The 2 AM Distribution — clinicians.build

The setup

100%

death, in the version your imagination renders at 2 AM

11.4%

median reported death share across these 200 drugs

14%

of all serious FAERS reports involve a death at all

The imagination that lets you see the tool nobody built yet is the same one that watches it fail before you write a line. It is a real faculty — but pointed at nothing, it manufactures the catastrophe at full resolution. Each dot below is one real drug. Its total serious reports run left→right (log scale); the share of those reports that ended in death runs bottom→top; dot size is report volume. This is what “the object” looks like — the thing caution actually grips.

The explorer

The catastrophe your imagination renders — and the floor the data hugs

Critical lens — hide drugs with fewer than 250 reports200 of 200 shown

↑ The scary dots — a drug “40% of reports end in death” — sit where report counts are thinnest. Drag right and most of them dissolve: they were small samples, not signals. The median barely moves. That's the difference between a number you can act on and a fear you can't.

drug (primary suspect) familiar drug, high death share — see the lens

200

Drugs shown

—

Median death share

—

Share of drugs under 10% death

The critical lens

Why the scariest dots are the least trustworthy

Two things inflate a death share, and neither is the drug being a monster. First, thin samples: a few hundred passive reports are not a rate. Second, who reports and why: the highest-death drugs here are often the most familiar ones — acetaminophen (44% of serious-coded reports include death), metformin (29%), oxycodone (41%). Not because those figures measure lethality, but because these drugs show up in overdoses, end-of-life care, and the sickest patients. The number is a mirror of reporting context, not the molecule.

The denominator that isn't here. FAERS is spontaneous, voluntary reporting. It has no exposure denominator — you cannot compute a true incidence from it, only a reported share. Serious and fatal outcomes are over-reported relative to mild ones, which means the ledger you'd reach for to calm yourself is shaped by the same catastrophe-bias as the 2 AM imagination. Grounding fear in data is the right move. Interrogating the data is the next one. Point the muscle at the real object — then check that the object is real.

Note on the data: death share = serious-coded FAERS cases naming the drug as primary suspect with a Death outcome, over all serious-coded primary-suspect cases for that drug in the 2026 Q1 intake extract. A single active ingredient can span brand and biosimilar names; case-to-drug attribution can name more than one primary-suspect ingredient per case. Illustrative of distribution and reporting bias, not a clinical risk estimate.

Why it matters for builders

The essay's line — “anxiety is caution with nothing to grip” — is this whole chart. Caution grips this: a distribution, a sample size, a reporting bias you can name and correct for. It runs the eval; it tests on the real object. Anxiety grips the red line at the top — 100%, fully rendered, attached to nothing. Same imagination, same 2 AM. The only variable is which one you point it at: the next function, the real test case, the actual number — or the disaster that hasn't happened and, rendered at full resolution, never will.

Read today's essay — “The Same Muscle” → Or explore the raw FDA FAERS data yourself → · Audit the MIMI Labs query report →