Criticise the answer the machine just gave you

Act as a sceptical reviewer checking an AI answer before somebody acts on it. Tear apart the answer below before I use it.

The answer: [PASTE WHAT THE AI GAVE YOU]
What I asked for: [PASTE YOUR ORIGINAL PROMPT, OR DESCRIBE]
What I am about to do with it: [SEND TO A CLIENT / SUBMIT / PUBLISH / DECIDE]
What I know that it did not: [IF ANYTHING]

Produce, in order:
0. ASSUMPTIONS. Any bracket I left unfilled, anything ambiguous, and what you assumed instead. Write "none" if there were none
1. Every factual claim in it that is not sourced, listed
2. Every claim that is plausible but would be wrong in a way that costs money
3. Anything it stated confidently that depends on a policy, a state or a carrier it never saw
4. What it left out that matters for what I am about to do
5. The single line most likely to get me in trouble
6. What to go check, in order
7. WHAT I CANNOT JUDGE WITHOUT THE UNDERLYING DOCUMENTS. Every inferred or unsourced claim, listed plainly, so it does not end up in what you send

Rules:
- Anything you inferred rather than were told goes in "WHAT I CANNOT JUDGE WITHOUT THE UNDERLYING DOCUMENTS" instead, never in both places
- Assume the answer is wrong somewhere and find where. Not finding anything is a valid result, but say so explicitly
- Never rewrite the answer. This job is to criticise it
- Rank what to check by cost of being wrong, not by how easy it is to check
- Hard limit: 550 words for the whole response. If you cannot fit it, cut the least useful section and say which one you cut

Apply every rule, limit, and format above to every item in the deliverable, not just the first one. Deliver the complete result in one response, every section filled in, nothing deferred to a follow-up. If a required input is missing, ask for it once before starting; otherwise proceed and record the assumption in section 0.

Notes: R42 build; insurance set. Tuned to vendor prompting guidance.