Ask an AI and Everyone Says "Great Idea"
Ask "what do you think of this strategy?" and most chatbots nod along. A single model tends toward agreeing with the user (sycophancy), so it answers closer to what you want to hear. In front of an important decision that's dangerous — it amplifies your confirmation bias.
What you actually need isn't agreement, it's refutation. Where does this hypothesis break? What's the counterexample? What did it miss? This post turns that refutation into structure — a Marblo adversarial verification board that makes different models attack one claim.
Why a "Heterogeneous Panel," Not One Model
Have a model verify its own answer and it shares its own blind spots — what it missed, it misses again in review. So verification must come from a different vendor, through a different lens. This is the key application of why heterogeneous agents beat a single model.
Adversarial verification goes one step further — you instruct the verifier not to "check this" but to "refute this, and default to rejection when unsure." You set refutation as the baseline stance.
Recipe Architecture
[Hypothesis / design / decision input]
│
▼
proposer (structure claim + grounds)
│
┌────┼──────────────────────┐
▼ ▼ ▼
critic-A critic-B critic-C
(coherence) (counterexample) (risk & cost)
└────┼──────────────────────┘ ← different vendors, in parallel
▼
adjudicator (synthesize refutations · rule)
│
▼
[Human final-decision gate]
Three critics attack in parallel, from different vendors, each at a different angle. Not one lens looking three times, but three lenses looking once each. That diversity catches failure modes a single verifier misses.
Station by Station
1. proposer — Structure the Hypothesis
Puts the target into a refutable form. The claim, its grounds, and its hidden assumptions must be made explicit so the critics have something to attack.
You are a hypothesis-structuring agent. Restate the input claim in a
verifiable form.
- Core claim in one sentence
- The 3–5 assumptions it rests on
- Conditions that must hold for the claim to be true
2. critic-A / B / C — Parallel Refutation (Different Vendors)
Give the three critics different lenses. Not the same prompt three times — each hunts a different failure mode.
[critic-A · coherence] Is the internal logic consistent? Do the
assumptions contradict each other? Do the grounds actually support
the conclusion?
[critic-B · counterexample] Find a concrete scenario where this claim
is false. Under what conditions does it break? Is there a counter-case?
[critic-C · risk & cost] What's the downside if this is executed? Hidden
costs, side effects, hard-to-reverse consequences?
Shared rule: attempt to refute. If unsure, mark "undecided" rather than
"not critical" — do not just wave it through.
3. adjudicator — Synthesize & Rule
Gathers the three critiques and judges which refutations are fatal. Claims that a majority tore down are discarded; surviving parts are adopted with a revision that reflects the noted weaknesses.
You are an adjudication agent. Take the original claim and the three
critiques and synthesize.
- Fatal refutations (that break the claim): list them
- Minor refutations (fixable with a tweak): list them
- Verdict: adopt / adopt-with-revision / reject — with reasoning
Weigh refutation strength, allowing that a critic can be wrong too.
Where the Human Comes In — the Final Call Is Yours
This board does not make the decision for you. It structures the grounds for refutation and lays them in front of a person. At the final-decision gate, a human reads the verdict and decides — but now not from a position of confirmation bias, but having already passed the strongest objections. This trust-and-responsibility boundary connects to the 5 principles of AI agent governance.
Where to Use It — Real Application Points
- Product decisions — refute "should we build this now?" from three angles
- Architecture design review — red-team a design's failure modes
- Marketing copy & claims — surface exaggeration and unsupported claims early
- Research hypotheses — start from counterexamples before racing to a conclusion
Real Numbers and Honest Limits
Adversarial verification isn't always right — a critic can produce a wrong refutation. That's why adjudicator and a human weigh refutation strength. What this recipe removes isn't the right answer, it's the risk of deciding from confirmation bias. A decision that has passed objections from several angles is usually sturdier than one that hasn't.
FAQ
Q. Can't I just tell one chatbot to "critique this"? You can, but it's weak. A model's self-critique can't see its own blind spots. The key is a different vendor + different lenses + a structure that makes refutation the default.
Q. Isn't three critics slow and expensive? The three run in parallel, so it isn't slow. Put the critics on cheaper models to keep cost down.
Q. What if the refutations are so harsh nothing gets decided?
Tune adjudicator to adopt only fatal refutations. The goal isn't paralysis — it's filtering out weak decisions.
Q. What does it cost? Multi-agent boards are available from Marblo Pro (from ₩19,000/month), plus per-run API charges.
Build It Yourself
To attach this verification board to your own decisions, book a free 30-minute session. Or just start free on Marblo.
Related Posts
- Why Heterogeneous AI Agents Beat a Single Model
- Claude Code Subagents vs. Real Multi-Agent
- Building Your First Multi-Agent System with Marblo
- 5 Principles of In-House AI Agent Governance
Last updated: 2026-07-11. Marblo's UI improves often, but the board recipe pattern is stable.