The grounding gate is the one place grounded decides whether to speak. It stands between the evidence and the answer and closes two failure modes with plain, repeatable checks. Before a single generation call is spent, it asks whether retrieval was strong enough: no hit, or a best chunk below a tuned cosine floor of 0.68, and the system abstains rather than pay a model to guess. After generation it runs four checks against the evidence it was actually given: the model must have claimed it could answer, every claim must carry a citation, every cited id must be one of the chunks or graph paths that were retrieved, and no protected topic may have leaked into the answer. A citation to something the model was never handed is rejected outright, and so is a leaked term.
That floor is the knob that trades recall for honesty. It was tuned against a labelled eval set where answerable questions score above 0.70 and out-of-corpus probes sit at or below 0.65, so 0.68 lands in the gap with margin. The model proposes the sentence; the gate decides whether it ships.
The load-bearing idea here is that abstention is a form of honesty, not a failure. When the evidence is thin the gate returns a flat "I don't have that in my public notes" instead of a fluent sentence no source supports. A confident fabrication and an honest refusal look equally polished on the page, and only one of them is safe to trust. The gate is deterministic precisely so the choice between them is never left to the mood of the model.