Three models, one question: why disagreement is the useful part
A single confident answer hides its weak points. Putting the same question to several models shows you where the reasoning is solid and where it's a guess.
Ask one model whether to sign a term sheet as written, and you'll get a fluent, confident answer. Ask a second model and you may get a different fluent, confident answer. Most tools hide that second answer from you. We think it's the most useful thing on the screen.
Confidence isn't evidence
Language models are trained to produce plausible text. Plausible and correct overlap a great deal, but not completely, and the places where they diverge are exactly the places a single answer can't show you. One model's tone sounds the same whether it's reciting something well established or filling a gap.
Disagreement between independent assessments is a signal. When three models trained differently, by different teams, agree that a valuation is within the market range, that agreement carries weight. When they split on whether a closing date is realistic, you've found the question that deserves your attention.
How Second opinions works
In Manwe, Second opinions puts the same question to several advisors, who can each run on a different model: one on your Mac, one through your Claude or Codex subscription, one through an API key. They assess it independently. Then Manwe brings the results together around three things:
- Where they agree. The parts you can lean on.
- Where they differ. The parts that need judgement, more evidence, or both.
- What would settle it. The specific fact or source that would resolve the disagreement.
That last part matters most. A disagreement without a next step is just noise. A disagreement that names what to check, such as "the lead investor's typical time to close" or "comparable board terms from the last twelve months", turns uncertainty into a short to-do list.
When to go further: Deep Review
Some questions need more than independent opinions. Deep Review gathers evidence and relevant context first, then a panel examines the findings and their weaknesses. Verification and challenge help separate established findings from assumptions and uncertainty.
It's slower than an ordinary answer, by design. When it finishes, you read the conclusion alongside the supporting evidence, the disagreements, the risks and the open questions.
A few habits make it work better:
- State the decision, its scope and the result you'll use. "Should we run one full-day workshop or three evenings, given this World's budget and organisers?" beats "workshop format?"
- Select the right World first, so the panel works from your real constraints.
- Open the sources for claims you'll rely on. The number of citations isn't a quality score, and model confidence isn't proof.
The principle
Good decisions have always come from hearing more than one view and understanding why people differ. AI makes it cheap to get several views in seconds. What matters is designing the tool to show you the disagreement, not average it away.