Triall

Triall is a specialized AI tool designed to detect and eliminate hallucinations in outputs generated by large language models like Claude, ChatGPT, and other AI systems. It operates through a multi-model peer-review mechanism that ensures only verified, factually grounded answers survive. Unlike standard single-model responses that often produce confident yet incorrect information, Triall orchestrates three independent AI models to answer the same question blindly. Each model’s response is then peer-reviewed by the other two, with reviewers explicitly checking for sycophancy, fabricated details, unexamined assumptions, and unwarranted confidence. The process begins with pre-analysis: before any model answers, Triall dissects the user’s query to identify hidden assumptions, classify the question type, and pinpoint potential pitfalls. Then, three models from different AI providers—each with distinct architectures and failure patterns—independently generate answers. These answers are anonymized and subjected to blind peer review, where reviewers flag overconfidence, invented specifics, and unquestioned premises. A convergence analysis step further guards against “relevance hallucinations”: when all three models agree confidently but offer no evidence, Triall marks the output as a high-risk hallucination, because agreement without supporting sources is the most dangerous kind. To strengthen reliability, Triall integrates real-time web search results at the start, so models work with current information rather than outdated training data. After initial answers are collected, an adversarial refinement loop begins: the best candidate answer is attacked by an adversarial critic model, which tries to poke holes, and then another model refines the surviving parts. This iterative