Answer 400 short statements with Yes, No or Not sure, and your belief style on 5 axes, plus your match with 15 religions and worldviews, updates live with every answer. It isn't a test that guesses your religion. For each item, how 15 traditions would answer is written down in advance and compared with your answer, so the result is less "You are a Buddhist" and more "Your answers overlap with these traditions this much." It runs as a single page with no server and no model calls.
At first the address was /god
The slug in the first commit was god. It looked like an app that asks about God, and the early screens really did follow every answer with a screen like "Why do you think so?" Only after pulling out those follow-up screens, changing the address to /worldview and swapping out the whole item bank did it take its current shape. Answers were also cut from a 5-point scale to three: Yes, No, Not sure. You'd think a shorter scale means less information, but in practice the problem of "3" meaning something different to every person goes away, and the math gets more honest.
Even opposite answers had a 30% floor
After it went public, the first thing the user pointed out was the numbers. Answering as a hard-line atheist, Christianity and Islam still came out in the 30s and 40s. That's strange, they said.
Opening the code, the match was computed as Σ weight×match / Σ weight×max match. With only the maximum in the denominator, the floor wasn't 0. On items where a tradition's expectation is "mostly yes" (a weak expectation), even answering exactly the opposite leaves 25–33% on paper. Stack that floor across 400 items and you get a screen where you overlap 30% with everyone. I changed the denominator to max match − min match and nailed the floor to 0. Now 0% means you flipped every single one of that tradition's expected answers, 50% is neutral, and 100% is a full match.
Denying reincarnation was raising the Christianity score
Even with the scale fixed, the atheist's Christianity was 41%. This wasn't a scale problem but a weighting one. Item weights used a single number: "how much answers split across traditions" (the standard deviation).
I took a hypothetical respondent answering exactly as atheism is expected to and pulled the 132 items where they gave the same answer as Christianity, sorted by weight. The top looked like this.
- Humanity's basic problem is being bound by karma to repeated rebirth: Christianity −2, atheism −2
- After death, we are reborn in another body: Christianity −2, atheism −2
- Multiple gods or divine beings really exist: Christianity −2, atheism −2
These items separate Hinduism and Buddhism, so they got high weights, and since Christianity and atheism both say "No" there, the match went into the Christianity score. Denying reincarnation doesn't bring you closer to Christianity, though. I changed the weight from a value shared by all traditions to each tradition's own discriminating power, |that tradition's expectation − mean across traditions|. If Christianity gives the average answer on an item (if it isn't a distinctly Christian position), it barely counts toward the Christianity score. Christianity dropped from 41% to 30%, Islam from 34% to 25%. Meanwhile, for someone who answers as a Christian, Judaism, Islam and the Bahá'í Faith still land at 75–77%. Relatives stayed relatives; only the distant ones moved further away.
One first answer sent the belief style to an extreme
The "Your belief style right now" section at the bottom got flagged too. Shouldn't there be a middle, and in-betweens, instead of jumping to an extreme all at once? The cause was computing axis scores as a simple average. Answer the first item on an axis and that axis score is pinned straight to ±1.00, the maximum extreme. Most of the 9 archetypes' centers are also ±1, so the moment it picked the nearest type, an extreme name popped out.
I added 2 to the denominator: Σsign/(n+2). One answer gives 0.33, three give 0.6, twelve give 0.86. It's a common way to pull toward the center when the sample is small, and in code it's one line. The display was also split four ways. If the axes are still near the center, Centrist; if the top two types are about equally close, Between two types; once it's clear, Single type; and if even the nearest type points the other way on half of its axes, No clear type yet. The last one matters. It's the spot where no name is forced on you.
I removed the 78% figure
"Closeness 78%" was shown next to the type. The user asked why answering as a Christian only gave 78%. Opening the calculation, that number was 1 − distance/2. The trouble is that the distance depends not on the person but on where the 9 centers were placed. The axis vector of someone answering as a Christian sits at −0.38 on the norms/practice axis, that is, toward the inner/autonomy side, while the Guardian of Tradition center is at +1 and Conversation with God is at −0.5. It's a hybrid that sits on none of the centers. 78% didn't mean "a worse fit"; it meant "someone caught between centers."
I removed the percentage and changed it to "4 of the 5 axes this type defines point the same way." Better to show only what can be counted. Making the change actually exposed a hole in the data. Answering as a Jain, even the nearest type matched only 1 of 5 axes. None of the 9 types had a place for extreme asceticism and nonviolence. I made a tenth type, Ascetic Purifier, to fill it.
A machine sweep of 5,600 item cells
Last, I checked 400 items × 14 traditions = 5,600 cells for points or penalties wired to the wrong tradition. It's more than a person can read, so I wrote a script that sweeps it four ways.
