We built an AI system that quotes specific numbers — financial figures, regulatory filing data, transaction details — back to users in a regulated domain. Every wrong number is a legal problem, a reputational problem, or both. After watching off-the-shelf LLMs confidently invent revenue figures from public filings, we built a grounding verifier that audits structured claims against the underlying tool output before the response leaves the server. Unverifiable claims get flagged and surfaced; never silently passed.

This is the practitioner walkthrough — what we tried first that didn't work, why each approach broke in instructive ways, and what the fail-closed instinct really means when wrongness is expensive. The three failed approaches before we landed on what works: regex pre-filters that missed semantic claims, LLM-as-judge audits that just doubled the cost of being wrong, and strict schema validation that broke as the response space evolved. The principled scope of structured verification — what kinds of claims a per-claim verifier can defend, and what kinds it structurally cannot. The specific failure category we caught early, including the time the model invented an entire insider sale that never happened.

By the end you'll have a concrete framework for thinking about verification in your own production AI systems — what the architectural tradeoffs are, what bug shapes to expect in week three, and the difference between bolt-on hallucination mitigation and architectural verification. The case study is finance, but the principles generalize to any domain where AI output is consumed as fact.

If you've ever shipped an AI feature and then prayed nobody would quote it back, this talk is for you.