How we read you

What WETI is not. And why that matters.

Why culture, not just personality?

WETI maps the cultural wiring you both brought into the relationship — before it maps as conflict.

Personality tests tell you who you are. Attachment tests tell you how you attach. Neither maps the cultural layer that sits between two people who grew up in different worlds. Hofstede, Hall, Trompenaars, and Schwartz — together more than fifty years of cross-cultural fieldwork — all point to the same finding: what breaks a relationship is often not that two people are wrong for each other, but that two cultural grammars have been running in parallel with no one ever comparing notes. WETI is built specifically for that gap. It profiles both partners on the dimensions where cultural background actually diverges, and gives them a shared vocabulary before those differences become resentment. The framework names and numeric scores never reach the prose itself — only the behavior they translate into.

The five patterns WETI maps in cross-cultural relationships

These are not traits of any nationality or ethnic group. They are communication defaults that people raised in particular cultural contexts tend to internalize. They are where WETI's six measured axes land in real relationships.

1 · High-Low Context

Hall 1976

People raised in high-context cultural environments tend to carry meaning in relationship, tone, and timing rather than explicit words. People raised in low-context environments tend to trust the literal surface of what is said. When these two defaults meet in the same relationship, one person reads the other as evasive, the other reads their partner as blunt. Neither is wrong. They are on different signal frequencies. WETI's V axis (Voice) places you on this dimension.

2 · Family Embeddedness

Hofstede 1980 · Trompenaars 1997

People raised in collectively oriented cultural environments often experience a relationship as a connection between two family systems, not just two individuals. People raised in individually autonomous environments often treat it as a private matter between two adults. The question of whose parents come first, or whether to live near family, never gets discussed explicitly because each partner assumes their answer is common sense. WETI's S axis (Self) measures this default.

3 · Emotional Expression

Trompenaars 1997 · Gottman 1994

People raised in emotionally neutral cultural environments often read restraint as respect. People raised in affectively expressive environments often read restraint as distance or withdrawal. The person who goes quiet is not punishing anyone. They are handling the situation in the way that feels normal to them. What feels normal was learned. WETI's H axis (Heart) and the Gottman conflict signal together measure tension here.

4 · Financial Transparency

Hofstede 1980 · Schwartz 1992

People raised in cultural environments where family financial pooling is the default often treat financial openness as a form of trust. People raised in environments where financial independence is the default often treat a partner's question about income as a boundary violation. How much goes to family each month, whether to merge finances at all, these conversations never happen because both partners think the other already knows the script. WETI's B axis (Boundary) measures where you sit here.

5 · Time Orientation

Hall 1976 · Hofstede 1980

People raised in long-term oriented cultural environments often read a clear five-year plan as a sign of seriousness. People raised in present-oriented or polychronic environments often read “let's see” as freedom, not avoidance. “We'll figure it out” is a promise to one partner and a deflection to the other. WETI's P axis (Pace) places you on this dimension.

The same math, applied to different objects

WETI publishes three numbers across its products. They look different on the surface · a 5-band label on the personal-deep matrix, a precise integer on the pair report, a 0–100 dynamics index on the trio. They are not three different methods. They are the same canonical calculation applied to three different objects.

The shared handle · cosine on 6 axes

Every test result is a vector of six numbers · V (Voice) · H (Heart) · S (Self) · P (Pace) · W (Warmth) · B (Boundary) · each in the 0–10 range. The angle between any two such vectors is what cosine similarity measures. When the angle is small the vectors point the same way, the similarity is near +1 · when the angle is wide the similarity drops toward 0 · in the rare anti-aligned case it can go negative.

The same √-stretch then maps cosine to a 0–100 display percent (preserves the −1 / 0 / +1 pivots, pulls real-human mid-cosines toward the ends so a 90 actually feels rare). Every score below starts here.

Why a band on the personal-deep matrix but a precise integer on the pair · the pair score is a measurement between two specific people who both took the test. The personal-deep matrix row is a measurement between you (specific) and a centroid (an abstraction). Showing a precise integer against an abstraction over-promises · showing a band tells the truth · the score lives in a range and a specific person of that type drifts within it. Data precision must match display precision. The math is the same · the typography is not.

