Back to home
ACE Engine v2.37 AI Models3-Round Debate

Methodology

The ACE Engine (Adversarial Consensus Engine) is our multi-agent reasoning framework. It forces AI models to argue against each other instead of agreeing with you — producing decisions you can actually trust.

1. The Problem with Single-AI Advice

When you ask one AI model a question, it gives you one perspective. Worse, most models are trained to be agreeable — they tell you what you want to hear. This is called sycophancy bias, and it's the single biggest problem with AI-assisted decision making.

The ACE Engine solves this by forcing 6 different AI models — each with a different role and perspective — to argue against each other in 3 structured rounds. A 7th model acts as an impartial judge.

The result: Instead of one biased opinion, you get a stress-tested verdict where weak arguments have been attacked, defended, and scored by an independent arbiter.

2. The 4 Phases of ACE

Every session follows the same structured protocol — no shortcuts, no skipped steps.

🎯

PHASE 1: Focus Lock

Round 1 — Independent Analysis

Each AI expert is given a specific role (Visionary, Devil's Advocate, etc.) and analyzes your question independently. They don't see each other's responses. This prevents groupthink and ensures genuine diversity of perspective.

⚔️

PHASE 2: Adversarial Challenge

Round 2 — Cross-Examination

Now each expert reads what all the others said in Round 1 — and attacks the weakest arguments. The Devil's Advocate targets the Visionary's optimism. The Financier challenges the Technologist's cost estimates. Weak reasoning gets exposed and dismantled.

🛡️

PHASE 3: Defense & Concession

Round 3 — Final Arguments

In the final round, experts defend their surviving arguments and concede where they were wrong. This is critical — it produces intellectual honesty, not stubbornness. Each expert gives a final score (0-100) for the original idea.

🏛️

PHASE 4: Judicial Arbitration

The Judge — GPT-4o

GPT-4o acts as an impartial judge. It reads all 3 rounds of debate, weighs every argument, penalizes attacks that were never refuted, and delivers a structured verdict with a consensus score (0-100), confidence level, and concrete recommendations.

3. The 6 Expert Roles

Each expert has a fixed perspective and a natural bias. The biases are intentional — they create productive conflict.

🔮

The Visionary

DeepSeek R1
Thinks like: A VC evaluating a Series A
Natural bias: Optimistic — sees the 10x opportunity
Central question: "What could this become at scale?"
Approach:Identifies market creation potential, network effects, and long-term competitive advantages.
⚠️
Known blind spot: Can underestimate execution difficulty and operational costs.
⚙️

The Technologist

Kimi K2
Thinks like: A CTO who has built 3 startups
Natural bias: Pragmatic — focused on what can actually be built
Central question: "Is this technically feasible and scalable?"
Approach:Evaluates architecture, build time, tech debt, and failure points at 10x scale.
⚠️
Known blind spot: Can over-optimize for technical elegance at the expense of business value.
😈

The Devil's Advocate

Llama 4 Maverick
Thinks like: A short-seller researching why this will fail
Natural bias: Skeptical — finds every flaw and failure mode
Central question: "What is the most likely cause of failure?"
Approach:Pre-mortem analysis. Identifies single points of failure, market risks, and hidden assumptions.
⚠️
Known blind spot: Can miss genuine breakthroughs by focusing only on tail risks.
📊

The Market Analyst

google/gemini-2.5-flash-preview-05-20
Thinks like: A McKinsey partner
Natural bias: Data-driven — needs numbers to believe
Central question: "What do the market data say?"
Approach:TAM/SAM/SOM analysis, competitive landscape mapping, go-to-market strategy evaluation.
⚠️
Known blind spot: Can over-rely on historical data and miss category-creating opportunities.
💰

The Financial Strategist

Qwen 3 (235B)
Thinks like: A CFO with private equity experience
Natural bias: Conservative — focused on unit economics and survival
Central question: "Does the math work?"
Approach:Unit economics, burn rate analysis, revenue model validation, path to profitability.
⚠️
Known blind spot: Can kill growth opportunities by demanding premature profitability.
⚖️

Ethics & Risk

Mistral Medium 3
Thinks like: A compliance officer at a top law firm
Natural bias: Cautious — focused on what could go wrong legally and reputationally
Central question: "What are the regulatory and ethical risks?"
Approach:GDPR compliance, bias assessment, privacy risks, reputation impact, regulatory exposure.
⚠️
Known blind spot: Can slow down innovation with excessive caution.
🏛️

The Judge — GPT-4o

Impartial Arbiter

The Judge never participates in Rounds 1-3. It only reads the complete debate transcript and delivers the final verdict. It weighs arguments by strength, penalizes unrefuted attacks, and produces a consensus score with concrete, actionable recommendations.

