The platform's Treewalker engine inspects top-five logprob candidate distributions at every token position during baseline ungrounded generations. When an alternative token exceeds a defined probability threshold, the engine splices that token into the preceding text prefix and prompts the model to complete the divergent branch into a full concept or product name. Running this recursive exploration across multiple parallel iterations surfaces latent brand associations and secondary entities embedded in the model's parametric weights that would otherwise remain masked by standard top-choice sampling.