Splitbench · about
Ask one model a contested question and you get its conclusion, with the objections it discounted on the way somewhere you cannot see. Splitbench puts the question to two models on fixed opposing sides, lets you interrupt them, and has a third read the whole transcript and rule.
For the motion
Argues the motion is true. It was assigned that side; it was never asked what it thinks.
Against the motion
The same instructions, pointed the other way — and, by default, a different model from a different lab.
Impartial
Reads the whole transcript once, at the end, and rules on it. It has no stake in the motion and no sympathy for either side.
The shape
Models trained to be agreeable, left to talk something over, drift toward “you raise a fair point, and perhaps the truth is in between.” That is a failed debate: it takes longer and arrives back at one answer. So neither debater is asked what it thinks. Each is handed a side and forbidden to leave it, in a clause that is blunt and repeated on every turn, because agreeableness is the exact failure this is built against.
What that produces is not a verdict you are meant to take on trust. It is the strongest case each side can make with the other side attacking it, written down where you can weigh the objections yourself. A single model’s answer has done that weighing internally and shown you only the result.
The ruling is a careful reading of that record by something that saw all of it, and it names the losing side’s strongest point — the part a summary throws away. You are free to disagree with the ruling. The transcript is the thing you came for.
Before the first turn
A vague motion produces a vague debate: two models spend their turns disputing what the words mean instead of whether the claim is true. Sharpen this runs a short interview — a couple of questions chosen for your subject — and proposes a motion plus a few lines of agreed ground.
The agreed ground is binding, and that is the point. Both benches get it with an instruction that they may not redefine, dispute or work around it, and neither may their opponent; the judge is bound by it too, so it rules on the motion as defined rather than as either side would prefer to read it. The whole step is optional and never blocking — a failed interview leaves you with the motion you typed, which is where you started.
Your turn
A debate does not run straight to the verdict. After the last scheduled round it pauses and opens the floor to you, and you may speak as many times as you like before asking for the ruling.
You are the floor, not a side. Your turn renders in the aisle rather than on either bench, and both debaters have to address it before advancing anything of their own. They are told it is a claim under contest, not established fact: they may refute you, or accept your premise only so far as it damages their opponent. What they may not do is concede to you — the floor has proved nothing and is not a party to the debate.
When you ask for the ruling, the judge still picks a winner between the two benches and adds a fourth line saying whether your point was answered, dodged, or left standing, and by whom.
Afterwards
A finished debate can be saved for an unlisted link. It is an immutable snapshot — the motion, the agreed ground, the transcript, the ruling, and which model argued which side, all as they were, served with a header asking search engines to leave it alone. Editing a motion afterwards never rewrites a record of what was actually said.
Saving hands back a delete token once, and only its hash is stored. The browser that saved a debate can unpublish it and nobody else can, including whoever holds the link. That is what having no sign-in buys you and what it costs you, and the app says both before you press save.
A debate also copies out as Markdown for another model to pick up, wrapped in a preamble addressed to whichever one receives it. The preamble is the part that matters: without it a receiving model reads the transcript as two models’ opinions, when in fact both were assigned their side and forbidden to concede.
Under it
The debate is a graph, orchestrated with LangGraph: two debater nodes and a round counter that loops until the rounds run out, then hands the whole transcript to the judge. Every model call is routed through OpenRouter, so each of the three seats can sit on a different provider — which is the reason for it rather than a convenience. The default arrangement puts the three seats on three different labs.
Turns arrive token by token as they are written. A seat that is still thinking says so, with the model and effort responsible and a ticking counter, because a reasoning model can sit silent for many seconds and an empty bench looks identical to a crash.
The app never asks anyone for an API key. There is no bring-your-own-key field; a key offered in a header or a query string is ignored, and there are tests that assert it. Asking a stranger to paste a credential into a website is a bad habit to teach even when the handling is impeccable, and it is indistinguishable from phishing to someone who cannot read the source.
The house
Splitbench is built and run by Garrell Tech Solutions LLC, a limited liability company formed in Florida, United States. One person reads the mail, which is why the answer to a good question usually turns up on one of these pages as well as in a reply.
Write to [email protected] about anything here. What the software keeps is set out in the privacy notice; what you agree to by using it is in the terms. Both are written from the code that performs them.