What this stage does
Frame the scientific question, endpoints, population, and statistical approach that anchor everything downstream.
Where AI applies
Endpoint mining across historical trials, synthetic control arm exploration, and reasoning over prior evidence to pressure-test the hypothesis.
Under the Hood · 5
5 capability-angle deep-dives generalizing this essay's pattern for AI and software engineers.
GovernanceEvaluationHypothesis
The Question You Didn't Ask Gets Asked at the Worst Possible Time
A decision made before the outcome is visible can't be reverse-engineered from it.
GovernanceHypothesis
The Decision You Didn't Log
Architecture Decision Records for choices whose reasoning would otherwise survive only in memory.
RAGEvaluationHypothesis
A Design That Argues With Precedent
Retrieval that hunts for disagreement, not confirmation.
Digital TwinsGovernanceHypothesis
Simulate Against Honest Priors, Not Convenient Ones
A simulation with a flattering input is a forecast that was told the answer.
EvaluationGovernanceHypothesis
Write the Loss Narrative Before the Win Narrative
A premortem is honest doubt with social permission attached.