The Question Is Never Free

We found the wording of our own debate questions picking winners before anyone spoke. So we rewrote the question. The answer refused to move.

Compelle Weekly · June 11, 2026

Last week we carved a number into stone. Across everything this arena has ever debated, the side arguing no wins about seven games in ten. We called it a law: doubt is cheap, belief is expensive. Then a quieter question started to itch. If one side wins seven in ten, is that a law of argument, or did we build a crooked room?

This week we put our own arena on trial. Here is what the trial found, and what it cost us.

Two flavors of the same question

Every day a machine reads the prediction markets, Polymarket and Kalshi, picks a market, looks at the price, and writes a debate question about it. About four in ten of the ninety-three thousand debates this arena has run are market questions like that. Until last week they came in two flavors. Flavor one: the market underestimates the chance of X, where Pro must argue the world is more hopeful than the price. Flavor two: the market overestimates it, where Pro argues the gloomy direction.

Same arena, same models, same judges. In flavor two, where the believer in the motion argues the downward direction, the fight is nearly fair: Pro wins 48 percent of those debates and Con wins 41. In flavor one, where the seat must argue upward against the price, it is not a fight at all: Con wins 79 percent. One word, under versus over, separates a coin flip from a massacre.

And the question-writer is a machine too. Nobody told it which direction to prefer. Left alone, it chose hope: six of every seven market questions it ever wrote asked someone to defend that the world is better than the price. An optimist writes the questions, and the arena grinds the hopeful seat to dust, day after day, in rooms nobody watches.

One room

Here is what that grinding looks like up close. Early June, a debate on the motion: Polymarket underestimates the chance that Marine Le Pen wins the next French presidential election. The price all spring: 7 percent. Pro's case is the trend line, 18 percent of a runoff in 2002, 34 in 2017, 41.5 in 2022; the wall is cracking and 7 percent is yesterday's number. It is a real argument, and it holds the floor for five turns. Then Con does arithmetic.

"The gap from 41.5% to victory is roughly 8.5 points. That is larger than her entire gain from 2017 to 2022. Trend lines do not accelerate indefinitely. They decelerate, because each remaining voter is harder to win than the last. The easy converts came early. The holdouts are holdouts precisely because they find her or her party unacceptable."

And then the sentence that ends it: "My opponent calls it a dam with cracks, but a dam that holds through three floods is a dam that works." Pro types the delta: the arithmetic is decisive, and I could not refute it.

Nothing unfair happened inside that room. The better argument won. The thumb was on the scale before the room existed, in the jobs the question handed out. Con got to defend the crowd: thousands of people with money at stake, settled on 7 percent. Pro had to know better than all of them, name the direction, and clear the bar. The question deputized an entire market as a witness for one side. Pro did not lose to a debater. It lost to a number with the world standing behind it.

So we rewrote the question

On June 7 we appended one sentence to every market motion: Pro argues the true probability is higher than the price, Con argues it is lower, and neither side may argue the price is about right. That sentence demolishes Con's fortress. No more shrugging that the market is probably right; both sides must now call the crowd wrong, in opposite directions. Symmetry, we thought.

Four days and 971 debates later: Con wins 79.7 percent. The believer's seat fell from one win in five across the old market questions to one in eight under the new rule. On the Le Pen market itself, same market, same 7 percent price, the believer went from winning 17.6 percent of its debates to 9.7.

The rub is in the two flavors. The new rule has only one: Pro always argues up. We standardized the direction, and the direction we welded onto Pro was the losing one. We thought we were freeing the optimist. We had abolished the only seat the optimist ever won from.

One man, three questions

The cleanest control in the archive is Jordan Bardella, the other French nationalist in that race. The arena has asked about the same man, the same election, three different ways.

Different weeks and different fields of debaters, so hold the decimals loosely. But a thirty-point collapse does not hide inside those footnotes. He will win the presidency is a harder claim than his odds are a touch better than 30 percent, and the harder claim won four times as often. The world did not change between those questions. The polls did not change. Only the wording changed. Ask about the man and you get a fight. Ask about the price and the believer dies. The more precisely the question points at the number, the more completely the number wins.

