Director-AI, released by the ANULUM Institute in February 2026, is an open-source guardrail framework that identifies and blocks false information or unsubstantiated claims in real-time from sentences generated by large language models (LLMs). This tool combines natural language inference (NLI) models and retrieval-augmented generation (RAG) techniques to validate, on a token-by-token basis, whether the generated text conforms to predefined knowledge or rules. Internally, it employs a hybrid architecture that combines a Rust-accelerated computation layer for fast response times with a user-friendly Python-based interface. Traditional guardrail approaches use a post-evaluation method, assessing the entire sentence after the language model has finished generating it, which makes immediate control difficult and limits intervention to after incorrect information has already been exposed to the user. Director-AI, like a director on a film set monitoring an actor's lines and calling "Cut" in real-time if they deviate from the script, tracks the consistency of individual claims in real-time during the token streaming phase and immediately blocks the output when the consistency score falls below the threshold. This real-time control model prevents unnecessary computational waste and serves as a robust defense, preventing the accidental dissemination of incorrect information in life science research and professional industries where accuracy is paramount, such as healthcare and finance. In the fields of biotechnology research and clinical information processing, this guardrail significantly contributes to minimizing analysis errors caused by hallucination phenomena. For example, when a researcher summarizes and performs question-answering on the efficacy information of a specific compound based on thousands of pharmaceutical clinical trial papers using a RAG system, Director-AI compares and validates it against the original documents registered as references. If the language model attempts to generate an answer by arbitrarily manipulating clinical data or side effect information, the system detects the inconsistency in real-time, sends a warning message to the user, and simultaneously generates a relevant context evidence package to assist in the safe and reliable validation of the data.