Triall

About Triall

Triall runs a question through three frontier models under an adversarial protocol and returns the answer that survived it, with a verdict on how well it held. The sequence is deliberate. First it grounds the question by pulling real web sources and feeding them in, so the models reason from evidence rather than recall. Then all three answer independently and blind — isolation is the explicit design goal, because models that can see each other anchor on the first answer and their errors correlate. The three answers are then stripped of authorship and handed back to all three models, each of which attacks and ranks every answer without knowing which one it wrote. The combined rankings pick a winner, which is then rewritten to withstand the sharpest criticism raised against it. What you get back is the surviving answer, a verdict describing how well it held up, and the sources. It is a directly useful shape for research and high-stakes questions where a single model's confident wrong answer is expensive. Access is through the web app, an MCP server or an API, so it can be called from inside Claude, ChatGPT or any MCP-capable client. Pricing runs on credits with both a subscription and a never-expiring pack option, and three free sessions need no signup at all.

Key Features

Tags