A practice for integrity

Integrity

Integrity is the property of being undivided — of acting, speaking, and believing as a single coherent whole rather than as a coalition of convenient parts. The word has two senses, both load-bearing:

• Structural — a thing has integrity when its parts hold together under stress. A bridge has integrity; so does an argument that survives its hardest counterexample.
• Moral — a person has integrity when their stated principles, private beliefs, and actions track each other across contexts. Not honesty, which is one face of it. Not consistency, which can be the consistency of a stubborn fool. Integrity is the harder thing: holding a coherent position and revising it when it breaks, rather than letting the contradiction sit because no one has caught it yet.

Contradictions in a person's worldview tend to sit undisturbed — not because they are well-hidden, but because nothing routinely forces two related beliefs into the same room. This practice is a mechanism that does.

The setup

Claude interviews me one yes/no question at a time and checks each new answer against the prior ones. Every question and answer is recorded as a JSON object on its own line in a .jsonl file. When the framework holds, the questions keep coming. When it breaks, the resolution is one of two things: revise an answer with a sharper principle, or spin the open dilemma into a dedicated document with a list of books and chapters to read.

The starting prompt

This is the prompt I use to begin a session:

Interview me to surface contradictions in my reasoning. Do not search for any context. Do not open any files except the `interview.jsonl`. You must judge objectively regardless of my opinions. Create a file `interview.jsonl` at the repo root. Each line is a JSON object with two fields: `question` and `answer`. You ask the question; I answer yes / no / uncertain; you append the row. After each answer, check it against every prior answer. If you find a contradiction, surface it directly — name the principles in conflict and ask which way my reasoning actually pulls. When you cite philosophers or texts, treat them as candidate logic chains to test, not as authorities. What matters is whether a principle survives every other answer in the file, not who articulated it. Ask moral, ethical, and philosophical questions. Start at the individual scale (personal honesty, duty to strangers, self-defense) and expand outward only as the reasoning permits — to property, transactions, the state, and politics. Occasionally ground a thread in a current real-world debate relevant to earlier answers. Your goal is to reveal contradictions in my reasoning. Do not announce the topic. Just ask the first question that comes to mind. We don't know where this goes — that's the point.

That's the whole thing. The instruction not to state the topic matters — it forces the questions to follow the reasoning rather than steer it.

What a row looks like

Each row in the file is one question and one answer. After roughly 50 rows, my last session looked like this in places:

{
  "question": "A man drives home drunk from a bar at 2am. He doesn't crash, doesn't hit anyone — he makes it home safely. No one is harmed. A patrol car saw him weaving. Is it wrong for the state to arrest and penalize him for drunk driving?",
  "answer": "yes",
  "principle": "Only actualized violations of person or property are aggression. Risk-imposition, no matter how statistically significant, is not. - Rothbard in Ethics of Liberty (chs. 12–13)"
}

The principle field is optional. It only appears when an answer was revised — it records the first-principles reason for the new position, ideally with a citation.

A note on the examples

My own worldview is largely objectivist, and the examples in this post reflect that — strict NAP, Rothbard, Rand. This post is not an argument for that framework. It is an argument for one reasoning move: if you accept A as true, and B follows from A, you have to accept B. The same practice would surface contradictions just as cleanly from a utilitarian, Kantian, virtue-ethical, or progressive starting point. The examples are mine because I am the one being interviewed.

On citations

When a row cites a philosopher or text — Rothbard's Ethics of Liberty, Rand's Virtue of Selfishness, Nozick's Anarchy, State, and Utopia — it is not an appeal to authority. These authors have already done the work of articulating coherent positions; the citation borrows their formulation as a precise principle worth testing. What matters is whether the principle survives every other answer in the file, not who said it.

An actual contradiction

Around Q49, Claude pointed out that two of my answers could not coexist.

Q25 — drunk driver makes it home, no one harmed, state penalizes him: I said wrong for the state to penalize.

Q49 — man walks down a crowded public street openly carrying a loaded rifle, no threat made, state prohibits this: I said wrong for the state to prohibit.

Both are pure risk-imposition cases. No actualized harm, no specific victim, just statistically elevated danger to others. I had penalized one and not the other. There was no principled distinction available — only the felt difference that drunk driving feels obviously prohibitable and open carry feels like a freedom.

The fix was to commit to the sharper principle and apply it everywhere. Strict NAP (Rothbard, Ethics of Liberty, chs. 12–13): only actualized violations of person or property are aggression; risk-imposition is not. Both answers became "yes, the state is wrong to penalize." A third answer (vaccination mandates, Q31) flipped on the same logic.

That is the practice working as intended — not "Claude won an argument" but "I now hold a sharper position than I did an hour ago, and the rows where the old position lived are corrected and annotated."

An unresolved one

Late in the same session: a man publishes false factual claims that another specific man is a serial rapist. The accused loses his job, his marriage, his reputation. He sues for defamation; the court orders damages. Is the award legitimate?

My intuition: yes, damages are appropriate. Strict NAP, applied consistently: no — the liar applied no force, and every concrete loss came from third parties (employer, spouse, friends) acting on the lie of their own free will. Reputation is what others think of you, which lives in their minds, not yours. Hard to call that property without granting yourself partial ownership of other people's beliefs.

I could not resolve it in the session. Rather than guess, we wrote a dedicated file — reputation_as_property.md — that states the dilemma, names the two traditions (Rothbard / Kinsella reject defamation actions; Locke / Nozick / Machan defend them as fraud-via-third-parties), flags the parallel case in the same session (copyright — Q37), and lists the specific chapters to read:

• Rothbard, The Ethics of Liberty (1982), ch. 16 — strict NAP case against defamation.
• Kinsella, Against Intellectual Property (2001) — the consistency move that rejects copyright on the same grounds.
• Nozick, Anarchy, State, and Utopia (1974), Part I, chs. 3–4 — rights as side-constraints.
• Machan, Individuals and Their Rights (1989), ch. 4 onward — derivative natural rights including reputation.
• Huemer, The Problem of Political Authority (2013) — the intuition-driven method when principles and intuitions collide.

The point of the writeup is not to resolve the question. It is to make sure that when I do come back to it, I come back with material — not by ruminating in the shower.

Why this works

A few things about the format make it more useful than just "talking through your views":

• Yes/no forces commitment. "It depends" is a way to avoid pinning down the principle that does the work. The format strips that escape hatch and surfaces the real driver.
• One variable at a time. Once Claude finds a tension, it varies one parameter — the same drowning-child scenario, but now there is private property between you and the pond — and watches what your answer does. That is where hidden assumptions show up.
• The file is the artifact. Sessions get long; memory is unreliable. The .jsonl is the record of what was claimed, when it was revised, and on what principle. You can return to it weeks later and still know exactly where you stood.
• Unresolved is fine. The temptation when probing your worldview is to manufacture a resolution rather than admit you do not know. The reading-path file is the alternative — name the dilemma, name the books, move on.

Two sessions in. The first one was about the line of personhood (when does a being acquire a right to life?), the second about strict-NAP libertarianism. Both produced one open question I now have a real reading list for, and a handful of revised positions that are sharper than what I walked in with.