The Verification Agent Designer

Design a verification step for this agent system.

What is being verified: [DESCRIBE THE OUTPUT]
What "correct" means: [DESCRIBE]
Cost of an incorrect result passing: [DESCRIBE]

Assess first:
1. Should verification be a separate agent, a step in the same
   loop, or deterministic code?
2. What can be verified deterministically and should not involve a
   model at all?
3. What genuinely requires judgement?

Then design the verification:
4. What the verifier sees — the output only, or the process too?
5. Fresh context or shared context, and why
6. What it checks, as an explicit list
7. Its output format: pass/fail, findings, or a score
8. What happens on failure — retry, escalate, or reject
9. How many verification rounds before giving up
10. What prevents the verifier and the producer sharing the same
    blind spot

That last point matters. A verifier running the same model with the
same context tends to agree with itself. Where independence
matters, what makes the verification genuinely independent?

Also: does the underlying model already self-verify? Some models do
this well by default, and adding an explicit verification step
duplicates work. Tell me how to check before building one.