Quorum puts several frontier models on the same question, has them read and challenge each other, then returns one answer. That is a real cost in money and in waiting. It is worth paying on some problems and wasteful on most. This page is the honest version of which is which.
There is no category of subject matter where deliberation wins. There is a shape of question, and it shows up in software, in law, in hiring, in operations, in medicine, in nothing in particular. Three tests. A question needs all three, not one.
| Test | Passes when | Fails when |
|---|---|---|
| Genuine disagreement is possible | Two competent people could read the same facts and land in different places | There is one correct answer and the only question is whether the model knows it |
| Being confidently wrong is expensive | A plausible, fluent, incorrect answer gets acted on before anyone checks | An error is obvious on sight, or costs a few seconds to correct |
| The answer is judgement, not retrieval | It turns on weighing trade-offs, reading context, choosing between goods | It turns on recalling a fact, applying a rule, or reformatting an input |
The mechanism follows from the first test. A panel only adds something when the models actually diverge; if they all say the same thing, you have paid several times over for one opinion. Divergence is what the critique round and the synthesis pass are there to exploit.
Two grounding measurements from Quorum’s own traffic, so this is not purely assertion. Across 3,947 classified questions, difficulty scored 45% under 0.35, 33% between 0.35 and 0.65, and 22% above 0.65, with a mean of 0.456. And across 3,845 deliberations, a critique step fired on roughly three runs in four (2,678 discussions in the third seat, 2,565 in the second, 1,495 in the first).
Read those two together and the honest summary is: most questions people bring do not need this, and on the ones that reach a panel, there is usually something for the panel to disagree about. Both halves matter.
Deliberation inside a process. Not a person opening a chat window, but a decision that gets made again and again, by whoever is on shift, where the cost of the decision drifting is higher than the cost of making it.
Approvals. Reviews. Triage and routing. Escalation calls. Anywhere a judgement is repeated at volume and consistency is part of the output, not just accuracy on any single case.
Three properties make this the strongest fit we know of.
Two practical notes. First, do not send the whole queue. Classify each item and escalate only what earns it; the pattern is written up at the escalation router. Second, every deliberation returns a receipt showing which engines ran and how much they disagreed, which is what makes a workflow auditable after the fact rather than merely automated. See the API reference.
We have not run a controlled study of Quorum inside somebody else’s production workflow. The reasoning above is sound and the mechanism is real, but if you want proof for your process specifically, the only way to get it is to measure your own before-and-after. We would rather say that than imply evidence we do not have.
Deliberation supporting a person who is making the call, rather than making it for them. The person stays the decision-maker. The panel is there to widen what they are looking at before they choose.
This is a different job from the workflow case, and it wants a different output. In a workflow you generally want one settled answer. Here, the disagreement is often the useful part. If three strong models split on a question, that split is a fact about the question, and a single confident answer would have hidden it from you.
Where that helps:
The honest failure mode of this category is that it can become expensive reassurance. If you already know what you are going to do, a panel will usually find a way to agree with you, and you will have paid to feel better. That is covered below.
On the consumer surfaces, this is why Panel exists as a distinct thing from Chat: it shows you the individual positions rather than only the resolved answer. Help covers what each surface does.
This section is not a disclaimer. Most questions belong here, including most of the ones sent to Quorum, and knowing that is more useful to you than another paragraph of advocacy.
Start with the arithmetic, because it settles several cases on its own. Classifying a question is not the same operation as answering it, and the gap is very large.
| Operation | Time | Cost |
|---|---|---|
| Free pre-scoring classification | ~0.9 s | ~$0.00006 |
| Full deliberation, median (n=3,845) | 28.5 s | panel pricing, see Pricing |
| Full deliberation, p90 | 88.2 s | |
| Full deliberation, p99 | 229.3 s |
Roughly three times the wait of a single frontier model at the median — our measured baseline is 9.0 s against 28.5 — and close to four minutes in the worst percentile. Now the cases.
One structural caveat that applies everywhere: deliberation reduces the chance that a single model’s blind spot survives to the answer. It does not eliminate it. Models trained on overlapping data share some errors, and a panel can be wrong together, fluently and unanimously. Where being wrong is genuinely dangerous, a panel is a better input to a human check, not a replacement for one.
A heuristic you can apply without running anything.
Imagine handing the question to two people who both genuinely know the field. Would they answer the same way? If yes, one model is enough; you are retrieving, not deliberating. If they would argue, and if you would want to hear the argument, that is the shape.
Picture the answer coming back confident, well written and wrong. What happens next? If somebody spots it in seconds, deliberation is not buying much. If it gets acted on, quoted, shipped or filed, the panel is paying for itself in the cases where it catches the divergence.
Ask what else is happening in the thirty seconds. If the answer is “a person is watching a cursor blink”, this belongs on a single model. If the answer is “the item is in a queue and nobody is waiting”, the wait is free.
All three tests are judgements about your own traffic, and judgements about your own traffic are frequently wrong. Classification is free and takes about 0.9 seconds, so run a representative sample of your real questions through it and look at the distribution before you decide anything. Compare against ours above if it helps. If your shape looks nothing like ours, that is information about your product, not a mistake.
/docs/concepts/where-deliberation-thrives
The strongest evidence we have that deliberation beats a single model is our own published efficacy figure: a 63.9% win rate against a single model, n=577, ±3.9 at 95% confidence, across four batches. Method and raw runs are at Testing.
Be careful how much weight that carries. It is an aggregate over a mixed set of questions. It does not say deliberation wins 63.9% of the time on your workflow, or on legal review, or on triage, or on any one problem type. It says that over that sample, more often than not, the panel’s answer was preferred. A win rate near 64% also means roughly a third of the time the single model was judged at least as good, which is consistent with everything on this page: a large share of questions do not need a panel.
Per-domain and per-mode breakdowns are the obvious next thing to measure and we have not published them. When we have them, they will go on Testing whether or not they flatter us.
The escalation router for routing only the questions that earn a panel · Pricing, explained for what a call costs · Testing for how the efficacy numbers are produced · Help for the consumer surfaces · API reference for the endpoint contracts.