Make the strongest case against your proposal.
Bring a recommendation, thesis, or plan. Devil's Advocate reviewers challenge its assumptions and show which evidence could change the conclusion.
Investment Thesis Stress-Testing
Run every assumption through adversarial review before allocation.
Include the proposal, supporting evidence, alternatives, and known constraints. Tell reviewers what decision the work is intended to support.
By structuring the debate between multiple frontier models, you make competing assumptions explicit instead of relying on one fluent answer. The synthesis report states what must be true for the thesis to hold and preserves any surviving dissent. Adversarial pressure does not prove truth; it shows the fault line that needs evidence, a test, or a human decision.
How It Works
Choose Devil's Advocate as the review format. Reviewers receive explicit instructions to challenge the premise and examine weak claims.
A final model summarizes the strongest supported conclusion and the objections that remain. Use those findings to revise the proposal or identify evidence to collect.
Key Features
- gavelExplicit opposition: Gives reviewers a clear instruction to find weaknesses in the proposal.
- psychologyCognitive Diversity: Leverage the different training distributions of OpenAI, Anthropic, and Google models against each other.
- descriptionInspectable Review Record: Receive a structured, citeable artifact detailing every attack vector and defense.
What adversarial pressure changed in the public corpus
A lower consensus score can be the useful result.
55% → 27%
Warehouse humanoids
The panel moved from apparent middle agreement to a concrete trade-off: humanoid flexibility versus specialized automation's throughput, reliability, and maintenance burden.
63% → 40%
401(k) contributions
The question could not be responsibly reduced to one recommendation. Emergency savings, employer match, debt, income stability, and risk tolerance remained user-owned constraints.
50.9% completion
The cost of staying adversarial
Devil's Advocate completed less often than Refinement because its job is to keep the challenge alive. Use it when surfacing the failure condition matters more than reaching a fast synthesis.
These observations come from the LLM disagreement study: 558 completed debates, 9,507 claims assessed, and 2,476 model turns. It is an observational product corpus, not a guarantee that a model panel has found the truth.