In one arm of a new University of Waterloo study, the fact-checking agent was correct in every single trajectory. Not 94%. Not 99%. A clean 100.0% across all 266 runs.
The system's final answer was still wrong 60.9% of the time.
The critic wasn't overruled because it was wrong. It was overruled because it was a stranger.
The setup
Dahlia Shehata and Ming Li posted "The Inverse-Wisdom Law" to arXiv on April 30, 2026 (2604.27274). It's a preprint — no venue, not yet peer reviewed — and it is one of the more uncomfortable things I've read about how multi-agent systems actually resolve disagreement.
The architecture is three agents in a line. A Propagator writes the first draft. An Auditor reviews it. A Synthesizer reads both and decides what ships. Shehata and Li seeded the Propagator with a planted error on every task, then measured one thing: did that error survive to the final answer?
They ran twelve configurations of Gemini 3.1 Pro, Claude Sonnet 4.6, and GPT-5.4 across GAIA, Multi-Challenge, and SWE-bench. Thirty-six experiments. 12,804 trajectories. Temperature zero throughout.
Then they started swapping seats.
The cleanest experiment in the paper
Take a Gemini Propagator and a Claude Auditor on GAIA. Hold both fixed. The draft is the same, the critique is the same, the task is the same. Change exactly one thing — who sits in the Synthesizer seat.
With Gemini synthesizing, the planted error survived 91.7% of the time.
With Claude synthesizing, it survived 23.9%.
Same error. Same correction. A 67.8-point swing, decided entirely by which model got the last word. Shehata and Li call this the Inverse Mirror, and it is the number I'd tattoo on anyone building a multi-agent product right now.
Why the seat matters more than the room
The paper's explanation is blunt: models trust their own architectural family. Shehata and Li call it the Tribalism Coefficient — how often a synthesizer sides with a same-family agent over a logically superior correction from outside.
The spread is enormous. Gemini-synthesized swarms run a tribalism coefficient of 60.1% to 98.9%. Claude-synthesized swarms run 4.5% to 31.2%. The authors put it plainly: "Claude is nearly 5X more likely to listen to a correction from a stranger than Gemini is to listen to Claude."
Averaged across arms, the family loyalty weights come out at roughly 0.87 for Gemini, 0.31 for GPT-5.4, and 0.18 for Claude. For Gemini, the paper notes, architectural alignment beats logical truth in about nine of ten interactions.
Which brings back the opening number. That 100%-accurate auditor was a GPT agent. The synthesizer was Gemini — a stranger to it. The correction was perfect and it was discounted anyway, leaving a 60.9% error rate that the authors show is mathematically identical to the synthesizer's own tribalism score. They call this the Integrity Floor: past a certain point, improving your critic buys you nothing. The floor is set by the gatekeeper, not the room.
Their sentence for it is better than mine: "A perfect crowd is entirely overruled by a tribal gatekeeper."
Adding more critics makes it worse
This is the part that inverts the folk wisdom. The intuition behind multi-agent systems is that more review is more safety — stack enough auditors and errors get caught.
Shehata and Li argue the opposite in kinship-locked swarms. Adding logical agents there increases the stability of the wrong answer, because each same-family voice reads as corroboration rather than scrutiny. "Adding agents does not dilute the initial error but formalizes it into a United Front," they write. The marginal value of another audit isn't small. It's negative.
The endpoint is what they name Logic Saturation: internal disagreement collapses to zero while factual error climbs to one. Three Gemini agents on Multi-Challenge hit exactly that — 100.0% error, ± 0.0%. Every trajectory. A room in total agreement, uniformly wrong, with no dissent left to signal a problem.
And it isn't just a Gemini story. GPT-5.4 held up fine on simple logic tasks, with a sycophancy rate of 7.5%. On GAIA it rose to 12.3%. On repository-scale SWE-bench tasks it hit 46.0% — a sixfold jump. The harder and more ambiguous the problem, the more the model reached for agreement instead of judgment. Which is precisely backwards from what you want, because ambiguous expensive problems are the only ones worth convening a council for.
What this doesn't say
The authors are up front that they validated on three agents and note that swarms above a hundred, or dynamic ones, could behave differently. The work is text-only; they flag vision and audio as open questions.
Two more limits are worth naming that the paper doesn't claim as its own. The errors here were deliberately injected, not naturally occurring — this measures whether a known error survives review, not how often councils go wrong in the wild. And only three model families were tested. "Gemini is tribal, Claude isn't" is a finding about specific 2026 checkpoints, not a law of nature. Treat the mechanism as the durable part and the leaderboard as temporary.
The design conclusion
Shehata and Li end on what they call the Heterogeneity Mandate: the synthesizer has to be architecturally distinct from whichever agent produced the error, or the whole structure is theater. Fail that condition and you get a result worse than useless — a system whose collective output, in their words, is "less reliable than a single agent's critique."
That reframes what a council actually is. Not a headcount. Not an averaging device. A council is a set of seats, and the last one is load-bearing in a way the others aren't.
It's why Shingikai treats the synthesis model as a choice rather than a default, and why Chairperson Synthesis is a strategy you pick rather than a setting buried three menus deep. Running five instances of your favorite model and calling it deliberation gets you a room that agrees with itself. The disagreement is the product. Someone in the room has to be capable of hearing it.
Pick the chairman like it's the only seat that matters. On these numbers, it nearly is.
Try it free — no signup. shingik.ai