Prepared remarks are the easiest part of a debate to polish. A side can rehearse an opening, smooth the wording, and sound certain for several minutes without ever touching the hardest part of its own case. The format changes the moment a direct question arrives, one about evidence, about costs, or about who is actually affected if the proposal goes ahead.
Building a structured debate platform kept returning to that moment. The question round was not decoration added after the speeches. It became the test that decided whether an argument was built on reasons or only on delivery, and designing for it changed decisions about motion wording, turn order, and how a verdict should be explained.
Designing for questions, not just speeches
The first choice was precision before anyone speaks. A vaguely worded motion produces vague questions, because nobody can pin down what is being claimed. Terms have to be defined early enough that a question can point at a specific part of the case instead of drifting into a general disagreement about values.
The second choice was listening as a format rule. Many exchanges stall because each side answers a version of the other side that nobody actually holds. The build pushed for questions that quote the claim being challenged, in short form, before asking anything. It slows the exchange by a few seconds and saves minutes of talking past each other.
The third choice was separating evidence from confidence in the scoring view. A claim delivered firmly is not automatically a claim supported well. The interface work therefore kept the source and the reasoning visible next to the verdict, so a reader could see whether a strong score came from evidence or from a confident tone that was never tested.
What worked
Short, specific questions did the most work. A question that names one cost, one dataset, or one affected group forces a concrete answer. Broad questions invited broad speeches in reply, and the exchange learned nothing. When the question format nudged people toward one point at a time, the answers got shorter and more useful.
Steelmanning before challenging helped more than expected. Asking a side to restate the opposing point fairly, before questioning it, reduced strawman replies and made the follow up question sharper. Readers could also follow the thread without needing a summary written by either side.
Showing the working stayed valuable. When the verdict explained which answers held up under questioning and which ones shifted or conceded a point, readers reported trusting the outcome more, even when they disagreed with it. The reasons being visible mattered more than the final score.
What broke, or nearly did
Questions kept turning into speeches. The earliest version allowed long question turns, and participants used them to deliver a second opening statement with a question mark at the end. The fix was structural: limit the question to the claim it targets and the one thing it asks, and move any new argument to the speakers own turn.
Vague motions broke the question round first. If the motion could mean three different policies, every question could be dodged by switching meanings. That failure showed up in questioning before it showed up in scoring, which made the question round a useful early warning that the motion needed rewriting.
Scoring confidence was the quieter trap. Fluent answers scored well in early tests even when they avoided the question. The scoring prompts had to be adjusted to reward direct answers and penalise evasion, and even then the honest limit remains: a score is a summary, not proof. Publishing the reasoning alongside it is what keeps the score honest.
Lessons for builders
Design the challenge, not just the presentation. Any system that lets people state a case should also make that case answer a fair question in public. The challenge is where weak reasoning becomes visible and strong reasoning becomes easier to trust.
Make the question cheaper than the dodge. If answering directly takes more effort than evading, people will evade. Short questions, clear terms, and one point at a time lower the cost of a straight answer.
Leave readers with a method, not just a result. A debate that ends with minds unchanged can still be useful if everyone leaves able to judge the next argument better. That durability is the point of showing the questions, the answers, and the working together.
The project behind these lessons is AI Senatus, a structured debate platform where motions are argued from opposing sides, questioned in public, and judged with the reasoning shown.
