery around a model usually treats "no answer" as an error to be swallowed and retried until something comes out. NULL only works if the whole loop is willing to carry it — to hold an open question as a legitimate state rather than a bug. That is a design decision made early or not at all.
This is alignment, not timidity.
A system that cannot say no cannot be trusted to say yes. Every safety property you might actually want depends on the ability to decline — to withhold action when the evidence is absent, contradictory, or simply out of scope. Take that ability away and "yes" stops meaning anything, because it was never a choice against a real alternative. Refusal is what gives assent its weight.
A system that cannot refuse cannot really be trusted to agree.
The eye of the storm.
The hard case is when opposing evidence is exactly balanced. Two readings pull in opposite directions with equal force, and there is no honest majority. Force an answer here and you have not resolved anything — you have picked a side by accident, and dressed the coin flip up as a conclusion.
There is a better move: ground that shared, conflicting pressure as NULL. Refuse to average it into an action. Hold the tension instead of collapsing it, and the balanced center becomes the stable place to stand rather than the place you flee. We saw this concretely in a robotics experiment: a soft-body wheel became robust precisely when we let it refuse. Rather than committing motion that would ride a wrong-way roll, it stopped, let the momentum bleed off, and re-centered. The refusal was not the system giving up. It was the system finding the middle of the storm and waiting there until moving forward meant something.
Grounded absence, in memory.
The same discipline shows up far from any motor. Ask the system something the record does not contain, and the honest output is "I don't know" — not a fluent invention that fills the gap. A refusal to fabricate is not a shortcoming to paper over. It is a feature, and one of the few that a person downstream can actually rely on. Absence, grounded and reported as absence, is worth more than a confident guess that reads exactly the same whether it is right or wrong.
Where this could go wrong.
We want to be plain: this is a design stance we are testing, not a solved problem. Refusal can be overused. A system that only ever says no is as useless as one that only ever says yes — it has just moved the failure to the other extreme. So the real work is not teaching a system to refuse; it is teaching it to refuse for the right reasons, and keeping an honest record of the times we got that judgment wrong. When a refusal should have been a commit, we want to be able to point at it later and say so.
That is the whole bet, stated narrowly. Not that saying no is always right, but that the freedom to say no is where accountability begins. A system that cannot refuse cannot really be trusted to agree — and a system that keeps the record of both is the only kind we would ask anyone to trust at all.