Score interpretation · five tiers

One vocabulary across all three products. A 90 on the pair, a 90 on the trio, and the “usually clicks” band on the matrix all mean the same thing structurally · they sit in the same band of the same calculation.

The matrix bands round to the same five tiers · ≥70 reads as “usually clicks” (strong/rare) · 55–69 as “mostly gets close” (upper mid) · 40–54 as “shared frequency” (lower mid) · 25–39 as “prone to drift past” (strained) · <25 as “rarely lands” (heavy). One vocabulary · three products.

When two people aren't the whole picture

A pair reading maps the friction between two people. But some of the most important relationship dynamics only become visible when a third person enters the picture — the partner's parent, a third person in a love triangle, a team triangle at work, a best friend and a new partner who don't quite fit together. When there are three people, you are no longer dealing with a single edge; you are dealing with a topology — three edges, one triangle, and a set of structural roles that pair readings cannot name.

Triangle Dynamics Index · harmonic mean

The obvious way to score a trio is to average the three pair scores. That would be wrong.

Averaging treats a trio of (90, 80, 40) as equivalent to (70, 70, 70). It is not. In the first trio, one pair is cold and that coldness is felt every time the three people are in the same room. The relationship's weakest edge pulls the whole network down in ways that a simple average hides.

WETI uses the harmonic mean instead. The harmonic mean gives the weakest edge a disproportionate weight — which is exactly how tension works in a real group. One cold pairing drags the whole reading down. A fully balanced trio stays balanced.

Bridge member · weakest edge

Social network analysis identifies two structural roles that recur across almost every three-person group, regardless of domain (families, friendships, workplaces). WETI names both.

5 triangle shapes

Every trio is classified into one of five deterministic shapes, derived from the three pair scores and the bridge / weakest-edge calculation. Priority order when shapes overlap: Unbalanced > Bridge > Mirror > Aligned > Edged.

What a trio reading contains

A trio report has four chapters, mirroring the pair report's structure:

• ✦§I · Where the Three Came From · Who this trio is as a unit · the shape it takes · the trio relationship type name
• ✦§II · Three Conversations · Each of the three pairings read separately · where each pair strengthens or strains the group
• ✦§III · The Bridge and the Tension · Bridge member named · weakest edge named · the hidden contract that holds or breaks the group
• ✦§IV · The Triangle in Six Months · Where the current topology is likely to move given each person's individual scores

A trio report is generated from the three sets of answers as they exist at the time of generation. If any one person retakes the test, the trio report does not automatically update — consistent with pair report behaviour.

The seven frameworks behind the reading

Hofstede's Cultural Dimensions (Hofstede, 1980)

Hofstede's original work analyzed survey responses from roughly 117,000 IBM employees across more than 50 countries, gathered between 1967 and 1973, and produced a six-dimension model of national culture (individualism, power distance, uncertainty avoidance, masculinity, long-term orientation, indulgence). It remains the most-cited dataset in cross-cultural psychology. WETI's S axis (Self: collective vs independent) maps directly onto the individualism-collectivism dimension, which is the dimension Hofstede's data spoke loudest on. All six numbers for 15 seed countries are stored in our country baseline table and injected into the report prompt by name.

High- and low-context communication (Hall, 1976)

Edward T. Hall, the anthropologist who effectively founded modern intercultural studies, argued in Beyond Culture (1976) that societies differ in how much meaning is carried by explicit words versus implicit context, and in whether time is treated as a single line of appointments (monochronic) or a fabric of overlapping events (polychronic). Hall's work was qualitative, drawn from decades of field observation rather than survey data. WETI's V axis (Voice: implicit vs direct) and P axis (Pace: planned vs improvised) take their conceptual structure from these two distinctions. We carry a per-country high / mid / low context band in the baseline table; the monochronic / polychronic read is left to the model as Hall framed it (qualitative, not numeric).

