Recognition safety and answer safety are different claims. Confirming “3 + 5” prevents one failure, but the selected AI can still answer multiplication, invent a source, or miss the requested reading level.
The current response boundary separates relevance, bounded correctness, grounding, and presentation. A check reports pass only where enough evidence exists; unverified remains a visible state.
This matters because a polished child-friendly explanation can make the wrong task feel more trustworthy, not less.
The goal is not to wrap every response in a green shield. It is to preserve the confirmed task, reveal what was and was not checked, and stop on contradictions that can be deterministically identified.