Limitations

Every dimension in this assessment rests on published, peer-reviewed research. The assembled instrument does not. The specific questions, the weights, the type thresholds, and the forecast rules have not yet been through their own reliability and validity studies. That means:

• Your dimension scores are reasonable estimates, not calibrated measurements.
• Your type is assigned by transparent rules approximating published cluster profiles, not by matching you against a validated dataset.
• The forecast reflects the direction of published findings, but its cutoffs are provisional. Treat it as a structured conversation starter, not a measurement.

The attachment items follow the constructs of the Experiences in Close Relationships scale, and the personality items follow the Big Five, but they are original items written for this prototype. Construct-faithful is not the same as validated. A production version would license or adopt the source scales and re-check internal consistency.

The scenario reflections ask you to code your own behavior: did I criticize, did I shut down, did I accept my partner's peace offer? This is the weakest link in the design. People under-report their own worst moments, and one partner's honesty can differ from the other's. The mitigations in this version:

• Markers are phrased behaviorally and in the first person ("I used sarcasm"), which reduces vagueness though not social desirability.
• Reflections are private; your partner never sees your individual checkboxes.
• Both partners' reports are pooled, so one person's under-reporting is partially offset.

In Gottman's original research this coding was done by trained observers watching the couple, which is far more reliable. If your results surprise you in a flattering direction, be suspicious of them.

The foundational studies (Gottman and Levenson, the PAIR project, the VSA reviews) used predominantly white, middle-class American samples. There is direct evidence that context changes what counts as a harmful pattern: the same demand/withdraw dynamic that predicts trouble for wealthier couples predicts stability for poorer ones. Whether the ten dimensions, the seven types, and the forecast rules hold across cultures is an open research question. If your relationship sits inside strong cultural or religious structures the research did not sample, weigh the results accordingly.

Some things this assessment does not measure and cannot detect:

• Abuse and coercive control. A compatibility tool can actively mask an abusive dynamic: a controlled partner may score the relationship as harmonious because dissent is unsafe. If you feel afraid of your partner, monitored, isolated, or punished for disagreement, this tool is the wrong instrument. Talk to a professional or a domestic-violence service. This is the most important limitation on this page.
• Sexual compatibility. Deliberately out of scope for v0.1; it matters, and the research supports its inclusion in a future version.
• Addiction, untreated mental illness, and major life instability. These reshape everything the test measures and need direct professional attention, not a questionnaire.
• Honesty of participation. The test assumes both partners answer sincerely and respect the write-alone steps. It has no way to detect coached or performed answers.

You may have read that researchers can predict divorce with more than 90% accuracy. That figure did not survive our source verification and this tool does not claim it. The honest state of the field: the best instruments reach roughly 80% over five years, satisfaction predicts stability only modestly, some stably happy couples divorce with no visible warning, and 50-90% of couples maintain stable satisfaction over decades. Probabilities, not prophecies. Anyone who sells you certainty about your marriage is overselling.

• Paired sessions sync through an encrypted relay. In the two-phone mode your messages are end-to-end encrypted on your devices with a key that travels only inside the QR code or join code, never to any server. When the two phones can reach each other directly they talk directly; otherwise the encrypted messages pass through this site's own relay, which stores only unreadable ciphertext and deletes the session when you finish (stale sessions are swept within a day). This works on any networks, including different ones, and even far apart for part one. One honest caveat: the relay operator (us) could in principle attempt to brute-force a join code offline against stored ciphertext; the key derivation is deliberately slowed to make that expensive, and we delete sessions promptly, but treat "end-to-end encrypted" here as a strong engineering practice, not an audited guarantee. If everything fails, the one-device mode works with no connection at all; there, honesty depends on the pass-and-look-away protocol being respected.
• Contextual calibration is minimal. The socioeconomic moderation of conflict patterns is documented in the research but only lightly implemented here.
• Almost nothing is saved. Your answers live only in your browsers. One-device sessions save progress locally on that device so you can resume; paired sessions save nothing and cannot resume. Nothing reaches a server unless you explicitly opt in, at the end, to contribute anonymized scores to research validation; that donation contains dimension scores, tallies, and coarse context, and never contains names, written responses, or the private safety answers. There is still no account and no way to compare a retake against your first result, so print or save the report if you want to keep it.
• No professional in the loop. Established tools like PREPARE/ENRICH are administered with a trained facilitator who helps couples process hard findings. Here you are on your own with the report. If something in it stings, that is a signal to bring in a couples therapist, not to conclude the relationship is doomed.

• If there is fear, control, or violence in the relationship. Seek professional help instead.
• As a weapon: "the test says you are the problem." The unit of analysis is the couple, and every pattern it measures is a two-person pattern.
• As a tiebreaker for a decision you have already made emotionally. It informs judgment; it does not replace it.
• During an acute crisis (an affair discovery, a bereavement, a fresh major conflict). Results captured in a crisis reflect the crisis, not the relationship.

The validation roadmap, in order. Until these are done, the tool remains a structured, research-informed conversation, which is still valuable, but is not a measurement.

1. Content validity review by independent relationship researchers and clinicians.
2. A pilot with about 100 couples: item analysis, internal consistency, completion behavior.
3. A structure study with about 500 couples: confirmatory factor analysis of the ten dimensions, and re-deriving the couple types from our own data through latent profile analysis, accepting whatever the data says.
4. A coding study testing whether language-model analysis of the written scenario responses can match trained human coders, replacing self-coding if it can.
5. A longitudinal study following assessed couples for three to five years, testing whether the types, trajectories, and forecasts actually predict outcomes.
6. Moderation studies implementing socioeconomic and cultural calibration against outcome data.