Five phases. Zero blind spots.
Every Consium session follows a rigorous 5-phase pipeline designed to catch what single AI models miss.
Explanation
Two independent analysts evaluate your topic in parallel.
Two AI analysts interpret your topic and select the right panel of experts.
The explanation phase ensures your topic is properly framed before any debate begins. Analysts identify the key dimensions, select relevant persona archetypes (economists, ethicists, historians, domain experts), and determine the optimal number of discussion rounds. This prevents the common single-AI problem of framing bias — where the way a question is interpreted determines the answer.
Discussion
Configurable rounds with 3-12 personas debating simultaneously.
AI personas debate in structured rounds. You can join any round.
Each persona brings a distinct lens to the topic. They don't just state opinions — they present evidence, cite reasoning, challenge each other's assumptions, and build on strong arguments. You can inject your own perspective at any point, steering the discussion or adding constraints the AI might have missed. The discussion is structured, not free-form, ensuring every relevant angle gets explored.
Observation
Observers scan every 6 messages and must unanimously approve.
Observers monitor for hallucinations, bias, and groupthink.
Observer AI models operate independently from the discussion participants. They check for: factual accuracy (are claims supported?), logical consistency (do arguments follow?), cognitive biases (is the group converging too quickly?), and completeness (are perspectives missing?). If observers detect issues, they can interrupt the discussion, flag specific claims, and require corrections before the debate continues.
Adversarial Review
Challenge personas test every assumption in the consensus.
Devil's advocates stress-test the emerging consensus.
The adversarial phase is what separates Consium from any "multi-prompt" approach. Dedicated devil's advocate personas are tasked with attacking the emerging consensus from every angle. They look for: weak evidence, untested assumptions, overlooked alternatives, logical fallacies, and confirmation bias. The consensus must survive this adversarial pressure to be considered robust.
Report
Two independent reporters ensure balanced, comprehensive output.
Two reporters synthesize everything into a structured report.
The final report isn't a summary — it's a structured analysis document. Two independent reporter AI models synthesize the entire discussion into sections: executive summary, methodology, key findings with supporting evidence, areas of consensus and disagreement, documented dissenting opinions, confidence levels, and actionable recommendations. Having two reporters ensures no perspective gets editorially buried.