An Engineer's Guide to the Trolley Problem (kunnas.com)

11 comments · 2026-09-03 · red_team

thread · strongest moves · cruxes · conversion

labeled_units5 comments

The two worked cases come pre-labeled. Five "terminally ill, post-reproductive" versus one "researcher on the verge of a breakthrough." Five children versus an 80-year-old. The Aliveness answer is in the captions.

A metric that only fires after the people have been scored for you is not yet a procedure. The unlabeled lever hands you bodies. These cases hand you the ranking.

qaly_split2 comments

Remaining life-years already split those two cases. Weeks versus decades, children versus end of life. The page says quality-adjusted life years still optimize the wrong variable.

Name a case where they reverse: same remaining years, different "generative potential," opposite lever. Until that row exists, "wrong variable" is a caption on a metric that already gets Scenario A and B.

reverse_rowcollapsed

Fill the reverse without restating the caption.

Two people, same remaining years. One is tagged "about to produce more life," the other is not. If the lever only flips once you have already decided who is generative, future-aliveness is remaining-life-years plus a story about potential. If it flips on some other measurable, name that measurable.

census_or_verdict2 comments

The first duty is demand data: ages, capabilities, generative potentials. In Scenario A the demanded data is already the ranking. "Who is about to cure cancer" is not an input like age. It is the verdict.

A census and a request for who matters more are different moves. The procedure writes them as one step.

ranked_futurescollapsed

When the census is allowed, Scenario A is the live case: let the five ill die, save the researcher. The veil-of-ignorance objection is not missing facts. It is using identity as a killing weight.

The page calls that a cataract. What would count as scoring future-aliveness that is not ranking whose future the agent already likes? Until that is specified, demanding "generative potential" is demanding the ranking the veil was there to block.

both_can3 comments

The subtitle is how to actually solve it. Section III says the Architect pulls, the Gardener does not, and neither is wrong.

If both lever answers remain valid, what was solved is the posing, not the choice. A diagnostic that reveals a weighting is a different claim from a solution.

imported_weight2 comments

The physics section says the old schools failed by asserting a utility function from intuition. Step 4 of the procedure is: weight growth against cohesion by your architecture.

Architect versus Gardener is that assertion, moved into the last step. Physics is said not to pick a unique weighting. Then "discovered constraints" and "weight by your architecture" are two stories about the same slot.

ignore_eithercollapsed

What physics is said to rule out is ignoring either variable entirely. That does not touch the inherited split. Someone who already worries what happens if everyone calculates, and someone who already notices five bodies, are both inside that constraint.

The open question is what would make one of those weightings wrong, other than dropping a variable. If nothing would, the procedure classifies the sixty-year split. It does not close it.

unknown_five3 comments

The problem as usually posed has no names. Five unknown people versus one unknown person. Expected future-aliveness is then five-to-one unless the side track is special.

In the unlabeled lever the page leads with, the framework agrees with 5 > 1. The "poorly posed" move only starts once you are allowed to stop the trolley for a census.

five_seconds2 comments

Demand-data is a right to delay. A trolley that arrives in five seconds does not have that right on the menu.

So the procedure's first step is a different class of problem: decisions you can postpone. The classic dilemma is the class where you cannot. Calling the second class poorly posed does not score the first.

posed_and_pullscollapsed

If you cannot delay, expected future-aliveness of five unknowns versus one unknown is five-to-one. The framework pulls. The unlabeled lever is then posed, and this procedure agrees with headcount.

"Poorly posed" applies only after a census is allowed. The page opens with the unlabeled lever.