4. Model Diversity & Assignment

Each expert runs on a different AI model from a different provider. This prevents the "monoculture problem" — if all experts used GPT-4o, they'd share the same training biases and blind spots. By using 7 different models, we get genuinely independent perspectives.

ExpertSite DisplayOpenRouter IDProvider
🔮 VisionaryDeepSeek R1deepseek/deepseek-r1DeepSeek
⚙️ TechKimi K2moonshotai/kimi-k2Moonshot AI
😈 DevilLlama 4 Maverickmeta-llama/llama-4-maverickMeta
📊 MarketGemini 2.5 Flashgoogle/gemini-2.5-flash-preview-05-20Google
💰 FinanceQwen 3 (235B)qwen/qwen3-235b-a22bAlibaba
⚖️ EthicsMistral Medium 3mistralai/mistral-medium-3Mistral AI
🏛️ JudgeGPT-4oopenai/gpt-4oOpenAI

5. The Conflict Matrix

The most valuable insights come from structured disagreement. These are the built-in tensions that drive the debate:

🔮 Visionary💰 FinancierGrowth vs Survival

The Visionary pushes for scale. The Financier demands unit economics. The truth is usually the sustainable growth path between them.

⚙️ Technologist😈 Devil's AdvocateBuild vs Break

The Technologist designs the system. The Devil tries to break it. The result is the most resilient version of the architecture.

📊 Market Analyst⚖️ Ethics & RiskOpportunity vs Compliance

The Market Analyst sees the opportunity. Ethics & Risk flags the regulatory exposure. The balance is a market entry that won't get you sued.

Why this matters: When two experts with opposing perspectives reach agreement on a point, that point is extremely likely to be valid. When they can't agree, the Judge weighs the strength of each side's arguments.

6. Output: The Verdict

Every session produces a structured verdict with these components:

Consensus Score (0-100)

How much agreement exists across all 6 experts after 3 rounds of debate.

Decision (GO / NO-GO / CONDITIONAL)

The Judge's final recommendation based on the weight of arguments.

Confidence Level (LOW / MEDIUM / HIGH)

How much the experts agreed. Low confidence means the debate was highly contested.

Key Arguments (For & Against)

The strongest surviving arguments from both sides, after adversarial testing.

Actionable Recommendations

Concrete next steps — what to do, what to investigate, what to avoid.

7. Debate Modes

The ACE Engine supports multiple debate formats optimized for different decision types.

Default

🏛️ Classic Council

All 6 experts debate freely across 3 rounds. Best for business decisions, strategy questions, and startup validation. Covers 80% of use cases.

⚔️ Tribunal Mode

Splits the council into Prosecution vs Defense. Three experts argue FOR, three argue AGAINST. Best for high-stakes binary decisions.

Coming soon — Pro plan

8. Limitations & Honest Constraints

We believe in transparency about what the ACE Engine can and cannot do:

  • AI can hallucinate. All 7 models can generate plausible-sounding but incorrect information. The adversarial format reduces this (other experts catch errors), but doesn't eliminate it. Always verify critical facts independently.
  • Not instant. A full 3-round debate with 7 models takes 30-90 seconds. Quality reasoning takes time.
  • Not a replacement for human judgment. We provide structured analysis from multiple perspectives. You make the final decision. AI is an advisor, not a decision-maker.
  • Training data cutoffs. Each model has a knowledge cutoff date. For very recent events, supplement with your own research.
  • Domain expertise. The council works best for business, technology, and strategy decisions. For medical, legal, or financial advice, consult qualified professionals.

See it in action

Submit an idea and watch 6 AI experts debate it through 3 adversarial rounds — in real-time.

Start a free session