We let five AIs argue before we trust the answer
There’s a specific kind of decision where AI advice gets dangerous: the hard call. Should I raise prices? Kill this product line? Make this hire? These aren’t lookups — they’re judgment calls with real downside, and they’re exactly where a single confident answer is worth the least.
A language model, asked once, produces one plausible line of reasoning and commits to it. It doesn’t naturally generate the strongest case against itself. Neither do people — which is why every serious decision-making culture ever built converged on the same tool: structured disagreement. Red teams. Devil’s advocates. Someone whose job is to attack the plan before reality does.
A committee that actually disagrees
So for hard calls, Nerve doesn’t ask one model for the answer. It convenes several voices with genuinely different mandates and lets them argue. One builds the strongest case for acting. One builds the strongest case against. Others weigh in from different angles — cash, customers, timing. Then a devil’s advocate attacks whatever consensus is forming, on principle. The voices vote, and crucially, the dissent is preserved: a 4–2 verdict is reported as a 4–2 verdict, with the losing argument attached, not laundered into false unanimity.
Before any of that starts, the debate is grounded. Every voice argues from the same evidence pack — your actual numbers, each with an ID — and claims have to cite it. This isn’t five chatbots trading vibes. It’s five positions arguing over the same facts, which is the only kind of argument that converges on anything.
A 4–2 verdict is reported as a 4–2 verdict, with the losing argument attached — not laundered into false unanimity.
Delay is not free
Here’s the moment that convinced us the devil’s advocate earns its keep. In a pricing debate — should this company raise a product’s price 15%? — the majority position was the reasonable-sounding one: wait, gather more data, revisit next quarter. Prudent. Safe. The kind of answer you’d get from one model, asked once.
Then the devil’s advocate attacked the consensus, and its argument was simple: delay is not free. At this company’s margin and burn, every month of “waiting for more data” had a price tag the majority had quietly booked at zero. Choosing not to decide was itself a decision — an expensive one — and nobody had costed it. The final verdict still carried conditions, but it carried them honestly: a split vote, the dissent on the record, and the hidden cost of waiting made explicit instead of invisible.
That’s the pattern single-model advice misses constantly. The obvious risk gets analyzed; the risk of the safe option doesn’t, because nobody at the table was assigned to it.
A verdict you can audit
When the debate ends, you don’t get a vibe — you get an artifact. The question, the positions, the attack, the vote, and a verdict whose claims cite the evidence IDs they rest on. You can open it, disagree with it, and see exactly which premise you’re disagreeing with. If the voices genuinely deadlock, Nerve reports “no majority” instead of manufacturing a winner, because a tie is information and a fake consensus is a lie.
And because arguing is expensive — in tokens and in your attention — the debate isn’t how Nerve answers everything. Routine questions get computed answers with citations, in seconds. The full argument is reserved for the calls that deserve it: the ones where being confidently wrong costs real money.
Disagreement as a feature
The deepest thing we changed wasn’t technical. It was the definition of a good answer. A good answer to a hard call isn’t the most confident one — it’s the one that survived the strongest attack available, with the survivors and casualties both on the record. That’s what we’d want from a human advisory board. There was no reason to accept less from the machines.