The biggest catch was 7 signature items tied to an axis. For example, "Right belief matters more for salvation than right action" is affirmed only by Christianity, but it was tied to the norms/practice axis. An atheist answering "No" there is rejecting the very frame of salvation, not answering whether norms or the inner life come first. Yet the axis score counted it as one vote for norms/practice. For the same reason, Jain vegetarianism and ahimsa, the Bahá'í world community, Hindu and Shinto lineage-based faith, and Judaism's "faith that argues with God" were each pulling the wrong axis. On those seven items I removed only the axis tag and kept the expected answers. For telling traditions apart, they're actually the best items.
One cell turned out to be miswired. There are two items asking about devotion to the guru. One had Sikhism at +2; the other had identical values for the other 13 traditions but Sikhism at 0. "Sikh" means disciple, and its whole account of salvation centers on the guru, so I set it to +2.
The rest was clean. A check on who owns which proper terms, across eleven families (Kaaba/sharia/Quran, Trinity/Bible, Torah/Sabbath, Four Noble Truths/nirvana, ahimsa, guru, kami/torii, wu wei/yin-yang, Ahura Mazda, Bahá'u'lláh, Vedas/moksha), found no item where the protagonist was a tradition other than the term's owner. In pairs of items asking the same proposition forward and backward, no tradition expected the same answer to both. All 11 cells where the 3 nearest traditions were exactly opposite were genuine doctrinal forks: the Trinity, original sin, a sole savior, the finality of Muhammad, sharia, the Kaaba. Those are exactly where it should come out that way.
Adding Confucianism exposed the axes' weak link
To the question "Why is Confucianism missing?", I had no answer either. So it went in as the fifteenth. That meant assigning Confucianism's expected answer to all 400 items, and once that was done, Confucianism stood almost flat across the 6 axes. In particular, the norms/practice axis read 0.05, neutral. A tradition of ritual propriety (li) being neutral on the norms axis makes no sense.
Looking at what was pulling it down, the items pushing Confucianism toward the inner/autonomy side were the Sunnah, sharia, the Sabbath, halal slaughter, outward signs of faith, confession to clergy, and the duty of pilgrimage. All of them are other traditions' own norms. Confucianism answering "No" to them isn't choosing the inner life over norms; it just doesn't use those particular norms. The axis counted it as a vote for inner/autonomy. The same flaw I'd found in those seven items showed up again as soon as one more tradition was added.
So I wrote 16 new items on what Confucianism actually asks about: filial piety and mourning rites, the formalities of coming-of-age, weddings, funerals and ancestral rites, hierarchy and proper order, the principle of Heaven and the Mandate of Heaven, study of the old texts, ruling by virtue, and the duties of this life over what comes after death. The item count went from 400 to 416. I ran the same diagnosis on Hinduism and the Bahá'í Faith and wrote 12 more: the duties of one's station by birth, stages of life, rites of passage, communal festivals, the beginninglessness of scripture, consultation and regular gatherings, abolishing prejudice, and equality of women and men. That made 428. Then, rerunning the audit script, three items on sacred images, natural objects and ancestral spirits were flagged as uncorrelated with the transcendence axis. Same reason. I detached them.
The way types were chosen had a hole too. Distance was divided by the number of axes the type defines, so types defining fewer axes came out structurally close to everyone. I noticed when a three-axis type built for Confucianism was pulling in Hinduism too. It took two missteps to fix. When I had it pick by highest match rate, the Guardian of Tradition type, which defines many axes, swallowed four traditions instead. In the end I fixed the distance itself: axes a type doesn't define are treated as "neutral on that axis," and distance is measured across all six. On top of that, only types where at least half the defined axes point the same way remain candidates. Confucianism lands on Cultivator of Human Bonds with all 3 axes in the same direction, and for someone answering as a Confucian, Shinto follows at 79% and Taoism at 76%.
"I don't know" is not the same as "this question can't be asked"
The user asked: is that the button to press not only when withholding judgment but also when the question itself is meaningless? Measuring it, this wasn't something to let slide. Of 416 items, only 81, the ones asking about God, ultimate reality and the afterlife, count "Not sure" as a match with agnosticism. And those 81 were precisely the places where you'd want to press "this question doesn't hold up."
For a hypothetical respondent answering as an atheist, I switched just those 81 items to "Not sure." Agnosticism jumped from 85% to 96%, and second place switched from pantheism to agnosticism. "I don't know whether God exists" and "questions about God can't be true or false in the first place" are different positions, but one button was putting both in the same box.
I made a fourth option: Question doesn't hold · Reject the premise. At first I dropped this answer from every tradition's score, but running it, agnosticism still rose to 96%. Those 81 items were the ones dragging down the agnosticism score when an atheist answered firmly, so dropping them removed the drag and the score went up. So instead of excluding it, on items where agnosticism withholds judgment, it's counted as a mismatch. Withholding judgment and dismissing the question itself relate to agnosticism differently. Now if you pass with Reject the premise, it stays at 85%. Someone who answers every item with Reject the premise comes out at a 0% match with agnosticism too. That person is not an agnostic.