Trompenaars' seven dimensions (Trompenaars & Hampden-Turner, 1997)

Building on Hofstede and Parsons, Fons Trompenaars and Charles Hampden-Turner surveyed roughly 15,000 managers across about 50 countries and proposed seven bipolar dimensions of culture: universalism vs particularism (rules vs relationships), individualism vs communitarianism, specific vs diffuse (public and private separated or blurred), neutral vs affective (emotion held in or shown), achievement vs ascription (status by what you do vs who you are), sequential vs synchronic time, and internal vs external direction (shape the environment vs adapt to it). WETI's H axis (Heart: reserved vs engaged) draws from neutral vs affective, and the B axis (Boundary: open vs guarded) borrows the specific vs diffuse logic almost line for line. All seven dimensions are now held as per-country numbers (0-100, midpoint 50) for the same 15 seed countries as Hofstede. At report time we surface each country's top three most-distinctive dimensions (distance from midpoint, not absolute score) into the prompt as a behavior-rhythm calibration anchor underneath Hofstede + Hall + Schwartz.

The Big Five personality factors (McCrae & Costa, 1987)

The Big Five (openness, conscientiousness, extraversion, agreeableness, neuroticism) is the most-replicated factor structure in personality psychology, with cross-cultural support across more than 50 countries. We name it for honesty about lineage, not as a per-country lookup. The W axis (Warmth: slow-warm vs quick-warm) takes its conceptual structure from the warmth and gregariousness facets of Big Five extraversion (Costa & McCrae, 1992) — the facets where the literature is densest and the items translate most cleanly across languages. Big Five is an individual psychometric framework, not a country-level one; it has no normative country baseline and we do not invent one. Treat this as the W axis design lineage, not a calibration input.

Sternberg's triangular theory of love (Sternberg, 1986)

Robert Sternberg's triangular theory proposed that love rests on three relatively independent components, intimacy (closeness and connection), passion (drive and arousal), and commitment (the decision to maintain the relationship), and that different mixes of the three produce qualitatively different love kinds (consummate, companionate, empty, romantic, and so on). Our six love-type r-types (r9 through r14) each carry a locked (intimacy / passion / commitment) fingerprint sitting on real theoretical ground, not improvised by the language model. When a pair's match lands in the love category, the three numbers feed directly into the Layer 1 prompt so the relationship portrait grounds its specifics in this pair's particular triangle rather than a generic “love relationship” image. The framework name and the words intimacy / passion / commitment never reach the prose itself, only their translated imagery.

Schwartz basic values (Schwartz, 1992)

Shalom Schwartz's theory of basic values is the most-replicated cross-cultural framework on what people prioritise in life. It posits ten universal values (self-direction, stimulation, hedonism, achievement, power, security, conformity, tradition, benevolence, universalism) and finds that what differs across cultures is the relative priority of these ten, not whether the values exist. We run Schwartz on TWO layers, country and individual. Country layer: each of our 15 seed countries carries all ten values on a 0-10 scale, calibrated against Schwartz 2012 + European Social Survey + World Values Survey 7 country aggregates; getTopValues() picks each country's top three at report time. Individual layer: a five-item Portrait Values Questionnaire (Schwartz et al. 2001 short-form thinking) sits at the tail of every session, covering all four Schwartz value clusters (openness to change, self-enhancement, conservation, self-transcendence); computeIndividualValues() returns the reader's own top three plus the largest deviations against their country average in SD units. Both layers feed into the pair-report and personal-deep prompts. The load-bearing angle is NOT “the country top is X” but “this reader sits WHERE inside their country”. A self-direction-high reader inside the China baseline gets “the friction with your traditional family is structural, not a personality issue,” not a generic CN read. The framework name, the value-key strings, and the numeric scores never reach the prose itself, only the behavior they translate into. Verdicts like “you do not belong here” / “you are not really Chinese” are hard-banned no matter how large the deviation.

Gottman's four horsemen of the apocalypse (Gottman, 1994 · v2 dual-signal)

John Gottman's longitudinal research on married couples identified four interaction patterns that most strongly predict relationship dissolution, criticism, contempt, defensiveness, and stonewalling, with contempt being the single strongest predictor. We use two layers of signal to predict which one or two horsemen a given pair is most likely to summon: (1) baseline / indirect, the six axes plus Bartholomew anxiety / avoidance (the v1 formula); and (2) self-report / direct, the CS1 item that asks the reader to name their own default first move when they disagree with someone close, mapped onto one of four clusters (direct / mediator / avoidant / escalator). The two signals run independently in predictHorsemen() and converge into a confidence label, both signals agreeing on the same horseman returns “high”, single-signal returns “medium”, and both weak filters the horseman out entirely. The pair-report Layer 4 forecast paragraph runs vivid and specific at high confidence and runs restrained at medium, using the reader's self-reported cluster as the scene anchor (the door-click moment, the Sunday-morning sock complaint, or a reveal line when self-report and behavior disagree). The framework name and the four English terms never reach the prose, only the behavior they describe.

