Teaching supervised AI use with TopMVA

• Distinguish fluency from accuracy in AI output, and explain why a confident draft can still contain a fabricated case or number.
• Run a scoped, attorney-in-the-loop AI skill on synthetic facts and identify every flagged assumption in the output.
• Verify a legal citation from AI output against a primary source (Revisor, CourtListener) and articulate a personal verification habit.
• Apply confidentiality and competence duties to the decision of what client information enters any AI tool.

• 0–5 min — The failure mode: fluent, formatted, wrong. One synthetic example of a fabricated citation.
• 5–15 min — Live demo: run the intake skill on synthetic call notes. Walk through the [CONFIRM WITH ATTORNEY] and open-questions blocks.
• 15–25 min — What makes it safe(r): scoped instructions, refusal to invent facts, flagged assumptions, the verification protocol behind the authority file.
• 25–30 min — The non-delegable part: review, verification, signature. Q&A.

• The 30-minute arc, plus:
• +15 min — Verification exercise (§ 05 below) done live by the audience on their own devices.
• +10 min — Professional-responsibility discussion (§ 06 below): confidentiality, candor, competence, supervision of nonlawyer assistance.
• +5 min — The full five-stage chain: intake → records → chronology → evaluation → demand.

• 0–15 min — Framing and demo (compressed 30-minute arc).
• 15–45 min — Small groups install one skill each and run it on provided synthetic fact patterns. Each group logs every flag the skill raised.
• 45–65 min — Groups swap outputs and attack them: find the assumption that would hurt most if unverified.
• 65–80 min — Verification exercise on the citations in the outputs; compare against mo-authority.md.
• 80–90 min — Debrief: write a one-paragraph office AI-use rule you would actually adopt.

Use only synthetic facts, and say so on the slide. Never demo with a real matter, a real client name, or a real claim number — even "anonymized." Build a demo file the way this site does: invented names, a plausible crash narrative, round-number bills, and a visible label ("Synthetic facts — not a real matter"). The skill library demo panel shows the pattern, and each skill page carries a labeled synthetic sample you can reuse.

• Confidentiality — what client information may enter a given tool, under what terms of service, with what retention; the difference between a consumer chatbot and a firm-controlled deployment.
• Competence and candor — the attorney owns every fact and citation that leaves the office, exactly as if self-written; sanctions cases involving unverified AI citations make useful cold opens.
• Supervision — treating AI output like nonlawyer work product: useful, reviewable, never self-authorizing.
• Fees and candor to clients — whether and how AI-assisted drafting affects billing narratives.

1. Hand the audience one legal proposition from a skill output (e.g., a limitations period with its pin cite).
2. Have them find the statute on revisor.mo.gov and read the actual subsection — not a summary site.
3. Have them check one case through CourtListener: does it exist, is the reporter cite right, is there later history?
4. Compare findings against the entry in mo-authority.md — including whether it is marked verified or ⛔ OPEN.
5. Discuss: what would you have missed if you had trusted the draft?

Everything referenced is free to download: the five skills, the verification protocol, the research page, the citation formats, and the license. Speaker inquiries and questions about using these materials in a program: open an issue on GitHub.