Generative AIs still keep picking up speed in nearly every industry, but the trustworthiness of what they spit out is starting to worry people. The biggest headache by far is hallucination - when the model strings together something that sounds good but isn’t really true. In fields like health care, teaching, or money management, that slip-up can cause real harm, so no one wants to brush it off. Plenty of fixes have been tried already, yet most are just clean-up crews that look for false claims after the damage is done. This paper instead rolls out a forward-looking shield that tries to spot trouble before it even leaves the keyboard. Our approach mixes guessing-game math, relevance scoring, and feedback loops so the model itself learns what to avoid. Because of that, flaky sentences get tossed while still letting fluent, on-topic prose pass through. Tests show hallucination drops sharply without losing the smooth feel readers expect. Taken together, the work moves the field closer to generative models that people can safely trust.