A political test that tells you what it can’t see

What actually happens

There are fourteen short assessments. Most are real, published psychology questionnaires — the same ones used in academic research, cited, with links to the original papers. You can do one and stop, or do all fourteen. Nothing is required, no account is needed, and your answers stay in your browser unless you choose otherwise.

A short one takes two minutes, the longest about five. All fourteen, if you ever want them, is about 32 minutes at their shorter length and about 47 on the path you are actually put on — and there is no reason to do them together. Each produces its own reading, and the estimate on each is worked out from that survey’s own questions rather than a guess, rounded up.

What comes back is a three-layer picture. Underneath sit stable dispositions — how you react, what disgusts you, how much novelty you want. Above those sit values: what you think fairness requires, how much you defer to authority, whether you think the system is basically fair. On top sits the visible part: where you place yourself, who you trust, which political language moves you.

The claim the model makes is that the top layer is downstream of the others. Not that it is caused by them — that the two are connected, and that seeing the connection tells you something a label cannot.

What the quadrant quizzes miss

The famous political quizzes have a problem that almost nobody who takes them ever hears about. Here it is, without jargon.

Imagine two people, Dana and Sam. They believe exactly the same things. Both are moved by language about order and unmoved by language about equality. Dana answers honestly. Sam is agreeable — when shown a statement, Sam nudges everything up a bit, because Sam tends to agree with sentences.

Read straight off, Sam looks like someone stirred by egalitarian rhetoric. Sam isn’t. That number is Sam’s manners, not Sam’s politics — and if every statement in a test points the same way, nothing in the result can tell the two apart.

This is the single most cited flaw in the best-known political compass test, which carries far more right-coded statements than left-coded ones — so agreeable people drift rightward as a side effect of how the questions were written, not because of anything they believe.

The fix is unglamorous: every scale has to contain statements pointing both ways, so that a habit of agreeing cancels itself out instead of piling up. This project checks that automatically and refuses to build if a scale fails, apart from seven exceptions, each listed with its reason on the methodology page. (The technical name is acquiescence bias, if you want to go read about it.)

What it cannot see about you

Everything here has edges, and hiding them would make the confident parts worth less. Some of the honest limits, in plain terms:

• It cannot detect authoritarian instincts arriving from the left.
  The standard research scale for authoritarianism was built to measure deference to established authority, which is a right-leaning shape by construction. Aggression against authority in the name of justice is a different motion and that scale does not register it. This project ships its own measure aimed at the gap, but that measure is homemade and unproven.
• It has no idea what is normal.
  There are no real respondents yet. Not few — none. Every score is your answers run through published research, not your position among other people who took it. A percentage here is not a percentile.
• It cannot tell a considered answer from a fast one.
  It can tell whether your answers contradict each other, and it reports that. It cannot tell whether you meant them.
• Four of the fourteen assessments are this project’s own writing.
  They are Moral Foundations, System Justification, Historical Attribution, and Quote Resonance, and each says so on its own results screen. Two of them exist to reach things the published instruments structurally cannot. The other two carry a published theory's constructs on questions written here — which is a weaker thing than borrowing the instrument, and was labelled as the stronger thing until August 2026. Nobody has ever checked whether any of the four work.

Each assessment carries its own limitations note where you meet it, and the full list — with every source and its DOI — is on the methodology page.

You can check the model against real research

The model claims certain traits are connected. Rather than ask you to take that on trust, it will go and count.

Pick any two of the 41 traits and the app searches OpenAlex — a free, open index of scholarly work — for what has actually been published on each one alone, and on the two together. Three numbers, not one, because a joint count means nothing on its own: seven papers linking two subjects is a lot if each has eighty, and nothing at all if each has eighty thousand.

Anything worth returning to can be kept. The Bench holds what you keep and ranks it by how little has been published — alongside whether your own two scores go with or against what the research predicts, which is the one thing a literature search alone can never tell you.

It will also measure how much the answer depends on the exact words searched. This is not a detail. When a leading AI research tool was tested independently against four published evidence reviews, its searches found under 40% of the studies those reviews had included, where the reviews’ own searches had found about 95%. A search that returns nothing is a fact about a query before it is a fact about a field, so the query is printed beside every number here and you can rewrite it.

What this is, honestly

A public prototype, not a validated research instrument. Using it for actual research would need licensed instruments, proper translations, pilot testing, reliability analysis, a sampling plan, informed consent and ethics review — none of which this has.

It is also not a diagnosis, not a clinical tool, and not an opinion about whether your politics are correct. It has no view on that. It is a way of looking at the machinery underneath a position, including the parts of the machinery that are just how you happen to answer questions.

Related: how the model is built, with every instrument and source · The Field, what a sample would let it say · case studies, where the scoring engine is run against real documented public figures · privacy.