Every serious multi-agent stack bolts a verifier onto the end. A critic agent, a judge, a second pass whose job is to catch the hallucination before it ships. The assumption underneath is quiet and almost never stated: a checker can only help. Worst case, it does nothing.
A new paper (arXiv:2606.27409) says the assumption has a cliff in it.
The setup is the one everyone recognizes. Agents pass beliefs around a network and update toward each other. Somewhere in that network sits a corrector, a node grounded in the truth, whose job is to pull the group back when it drifts. The catch is the detail every real system has and most analyses ignore: the correction arrives late. "During this delay," the paper notes, "false claims can propagate through the agent network." By the time the referee speaks, the room has already moved.
Push the critic harder, do worse
Here is the part that should stop you. The author ran the correction at two strengths on a grounded factual task. A strong, grounded verifier reached the truth in 16 of 20 runs. A more forceful one — same grounding, just pushed harder — reached it in 8 of 20. Turning up the critic cut the hit rate in half.
That is not a rounding error or a bad seed. It is the shape of the whole result. A verifier that corrects too hard, too late, does not merely fail to help. In the paper's linear model of the dynamics, it destabilizes the system "through an oscillatory mode": the majority answer swings toward the truth, through it, and out the other side, then swings back. The council doesn't converge. It rings.
A verifier is a force, not a safety net
The reframe worth keeping is small and load-bearing. A verifier is not a safety net you can only fall into. It is a force. And force applied late, to a target that is already moving, overshoots. Strength and timing stop being free parameters you crank to taste and become a dial with a cliff on one side of it.
The paper even locates the cliff. When communication and verification share the same delay, their effects add, and at a delay of two steps the threshold for the combined "dose" lands on the inverse golden ratio, about 0.618. Below it, the group settles. Above it, the group oscillates. There is a number where a helpful correction becomes a destabilizing one, and it is not intuitively large.
Two reasons not to rip the verifier out
The first is grounding. The danger case in the experiments is the critic that forces a correction the agents must adopt wholesale, or a contrarian that isn't anchored to fact at all. A verifier grounded in the actual answer behaves. The forceful, less-grounded one is the one that halved the hit rate. The honest headline is narrow, and it is not "verifiers are bad." It is this: an ungrounded or overbearing critic is dangerous, and a grounded, moderate one is the thing you wanted all along.
The second is placement. A corrector is a budget. You cannot ground every node, so where you spend the few you have is a decision, and the paper shows it is an optimizable one — a supermodular objective where greedy assignment captures at least (1−1/e) of the best achievable coherence reduction. Put differently: a council is a topology, not a roster. Who checks whom, and when, is architecture.
Carry the limits honestly, because they are real. This is a theory paper with a small confirmatory experiment. The core analysis is linear and local, valid near the truth rather than everywhere. The controlled overshoot study runs on eight questions, and its strength is the near-total separation it shows — overshoot in 96 to 100 percent of runs at the worst delay, versus 0 to 4 percent when the correction is prompt — not a large sample. The five models tested (Qwen3.6-35B, Qwen3-14B, Mistral-7B, Phi-4, Gemma-4-12B) are small to mid-size open-weight models, not frontier systems. The placement rule, so far, is validated on the math surrogate, not on live agents. This is a sharp idea with a modest proof, not a settled law.
But the idea is the point, and it aims squarely at the part of a council most people treat as an afterthought. The roster debate — which models, how many — has had a year of attention. This is about the referee. When it speaks, and how hard, decides whether the same models land on an answer or orbit one.
The synthesizer has a cliff behind it
That is the argument we have been making about Chairperson Synthesis the whole time, now with a threshold attached. The layer that decides what to do with disagreement is not a summarization prompt you append at the end. It is the seat with a cliff behind it. A synthesizer that waits too long and then corrects too hard can take five models that were converging and set them oscillating.
A single model never has this problem. It also never catches itself. The fix was always more than one model in the room. This paper is a reminder that the room needs a referee who knows when to speak, not just a louder one.
You can watch a council work through a hard question, synthesis layer and all, for free. No signup. shingik.ai