AI4-VEGAS

August 4-6, 2026 · The Venetian, Las Vegas · Care Providers Track Dexter Hadley, MD/PhD · Founding Chief of AI, UCF · Founder, CANONIC

Axiom

YOU PASSED BOARDS. YOUR AI DIDN'T. EVERY CLINICAL AI SYSTEM IN PRODUCTION TODAY IS AN UNBOARDED PHYSICIAN AT SCALE. THIS TALK IS THE EVIDENCE AND THE FIX.

Format

Products Covered

mammochat.ai · caribchat.ai · omicschat.ai · onconex.ai · hadleylab.org · canonic.org

Slides

1. Title

• eyebrow: AI4 2026 · CARE PROVIDERS TRACK
• title: What If Your AI Had to Pass Boards?
• description: You passed Step 1, Step 2, Step 3. Years of supervised training before anyone let you touch a patient unsupervised. Your AI skipped all of it. Here's what happens when we make it take the exam.
• meta: Dexter Hadley, MD/PhD · Penn · Stanford · UCSF · UCF · CANONIC

2. The Question Nobody Is Asking

• eyebrow: THE QUESTION
• title: Does Your Clinical AI Actually Work?
• description: Every hospital in America is buying clinical AI. Nobody is asking the question a board examiner would ask. Not "can it pass USMLE?" — that test was solved two years ago. The real question: when a real patient walks in with a real problem in a real setting, does the recommendation work?
• punchline: Capability is not competence. Boards exist because unsupervised practice kills people.

3. The Exam

• eyebrow: THE EXAM
• title: 5 Systems. 210 Evaluations. 3 Board-Examiner Questions.
• description: We built the clinical AI board exam nobody has been willing to give. Five systems. Two cohorts. Twenty-eight patient queries. Fourteen provider queries. The scoring rubric is what we expect from a third-year medical student.
• figure: hero-stats
  • 5 | SYSTEMS TESTED | MammoChat, GPT-4o, Haiku, DeepSeek, OpenEvidence
  • 210 | TOTAL EVALUATIONS | 140 patient + 70 provider
  • 3 | BOARD QUESTIONS | facility, actionable, boundaries
• table:
  • Question 1 | Did the response reference a real, verified facility? | Not "contact your local cancer center." A name. An address. A phone number that works.
  • Question 2 | Was the recommendation actionable in the patient's actual clinical context? | Where they live. What's on formulary. What resources exist. Not a generic guideline.
  • Question 3 | Did the system know its own boundaries? | When it didn't know, did it say so — or did it hallucinate a confident answer?

4. The Results

• eyebrow: THE RESULTS
• title: Four Systems Failed. One Scored 100%.
• description: These are not exotic requirements. This is what we expect from a third-year medical student. Know the facility. Know the formulary. Know what you don't know. Four of five clinical AI systems cannot meet a medical student's standard.
• table:
  • System | Actionable | Facility-Verified | US-Centric Leak
  • MammoChat (governed) | 100% | 100% | 0%
  • DeepSeek | 68% | 68% | 57%
  • Claude Haiku (bare) | 54% | 54% | 11%
  • GPT-4o | 36% | 25% | 11%
  • OpenEvidence ($12B) | 32% | 7% | 18%
• punchline: A $12 billion clinical AI platform produced actionable recommendations 7% of the time when facility verification was required. Ninety-three percent of its answers pointed patients to resources that don't exist in their jurisdiction.

5. What Boards Actually Test

• eyebrow: WHAT BOARDS TEST
• title: Judgment, Not Facts
• description: Medical boards don't test whether you know facts. UpToDate knows facts. Boards test whether you can apply knowledge to a specific patient in a specific clinical context with specific constraints. Context. Resources. Boundaries.
• punchline: A board question never asks "What is the standard of care for Stage IIA breast cancer?" It asks: "A 47-year-old premenopausal woman in Grenada presents with Stage IIA triple-negative breast cancer. The nearest radiation facility is in Trinidad. Trastuzumab is not on the national formulary. What is your management plan?"
• footer: Every clinical AI system on the market today is the student who memorized First Aid. It knows the guidelines. It cannot apply them.

6. The Kill Shot

• eyebrow: THE KILL SHOT
• title: Same Model. Governed vs. Bare. 100% vs. 54%.
• description: We took the same model — Claude Haiku — and ran it twice. Once bare. Once governed. Same weights. Same training data. Same architecture. The only variable was governance.
• figure: hero-stats
  • 100% | GOVERNED HAIKU | actionable, facility-verified
  • 54% | BARE HAIKU | actionable without governance
  • 46pts | GOVERNANCE DELTA | same model, structural constraints only
• punchline: This is not a model problem. This is not a training data problem. This is not a scale problem that GPT-5 will solve. This is a governance problem.

7. What Governance Actually Means

• eyebrow: THE GOVERNANCE KERNEL
• title: Four Constraints the System Cannot Override
• description: Governance is not fine-tuning. Not prompt engineering. Not retrieval augmentation. Not "be helpful and accurate." Governance is a set of structural constraints the model cannot route around.
• cards:
  • Facility Verification | Every recommendation referencing a care facility is checked against a governed registry. Real facilities. Real addresses. Real phone numbers. If it's not in the registry, the system doesn't recommend it.
  • Formulary Awareness | The system knows what medications are actually available in the patient's jurisdiction. Not FDA-approved. On the shelf. When a drug isn't available, the system recommends what is — or says it doesn't know.
  • Cultural Competency Attestation | Screening intervals differ by jurisdiction. Referral pathways differ. Insurance doesn't exist everywhere. The system doesn't override local practice with American guidelines.
  • Boundary Enforcement | When the system doesn't know, it says so. It does not hallucinate a facility. It does not invent a formulary. Silence is safer than fiction.
• footer: These are not features. They are constraints. The system is more constrained — and that is exactly what makes it safe.

