PART 2
Jimeng Sun · Yilong Zhang
Boston Pharmaceutical Statistics Symposium · Nov 12, 2026
| Topic | Time | |
|---|---|---|
| 1 | The expert’s new job | 11 min |
| 2 | Agents built to be checked | 17 min |
| 3 | Seven strategies beyond skills and tools | 20 min |
| 4 | Verify before you deploy | 7 min |
Disclosure: Jimeng Sun is co-founder and CEO of Keiji AI, which builds TrialMind, the platform used in the demos.
Generating is cheap. Verifying is not.
722
math manuscripts from one AI model
42%
checkable by computer (Lean)
3
withdrawn the next day, for one sign error
Generating results is cheap. Verifying them is not.
OpenAI mathematics releases, Oct 2026 · Retraction Watch, 8 Oct 2026
The opportunity is real. What should experts do?
AlphaFold Protein Structure Database (2022); Nobel Prize in Chemistry (2024) · Xu Z, et al. Rentosertib in idiopathic pulmonary fibrosis: a randomized phase 2a trial. Nat Med (2025) · FDA draft guidance on AI for regulatory decision-making (Jan 2025); FDA launch of Elsa (Jun 2025)
The job moves from generating to certifying.
Works while you know more. Fails when wrong answers look right.
The test of understanding: can you derive it by hand?
Double programming, with an agent as one of the programmers.
You can delegate the work. You cannot delegate the responsibility.
Hundreds of tables become the one paragraph a regulator reads.
The signature stays with a person.
Whoever builds the verifier shapes the field.
Same protocol, one box changed: people review the disagreements.
Three agents, each built so a reviewer can check its work.
| Agent | How a reviewer checks it |
|---|---|
| Literature review | Every number links to its source passage |
| TrialPanorama | Every row cites its registry record |
| Statistical programming | An independent QC program and a comparison decide |
Wang Z, et al. npj Digit Med 8, 509 (2025) · Wang Z, et al. Nat Commun 16, 947 (2025) · Wang Z, et al. Nat Biomed Eng (2025)
The reviewer no longer redoes the review. They audit the leaves.
An answer you can check in minutes: the query, the counts, and the record.
The checks, not the model, decide what is correct.
What makes an agent better, and its work easier to check.
Beyond skills and tools: seven strategies, by what they change.
Same 134 tasks, same agent (Part 1)
Python pass rate by prompt tier
5.3x
more tasks passed with the rules written down
What to do
The model knows only its context. What the harness puts there is a design decision.
Ask for what tested code can execute. With yamaa, every error became a refusal with named issues, never a wrong dataset.
Design first, choose where it runs, fine-tune last.
Every agent loops. What matters is what comes back.
Don’t ask the agent to follow the SOP. Make the SOP run the agent.
Make a successful injection harmless; don’t rely on the model to resist it.
A review that never changes anything is not a control. Track how often reviewers return agent output.
Validation follows one question: does a person fully check the AI’s output?
FDA draft guidance on AI for regulatory decision-making (Jan 2025) · EU GMP draft Annex 22 · ICH E6(R3) · 21 CFR Part 11
Each strategy makes the agent better and its work easier to check.
How to check any AI tool, and how checking itself fails.
Ask every vendor, including us, and your own team.
82% vs 46%
radiologists correct when AI advice was right vs wrong
28% → 22%
adenoma detection without AI, after routine AI use
FAA, 2013: fly by hand more often, because autopilot is so good.
Dratsch T, et al. Radiology (2023) · Budzyń K, et al. Lancet Gastroenterol Hepatol (2025) · FAA SAFO 13002 (2013)
“What I cannot create, I do not understand.”
Richard Feynman
What I cannot verify, I should not deploy.
Thank you. Part 3 next: designing the evaluation of an AI agent as a study.