The grounded research engine drives an iterative discovery pipeline using gemini-3.6-flash over a configurable sequence of 1 to 12 research rounds (defaulting to 8), equipped with live Google Search and URL context tools to extract factual propositions about a subject brand. Round one initializes the search using the base grounded_memory prompt, while all subsequent iterations inject a cumulative list of previously discovered claims into the grounded_memory_round template to instruct the model to uncover only novel, unmentioned facts. Output extraction is strictly enforced through a structured RESPONSE_SCHEMA, returning discrete factual assertions paired with their supporting page URLs and titles up to a cap of 60 items per round. Each fresh fact is normalized and recorded with a rank-based confidence score into grounded_memory_facts and grounded_memory_sources, while an automated diminishing-returns stopping rule terminates the execution early if an entire round yields zero new distinct findings. Once the research phase finishes, the accumulated evidence is automatically handed off to the Fact & Gap Checker to reconcile live discoveries against parametric recall.