The results plotted in a cube
Seeing it only after finishing the whole thing is too late. The live panel shows the same map while you're taking the quiz, so you can see your dot move with each answer. A ranked list and a hexagonal radar alone don't really tell you "where am I?" So I made a box out of three axes and plotted the 15 traditions and the respondent inside it together. You can drag to rotate it, and tapping a dot opens that tradition's description. It's drawn directly on a canvas with no library.
The three axes weren't picked; they were measured. I ran all twenty ways of choosing three of the six axes and counted how many of the eight octants the 15 traditions fill. Transcendent↔Naturalist · Norms/practice↔Inner/autonomy · Certainty↔Inquiry was the most even, with six octants, and the groupings read sensibly: Islam, Judaism and Zoroastrianism in one octant; Christianity, Sikhism and the Bahá'í Faith in another; Buddhism and Taoism in another; Confucianism alone in one.
I tried principal component analysis and dropped it. The first principal component took 81% of the variance. The weights of all six axes fell between +0.34 and +0.44: effectively a single "overall religiosity." Plot on that plane and the dots line up in a row. It means the 6 axes are entangled with each other at 0.54 to 0.95. That's a limit of the model, not something to hide, so it's written right under the map.
Not 93%, but "60% chance of being first"
With the map attached, the remaining problem got clearer. The result screen shows one number, "Pantheism 93%," but nowhere does it say how solid that number is. Even when first and second are 3 points apart, it reads as settled.
I measured it. Taking the 60 items of a hypothetical respondent answering as a mix of atheism and Buddhism, I resampled with replacement and scored 200 times, and first place changed 80 times out of 200. Pantheism 60%, atheism 29%, Buddhism 6%. The 93% on screen is just one point; really it was a story of "about a 60% chance pantheism is first."
So I added a rank-stability box under the result. For the top four traditions, it shows both how often each came first and the range the match moved over. If first place holds 75% or more, it says "Almost the same even if the items were redrawn"; 45–75%, "First and second aren't clearly separated"; below that, "Your current answers can't settle the ranking." For 60 items, the whole result screen computes in 35 milliseconds.
Three of the six axes were measuring the same thing
After adding the map, I measured again how much the axes overlap. Even leaving out atheism, agnosticism and pantheism, which stand at one end of every axis, and looking only at the 12 religions, the revelation/tradition axis and the community axis correlated at 0.91, and revelation/tradition and certainty at 0.85. Different names, same measurement. I merged the three into one: Authority of tradition ↔ Individual inquiry.
Something was missing, too. Splitting the expected answers on all 428 items into principal components, the first is "revealed monotheism ↔ secular" (39%) and the second is "rebirth, karma, many gods ↔ one life" (18%). None of the six axes measured that second structure. That's why Hinduism and Jainism were bunched in the middle of the axes. I set up a new Rebirth ↔ One life axis, and its item consistency (0.91) is the highest of the five axes.
The average correlation between axes fell from 0.79 to 0.49. How well the five axes reproduce the distances between traditions across all 428 items barely moved, from 0.88 to 0.86. With the community axis gone, the Devotee of Community type could no longer be measured, so it became the Devoted Pilgrim, holding both a personal God and rebirth. People answering as Hindus or Sikhs land on this type. The map's default axes became personal God, rebirth and norms. Of the two combinations where the 15 traditions fill all eight octants, it's the one whose three axes are less entangled. The other three axes can be selected on the result screen.
What I couldn't confirm
The most important limit is this. The expected answers per tradition written for each item have not been validated against response data from actual believers. They are structured assumptions about each tradition's typical position, and differences between denominations within a tradition are lumped into a single "divided within the tradition" (0) cell. So this result is not a probability, not a diagnosis, and certainly not a ruling on which religion you belong to.
The 5 axes don't capture every tradition evenly, either. The Bahá'í Faith sits near the center on the authority-of-tradition and norms axes, so its type assignment is weak. The transcendence axis and the personal God axis still overlap heavily, correlating at 0.89 within the 12 religions. The wording direction of axis items is also lopsided in places. On the personal God axis, 18 items have "Yes" pointing toward a personal God and only 7 point the other way, so someone who habitually presses "Yes" gets pushed to one side on that axis.
And a hard-line atheist's Christianity score doesn't go below 30%. In the ethics and society areas, items remain where both sides genuinely expect the same answer. To widen it further, I'd have to force the lowest visible score down to 0%, which is a different kind of lie, so I didn't. Read 30% not as "still 30%!" but as "well below the neutral 50%."