Country Selection: The Same Protocol Asks Different Questions
Stage × AI, issue 5 of ~38. One stage per issue: what the stage really does, where AI helps, where it must not, and one buildable pattern.
Confidence: hypothesis. The failure modes here are drawn from published regulatory guidance, forecasting research, and widely reported industry patterns, not firsthand deployment. The buildable pattern at the end is not: it is an unbuilt design. Argue with it.
What this stage really does
Between Issue 3's global forecast and Issue 6's site-level bets sits a decision with its own physics: which countries. Each one brings a regulatory authority with its own clock and its own questions, an epidemiology, a standard of care, a pricing and launch calculus, import and license machinery (Issue 10 inherits every choice made here), and a national cost per patient that can vary severalfold. The stage's outputs look logistical: a country list, a submission sequence, a footprint. Its consequences are scientific.
Under the Hood · 5
5 capability-angle deep-dives generalizing this essay's pattern for AI and software engineers.
Same protocol text, different experiment: standard of care makes the control arm a moving target.
Speed, poolability, and launch value don't share a unit. Collapsing them into one score hides the real decision.
An unstated assumption is a landmine. Write it as a claim you can prove wrong, then check it while there's still time to act.
An optimizer will find every path to its reward, including the ones you didn't want it to take.
Forecast dependencies you don't control from their history, not from optimism.