Answers with receipts: grounding and citations

A grounded answer is one whose facts come from sources you supplied, not from what the model absorbed in training. In an enterprise that distinction is the whole game: the source is current, authoritative and reviewable; the model's memory is none of those. So you write an answer contract (use only these sources, cite each claim, say "I can't find that" when the sources are silent), you make the model cite at a granularity a reviewer can check (document, version, page, quoted span), and then you verify the citations in code after generation, because a citation that looks right is not the same as a claim the source supports. When verification fails, the answer is repaired, reduced or withheld, never shipped as is. Finding the right evidence is Week 02's problem. Using it honestly is this one.

• Grounded means the facts trace to supplied sources. The model's parametric knowledge supplies language and reasoning, not policy facts.
• Abstention is a feature. "The policy doesn't say" is a correct answer when the policy doesn't say.
• A citation is a claim about a source, and it can be false. Verify spans and numbers after generation.
• Faithful is not the same as relevant or correct. Measure each separately.
• Carry version and effective date end to end, or you will cite superseded policy with perfect confidence.