Most people grading an AI council are measuring the wrong thing. They ask whether the answer was right. A new paper from AWS's Generative AI Innovation Center and General Motors (arXiv:2606.29654, submitted late June) measures something almost nobody in the consumer-council world tracks: whether the system knew when to shut up.

They give it a name — acted-on accuracy. Not how often the council is correct across every question you throw at it, but how often it is correct on the questions it actually chose to answer. The gap between those two numbers is the whole story.

A council that always answers is a liability

Here is why the distinction matters, and why it isn't academic hair-splitting.

The failure modes of multi-model systems are documented, not hypothetical. Fill a council with several agents and hand them a problem where the key facts are split across them, and they tend to herd on what everyone already shares rather than surface what only one of them knows. In HiddenBench — a 2025 controlled benchmark built on the classic Hidden Profile setup — multi-agent LLMs scored 30.1% under distributed information, against 80.7% for a single agent handed everything. That is a council underperforming one model, badly, on exactly the kind of question a council is supposed to be good at.

Judges have their own tell. When a model grades output from its own family, it tends to reward its own style — which is why production evaluation teams now describe cross-family judging as something close to mandatory rather than a nicety. And plain majority voting can land on the wrong answer even when the individual agents were individually right.

Put those together and you get the uncomfortable case: a council can be confident, unanimous, and wrong, all at once. If it answers every question with the same steady voice, you have no way to tell that case apart from the good one.

The next gate isn't a better answer

For a few years, council design has been a race to make the answer better — more models, sharper synthesis, judges drawn from a different family than the thing they grade. This paper points at a different gate, and it sits on top of all of that.

The mechanism is an abstention rule with a spine. At each round of the debate, the system maps the exchange so far to a compact state, then computes a k-nearest-neighbor lower confidence bound on how likely that state is to be correct, using calibration data. It acts only when that bound clears a threshold the operator sets. Otherwise it defers to a human. The authors are careful about what they are promising: the guarantee is "conditional, not distribution-free" — it holds under stated assumptions, not as a law of nature.

The reported behavior is the part worth sitting with. Across six benchmarks and nine baselines, the method used 9–12% of its pre-declared error budget on the datasets it engaged, reaching up to 84% automation at 96% acted-on accuracy. On deliberately hard stress-test data, it deferred rather than force an unreliable answer. Read those as six-benchmark experimental results under the paper's own assumptions, not a universal promise. But the shape is unmistakable: the system got most of its accuracy by declining the questions it wasn't equipped to answer.

That is a council whose competence is measured partly by its willingness to stay quiet. A council that always answers is a liability. The skill is knowing when not to.

What this looks like when you can watch it

Most council products hand you a single resolved answer and no way to feel how load-bearing it is. You get the verdict, not the room.

The value of watching a council deliberate in real time is that the confidence bound this paper computes has a human-readable twin: visible disagreement. When several models land in the same place for the same reasons, that unanimity is a signal you can act on. When they fracture — one model raising an objection the others were structurally blind to, a different-vendor model refusing to sign off — that fracture is the human-readable version of a bound that failed to clear the threshold. It is the moment to bring judgment in, not the moment to trust the majority.

This is why the two decisions are worth separating. A synthesis layer — what Shingikai calls Chairperson Synthesis — decides what the council's position is. An act-or-defer certificate decides something different: whether that position is trustworthy enough to act on without a human. The first is a question about the answer. The second is a question about the answer's answer. Collapsing them is how you end up acting on confident nonsense.

The harder skill

Almost nobody has built the second gate. It is easier to add a ninth model than to teach a system to recognize the states where its own consensus is not to be trusted. It is easier to ship a product that answers every question with equal confidence than one that sometimes says "this one needs a person."

But the easy version is the one that burns you, because the failure is silent. A council that herds, or a judge that flatters its own family, produces an answer that looks exactly like a good one. The only defense is a system that knows the difference between a question it has earned the right to answer and one it hasn't.

Knowing when not to answer is harder than answering, and it is worth more than another model in the room.

Try it free — no signup. shingik.ai