Familiar projects, like better accuracy or clearer explanations, that make an institution better at deciding without asking whether it should decide, and what to ask instead.
The criterion
A project is low-leverage when it makes the institution better at exercising power instead of asking whether, how, and under what conditions it should exercise that power at all.
None of the projects in the catalog below is worthless in itself. Metric precision can matter. Explanation can matter. Participation can matter. Each becomes low-leverage when it takes the place of the harder question: when a great deal of effort improves how power is used, and leaves untouched who holds it and who can challenge it.
The catalog
Effort goes into accuracy, coverage, and personalization, as if enough fidelity would settle the political question. It will not. Even a perfect prediction would not give the institution the right to decide.
Ask instead: When the model of the person is wrong, or right but should not decide, can the person still challenge it?
The model of a person is not the person →Years go into making the ledger more precise while what it leaves out stays the same. Accurate accounting inside the wrong boundary still gives a false account.
Ask instead: Whose costs does the metric leave out, and who carries what it cannot see?
Falsified denominator →Dashboards, model cards, and disclosure notices explain the system and leave it unchallengeable. Knowing why you were harmed is not the same as being able to reverse the decision, and seeing the system is not the same as being able to challenge it.
Ask instead: What can the explanation change, and who can make that change happen?
Law VIII: an observation counts only if it can change the system →A person clicks approve at the end of a machine-generated workflow. Without time, information, authority, protection, or a way to reverse the decision, that person does not make the system answerable.
Ask instead: Can anyone halt it, override it, investigate it, and change the rule?
Meaningful control →Workshops, codes of conduct, and exhortations improve behavior while people still depend on someone's goodwill. Better character does not replace better structure.
Ask instead: Which dependency can be redesigned so that it no longer rests on someone's virtue?
Dependence without standing →Citizen panels, employee feedback, and workshops take a great deal of effort while the decision rules stay the same. If participants can force nothing, the process has absorbed their dissent and changed nothing.
Ask instead: What can participants force the institution to reconsider?
Corrective standing →A workflow that gets faster by pushing friction onto patients, claimants, contractors, or support staff has only moved the burden. Do not optimize a system until you know where its leftover errors land.
Ask instead: Where do the leftover errors land, and does the ledger count them?
Absorption as concealment →Years go into negotiating principles such as fairness, beneficence, and responsibility, while deployments move into more private parts of people's lives. A system that does an unjustified job well is still badly designed.
Ask instead: Why is this institution entitled to make this inference or decision at all?
What does not convert →Repeated errors get repaired case by case, with better support and more patches. Past some number, the exceptions are evidence against the rule that produces them.
Ask instead: Has the pattern of exceptions gone to the rule's owner as evidence against the rule?
Corrective learning evals →Whether the machine is agentic, aligned, or controllable matters less than the relationship the institution builds around it. A perfectly obedient algorithm can still carry great power that answers to no one.
Ask instead: An obedient machine does not make the operator answerable. What does?
Democratic vs. coercive governability →One more dataset, one better model, one more sensor, each imagined as finally removing the need for judgment, challenge, and exception handling. Complete representation is not the goal.
Ask instead: Is the system still legitimate when, as it will, it does not know?
Safe incompleteness →Nothing here says expertise is worthless. Expertise keeps its authority on the questions it is qualified to answer. What it loses is the step from knowing more to deciding for others. And nothing here says efficiency is worthless. The objection is to efficiency figures that leave out the people absorbing the errors. See also abstention, the choice not to decide, which most of these projects stand in for.
The permanent question
A project has restored leverage when getting someone wrong is no longer final. The exception travels back to the rule that produced it. The person who carries it can afford to challenge. The decision can be reopened, which is what non-finality names. The tools on this site are built to check that standard.
The about page states what domination means here, why no arrangement is exempt by design, and the discipline in brief. The glossary defines the ideas underneath: the excluded case, consequential contradiction, and safe incompleteness.