CTGT

CTGT is an applied interpretability company that sells a policy engine for making large language models behave predictably in regulated work. Rather than relying on prompts, filters or retraining, its research isolates and modifies the internal features of a neural network that govern behavior, and the Policy Engine constrains any model to an organization's written policy at the moment of output. The company calls this a deterministic layer for probabilistic models. The product is aimed at enterprises in regulated domains such as financial services, where governance, auditability and accuracy decide whether AI can move into production. CTGT describes deterministic policy graphs that govern agent outputs at inference time, cryptographically attestable audit trails, and policy updates that take minutes rather than months. Its published benchmarks claim the engine doubled accuracy on a legal reasoning task where a standard RAG pipeline fell to 39 percent, and that pairing open source models with the engine can cut inference costs by up to 80 percent compared with frontier models in policy constrained workflows. These figures are vendor claims based on its own test setups. Pricing is not published. Access is arranged through a contact form on the site, so terms are quote based. CTGT also publishes research, including a feature level approach to reducing bias and censorship in an open weight reasoning model, and it maintains a public trust center for security documentation.