How it works
Redteam AI runs your submission through a four-stage adversarial loop. Four different AI models — each from a different US-owned company — evaluate your decision from opposing perspectives. No model runs two consecutive stages. The Prosecutor and Refiner are randomly selected on each run to prevent epistemic lock-in.
The four-stage loop
Gemini reads your submission with live Google Search grounding — surfacing current market conditions, recent competitive moves, and public information about the entities you mention before the adversarial loop begins. It determines whether the submission is specific enough for meaningful analysis. If a critical field is too vague to produce specific output, it asks one clarifying question before the loop continues. This stage is fixed — Gemini runs it every time. Web search is not run on code review submissions.
Claude constructs the strongest honest version of your idea as it currently stands. It does not fix weaknesses — it presents the best case. It identifies the real customer, the mechanism of value, and the three critical assumptions the entire idea depends on. This stage is fixed — Claude runs it every time.
A randomly selected model — GPT-5.4 or Grok 4 — attacks the idea at its load-bearing points. Each run the Prosecutor is assigned one of three attack personas: The Assassin (targets the single assumption the idea cannot survive being wrong about), The Skeptic (demands evidence for every claim and exposes where none exists), or The Contrarian (finds the counterintuitive angle nobody has said out loud). Every risk must reference specific mechanisms, precedents, or real-world behavior. Generic risks are not permitted.
A different randomly selected model — Gemini 2.5 Pro or GPT-5.4 — synthesizes the Advocate's case and the Prosecutor's risks into a meaningfully improved version. It starts with the Fatal risk and does not avoid it. If that risk breaks the idea, it changes the idea. For each risk it decides: resolved, partially resolved, or unresolvable — with a specific experiment named for anything that cannot be resolved in analysis alone.
Prosecutor personas
Every Prosecutor run is assigned one of three attack personas. The persona is randomly selected alongside the model — adding a second layer of adversarial diversity beyond which AI system is attacking. The same submission attacked by The Assassin and then by The Contrarian will surface different risks.
The Assassin
Finds the single assumption the idea cannot survive being wrong about. Does not spread fire across the surface. Goes straight to the load-bearing point and attacks until it collapses.
The Skeptic
Demands evidence for every claim. Where evidence is missing, the claim fails. Where evidence exists, it asks whether it generalizes. Exposes the difference between what the submitter believes and what the submitter can demonstrate.
The Contrarian
Finds the angle nobody has said out loud. The uncomfortable read that the submitter has been avoiding. The framing that makes the idea look completely different from the one presented.
Three tiers of analysis
Round 1 results are always free. After reading the analysis you choose how deep to go. Each additional round attacks the previous round's refined version — not the original submission — with a different model, a different persona, and an explicit memory of every attack vector already exhausted.
Standard
1 roundOne full adversarial loop. Clarifier, Advocate, Prosecutor, Refiner. Results shown free. Decision record included with purchase.
Deep
2 roundsRecommendedTwo rounds. Round 2 Prosecutor reads the Round 1 Refiner's revised idea and attacks it with a different model, a different persona, and an explicit constraint: do not repeat any vector already exhausted. The idea that survives two rounds has been tested from two genuinely independent angles.
Gauntlet
3 roundsThree rounds. Grok joins as the fixed Round 3 Prosecutor — the final stress test on the twice-refined version. After all rounds a ReportAgent computes a survivability score and synthesizes what survived, what collapsed, and what persists across every round. The most thorough adversarial analysis available.
Gauntlet runs end with a ReportAgent synthesis — a survivability score (0–100) reflecting how much the idea changed under pressure and how many risks remain unresolved after three rounds.
Why the Prosecutor and Refiner rotate
Different AI models have different epistemic tendencies, training data emphases, and failure modes. A model that is consistently the Prosecutor would develop predictable attack patterns. By randomly selecting the Prosecutor and Refiner on each run — subject to the constraint that no model runs two consecutive stages — each analysis reflects a fresh independent perspective.
The Clarifier (Gemini + Google Search) is exempt from the no-repeat rule. Because its role is normalization and web grounding rather than adversarial reasoning, Gemini can appear as both Clarifier and Refiner in the same run without compromising epistemic diversity.
The model chain used for your specific run is logged and shown in your results and your PDF, so you know exactly which models evaluated your decision.
Five domains
VC Deal Review
Stress test an investment thesis before the partner meeting.
Founder Fundraising Prep
Find what investors will tear apart before they do.
Business Decision
Challenge a high-stakes strategic commitment.
Legal Strategy
Attack a legal position the way opposing counsel will.
Code Review
Find what could break, be exploited, or fail in production.