Why two signals: v1 used baseline alone, so a low-Warmth + high-anxiety profile pushed toward criticism even if the reader is actually a silent-exit type. CS1 catches that disagreement and lets the forecast land on the right scene. In Gottman's own observational work the combination “behavior + self-report” outperforms either signal alone, which is the same logic powering v2's jump in usable accuracy (~30% over v1 by inspection).

sim(u, c_k) = (u · c_k) / (||u|| · ||c_k||)
type(u)     = argmax_k  sim(u, c_k)

The six axes, and where each comes from

Why this many questions

The session length was a deliberate choice, not a round-number convenience. Three reasons. First, Krosnick and Presser (2010) and the broader survey-methods literature show that test-taking quality starts to deteriorate noticeably past the five-minute mark, with respondents satisficing rather than discriminating. Second, each session gives every axis enough weighted items to land somewhere other than dead-center. Third, a small number of items are reverse- coded so we can detect acquiescence bias (the tendency to agree with whatever is on screen), and silently down-weight users whose pattern looks like autopilot. The v3 expansion to a 32-item session draws from a 72-item bank via stratified sampling: every axis sub-pool is sampled in proportion, so the 32 items a given reader sees still cover the full instrument even though no two sittings look identical.

How the 12 types were set

We need to be honest here. The 12 archetypes are not the output of unsupervised clustering on real user data, and we do not claim they are. They are 12 hand-set centroid points in 6-D space, chosen to span the most behaviorally distinct corners of the space without leaving large empty regions. Each user's 6-D vector is matched to the nearest centroid via cosine similarity. As the user base grows past the threshold where unsupervised clustering becomes statistically honest (we estimate 30,000 to 50,000 completed assessments per locale), these centroids will be re-fit empirically and the type definitions re-published.

How the matrix is computed

Pair compatibility is rank-based, not threshold-based. After two users' vectors are compared, the raw similarity score is bucketed into one of five bands by its position in the live distribution (quintiles), so every user gets a balanced spread of “high chemistry,” “tension,” and so on across the type set. We chose this over fixed thresholds because users with polarized profiles (someone extreme on three or more axes) would otherwise see almost everyone as a “reverse-magnet.” That reading is technically true and practically useless. Rank-bucketing keeps the result interpretable.

The limits

• The user base is still small. Norms shown on the result page (for example, "you are calmer than 72% of users") are computed by combining a Bayesian prior with the live distribution. The prior is weighted heavily until N grows large enough to speak for itself. If you take the test today, your percentile is partly informed by what we expected, not only by who else has answered.
• Self-report has known biases. Mood-of-the-day, social desirability, and the way the question is phrased all leak into the answer. We do not statistically correct for these. We considered building a desirability-correction layer and decided the false confidence it would project was worse than the bias itself.
• The 12 types are categorical for clarity, but the data underneath is continuous. Edge cases will read as "kind of like X, kind of like Y." That is not a bug. If your nearest two centroids are within 3% of each other on cosine similarity, the honest answer is that you sit between them, and we say so on the result page.
• Cross-cultural calibration is a hypothesis, not a settled fact. Hofstede in particular has been challenged on what McSweeney (2002) called the ecological fallacy: country-level averages do not safely describe individuals from that country. We treat each of the seven frameworks as one input among several for exactly this reason. If any single one of them turns out to be wrong about a given dimension, the others still hold.
• Compatibility is descriptive, not prescriptive. A “low chemistry” reading does not mean a relationship cannot work, and a “high chemistry” reading does not mean it will. The matrix describes the friction surface, not the outcome. Outcomes depend on things this instrument does not, and probably cannot, measure.