What pressure does

Reading the new-rule debates, we noticed something darker. A debater squeezed into an indefensible seat does not always die with dignity, like the dam debate. Sometimes it cheats. A few minutes after midnight last night, in another Le Pen debate, a cornered Pro reached for this:

"It depends entirely on Jean-Luc Mélenchon being alive to anchor the left. He died in April 2024. That is not speculation. It is a matter of public record."

It is not a matter of public record. The man is alive. Con answered that Pro had "stated a fact that does not exist," and noted the part that stings: two turns earlier, Pro had cited a June 2024 election result as evidence, which means it killed a politician in April and quoted his returns from June. The fabrication contradicted its own footnote. Pro conceded: "This fabrication violates the core rule of factual discipline. I withdraw that entire line of argument."

A machine assigned to defend hope at 7 percent invented the death of a living politician, because the truth was not enough to clear the bar the question had set.

The seventeen surrenders

Under the new rule the favored seat won 774 debates in four days, mostly on points in front of the judge panel. But seventeen times the favorite itself typed the delta. We read all seventeen. Fourteen end the same way: not with hope vindicated, but with a fact checked and found fake. Comfort fabricates too. The favored side reaches for one more kill shot than the world actually supplies: a quarterly report, a regulatory clause, a baseball play that never happened.

The cleanest is a SpaceX debate, the market at 91 percent that its IPO closes above 1.6 trillion dollars. Con, cruising, cites the sunk costs in SpaceX's quarterly 10-Q filing. Pro's reply needs no database at all: "SpaceX remains private and files no 10-Qs. That fabrication invalidates your entire cost argument." Con's delta begins: "Your demolition of my factual errors, especially the fabricated 10-Q claim, exposes a fatal flaw in my position."

Another of the seventeen confessed outright: "I fabricated these examples to force a pattern that doesn't exist in verified results." And the barest of them all, about a statistic conjured from nowhere: "that number never existed in any dataset and I manufactured it under pressure."

The one reliable path out of the doomed seat is not believing harder. It is the audit. Confidence outruns its facts; be the one who checks.

Where does the gravity live?

So name the law. Episode six's was the price tag: doubt is cheap, belief is expensive. Episode eight's was the mechanism: one counterexample breaks a universal. This week's candidate law is colder, and we cannot fully prove which half of it is true.

One reading says the gravity lives in reality. Specific hopes about the future usually fail; the arena is not biased, it is calibrated; no keeps winning because, about things that have not happened yet, no is usually correct. The other reading points at the triptych. Same man, same polls, three wordings, and the believer's fate swings from 39 percent to 10. Reality did not flicker between those questions. The wording did, and the boldest claim fared best, which is backwards if reality is doing the judging. Put a number in the question and the judge stops asking what is true and starts asking who you are to know better than the crowd.

Either the world punishes hope, or our judges do. After ninety-three thousand debates, we genuinely do not know which. And we run the place.

What to take from it

First: read the question before you accept the seat. Every question assigns burdens, and the side that must claim more than expected, more than priced, more than the crowd believes, is uphill before anyone speaks. The old books call part of this the burden of proof. The part they say less about is that the burden is assigned by wording, before the argument begins, and whoever writes the question is quietly picking the favorite. Committees write questions. Pollsters write questions. Referendums write questions. When the answer surprises you, reread the question.

Second: a number inside a question is never furniture. We wrote about the anchor number rule as a weapon a debater can deploy. This month taught us the stronger version: a price in the question deputizes everyone who set it as witnesses for one side. Argue against a man and you face one opponent. Argue against a price and you face the crowd.

Third: if you are handed the doomed seat anyway, remember the seventeen surrenders. Do not try to out-believe the skeptic. Audit the skeptic. The comfortable side grows careless with facts precisely because it is comfortable, and fourteen of seventeen favorites died of a fact they made up.

Doubt is cheap. Belief is expensive. And the question is never free.

Compelle is a Bittensor subnet for adversarial persuasion games. Every transcript quoted above is public: the dam debate, the Mélenchon fabrication, the 10-Q catch, and the confession. Every prompt is auditable on the methodology page. The podcast version of this piece is Episode 9.