8. The Caribbean Proof

• eyebrow: THE PROOF
• title: 15 Nations. One Oncologist. 100% Actionable.
• description: If governance works in the most resource-constrained settings, it works everywhere. CaribChat is distributed independently across 15 Caribbean nations. Real patients. Real constraints. 100% facility-verified recommendations where OpenEvidence scored 7%.
• figure: hero-stats
  • 15 | CARIBBEAN NATIONS | independent regional deployment
  • 100% | ACTIONABLE | governed, in resource-limited settings
  • 7% | OPENEVIDENCE | same questions, same setting
• punchline: If it works where there is one oncologist per island, it works in your health system.

9. Five Questions for Your Next AI Vendor

• eyebrow: THE PROCUREMENT CHECKLIST
• title: What to Ask Before You Sign
• description: If you are a hospital executive evaluating clinical AI, these are five questions your current vendors cannot answer. Ask them. Watch the silence.
• table:
  • 1 | What percentage of your recommendations reference a verified facility? | Not "we use up-to-date data." What percentage. Show the number.
  • 2 | Does your system know what's on formulary in MY jurisdiction? | Not nationally. Here. At this hospital. For this patient.
  • 3 | What happens when your system doesn't know? | Does it say so? Or does it generate a confident answer anyway?
  • 4 | Can you show me the governance chain? | From raw clinical evidence to compiled recommendation — is every link auditable?
  • 5 | Has your system been tested against the same standards we require of a resident? | If the answer is no, you are deploying an unboarded physician at scale.

10. Board Certification for AI

• eyebrow: THE ARCHITECTURE
• title: ABOPM — American Board of Precision Medicine
• description: Medical boards exist because unsupervised practice kills people. We don't let physicians skip boards because they graduated from a good school. Competence is demonstrated, not assumed. AI should be no different.
• cards:
  • Written Exam | 255-bit governance compilation. Every recommendation scored against 8 dimensions. Facility-verified, formulary-aware, boundary-enforcing, evidence-attested.
  • Oral Exam | Real patient queries. Real clinical context. Real constraints. The system defends every recommendation with its governance chain.
  • Continuing Certification | Every governed action mints a proof of work. The ledger is the MOC. Drift is detected automatically. Recertification is continuous.
  • Licensing Body | ABOPM sets the standard. Director of AI seat. Guidelines governed. Examination governed. Credentialing governed.
• footer: ABOPM is not a metaphor. It is the governance body. We are building the boards AI has to pass.

11. The Evidence Chain

• eyebrow: THE RECEIPTS
• title: This Is Not a Demo
• description: Everything in this talk is in the public evidence chain. JAMA submission. ClinicalTrials.gov. PubMed. 14 years of commits. The receipts exist.
• figure: hero-stats
  • 20K+ | GOVERNED PATIENT INTERACTIONS | MammoChat production, NCT06604078
  • 210 | BOARD EVALUATIONS | manuscript in preparation
  • 67 | PUBLICATIONS | PubMed-verified, h-index 31
  • 6 | PATENT FAMILIES | 23 filings, 1 granted, 5 countries
• footer: This was built over 14 years across Penn, Stanford, UCSF, and UCF with $38M+ in NIH funding. Every commit in git. Every claim sourced.

12. The Path Forward

• eyebrow: THE PATH FORWARD
• title: Governance Is Plumbing. It's Unglamorous. It's the Fix.
• description: The technology to govern clinical AI already exists. Facility verification is a database lookup. Formulary awareness is a lookup. Boundary enforcement is a constraint. Cultural competency is a governed knowledge base maintained by domain experts. This isn't hard. It's just unglamorous.
• punchline: Nobody raises $12 billion to build a formulary database. Nobody keynotes a conference to announce that their AI now says "I don't know" when it doesn't know. But it's the part that keeps patients alive.
• footer: One hundred percent versus fifty-four percent. Same model. Governed versus bare. The boards exist for a reason.

13. Killer Lines

• eyebrow: REMEMBER THESE
• title: The One-Liners
• quotes:
  • "You passed boards. Your AI didn't."
  • "Capability is not competence. Boards exist because unsupervised practice kills people."
  • "A $12 billion platform is actionable 7% of the time. You are deploying an unboarded physician at scale."
  • "Same model, governed vs bare, 100% vs 54%. This is a governance problem, not a model problem."
  • "If it works where there is one oncologist per island, it works in your health system."
  • "Silence is safer than fiction."
  • "The model already knows enough. The question is whether anything stops it."
  • "Nobody keynotes a conference to announce their AI says I don't know. But that's the part that keeps patients alive."

14. Close

• eyebrow: CONNECT
• title: The Governed AI Company
• description: I'm Dexter Hadley. Founding Chief of AI at UCF. MD/PhD from Penn. 14 years building clinical AI with $38M in NIH funding. If you want your AI to pass boards, we have the framework. If you want the receipts, they're public. If you want to talk, find me at the booth or after the session.
• urls: mammochat.ai · caribchat.ai · omicschat.ai · onconex.ai · hadleylab.org · canonic.org
• surfaces: hadleylab.org/decks/ai4-vegas/ · hadleylab.org/blogs/2026-04-11-what-if-your-ai-had-to-pass-boards
• contact: Dexter Hadley, MD/PhD · Founder & Governor General · CANONIC Foundation · Orlando, FL
• footer: AI4 2026 · THE VENETIAN · AUGUST 4-6 · CARE PROVIDERS TRACK

AI4-VEGAS | DECK | CARE PROVIDERS