How this works, and what it isn’t

Worldview asks 43 forced-choice questions in three layers (a small number of respondents see one extra, conditional follow-up question — see “The blame question” below): Epistemics (L1) — how you decide what’s true at all, before politics or morality enter the picture, six axes with three questions each; Morals (L2) — which of six moral foundations (care, fairness, loyalty, authority, liberty, sanctity) you lean on when two of them conflict, every pairing of the six asked once; and Positions (L3) — where you actually land on economic, social-order, and authority questions.

You answer L1 first and L3 last, on purpose, so your political vocabulary doesn’t contaminate the earlier layers.

Then we do one thing no label-generating quiz does: we build a model that predicts your L3 answers from your L1 and L2 answers, and we show you the gap between what that model predicts and what you actually said.

It is not a measurement of inconsistency, bias, hypocrisy, or bad reasoning. A large gap does not mean you’re wrong. It means your political positions don’t follow, by this particular model, from the rest of what you told us — which is completely normal. Most people’s politics are shaped by lived experience, specific policy knowledge, group identity, and plain disagreement with a model’s assumptions, not purely by abstract epistemics and moral weighting.

A Euclidean distance between two hand-tuned, unstandardized vectors has no scale. “31%” means nothing without a distribution of real completed sessions to compare it against — printing it anyway would be exactly the false authority this instrument is built to avoid.

So below 500 completed sessions, the gap is reported per-axis and qualitatively: which of your positions follow from your foundations, and which don’t. Above that threshold, percentile phrasing unlocks — “your positions diverge more than roughly two-thirds of people who’ve taken this” — still never a bare number. This is checked live, against however many sessions actually exist, every time a report is viewed, not decided once at build time.

Each of the six epistemic axes is answered three times. If the two axes that matter most to your result each had their three answers split — two one way, one the other — instead of agreeing, we say so plainly instead of averaging the disagreement into a confident-looking label. A forced archetype on noisy input is the fastest way to turn an instrument into a horoscope.

One question branches. Everyone is asked whether someone who does something cruel deserves blame. If you said determinism is true elsewhere in the instrument (your choices were never really open) and you said blame is still deserved, we ask one more question: is blame deserved because they could have done otherwise, or because the act came from who they are, regardless of whether they could have done otherwise?

The first answer is a real contradiction with saying determinism is true, and we flag it. The second is compatibilism — a serious, coherent philosophical position with a long history — and we don’t flag it. A trap that fires on every compatibilist would be a bug, not a finding.

The mapping from L1/L2 to a predicted L3 position lives in one file, /config/derivation-model.ts, as a hand-set 3×12 coefficient matrix. Every single coefficient has a comment next to it explaining the reasoning behind it. It has not been fit to any response data — there isn’t a dataset of real answers to fit it to yet. It is one team’s best-effort hypothesis, encoded so it can be argued with, inspected, and eventually replaced by coefficients fit to real data without touching any other part of the app.

If you disagree with a coefficient, you’re probably right that it’s debatable — that’s the point of publishing it in the open rather than hiding it behind a black-box “algorithm.”

• We do not claim this is a validated psychometric instrument. It hasn’t gone through peer review, test-retest reliability studies, or factor analysis.
• We do not report false precision anywhere — no p-values, no confidence intervals, no decimal places on any constructed score.
• We do not tell you your worldview is correct, rare, sophisticated, balanced, or healthy. The instrument describes; it does not rank.
• The “nearest thinkers” are figures whose published positions land near yours on the three L3 axes. Proximity, not endorsement, in either direction — for you or for them.

Every question is plain data, editable without touching any scoring logic. The rules we hold every item to:

1. If a smart person can’t sincerely pick either side, the item is broken.
2. If you can guess someone’s politics from one L1 or L2 answer, the item is contaminated — cut it.
3. Concrete beats abstract every time.
4. Never let one option be the longer, more nuanced-sounding one.
5. Three items per L1 axis, minimum. The L2 round robin stays complete.
6. A trap that fires on a coherent worked-out position is a bug, not a finding.

The scoring logic is unit-tested against a perfectly coherent respondent, a maximally incoherent one, an all-A respondent, an all-B respondent, an isolated high-dispersion axis, straight-lining (detected by median response time, not mean, so one outlier can’t hide or fake it), both attribution-asymmetry trap patterns, and both branches of the blame question. Nothing about how your particular result was computed is hidden from you.