HOW IT WORKS · SEPTEMBER 2026
The same situation.
Your own perspective.
DyadStudy helps couples explore how their selected preferences, priorities, decisions and problem-solving approaches match across everyday scenarios.
The private tool is available. Its scenario questions have undergone AI-assisted content review and revision. That review is not a human expert panel or an empirical validation study. DyadStudy is a self-assessment questionnaire, not a research study, and no research is being carried out with participants’ answers. Each partner may separately allow possible future research use, which could only begin after independent ethics approval. Read the future-research information.
Read our AI review methodology ↓ · Four-model content review
What this test does—and how to use your results
What it does: DyadStudy compares the options each partner selects independently when given the same fictional situations. The questions concern preferences, priorities, decision-making and problem-solving approaches. Your overall and sixteen facet scores summarize matching responses to those questions, under their stated conditions. They are a starting point for discussing your perspectives.
What it does not do: the questionnaire does not establish relationship compatibility, predict marriage success, separation or divorce, assess relationship health or safety, diagnose a condition, or measure actual problem-solving skill. It does not reveal either partner’s motives, honesty, intelligence or worth. AI-assisted question review is not evidence that the scores have been validated with real couples.
Read the number as response agreement
A higher score means more matching selected responses within the scoring rules; a lower score means more different selected responses. Neither is a relationship grade. There are no validated “good,” “bad,” “compatible” or “incompatible” cutoffs, and the score is not a percentile against other couples.
For example, an overall score of 80% is the equal-weight average of sixteen facet agreement scores before display rounding. It does not mean you are 80% compatible, have an 80% chance of a successful marriage, or agree on 80% of everything in life. A facet score concerns only that facet’s included questions. The overall average can conceal substantial differences between facets, so read the breakdown alongside it.
Check how much information is behind it
Read the paired-answer coverage and any skipped or “Need more context” responses alongside your scores. Those responses are excluded from matching, not treated as disagreement; both requesting context is not counted as a substantive match. If a score is unavailable because coverage is insufficient, that is missing information, not a zero score or a sign of a relationship problem. The coverage rules and treatment of repeated questions are explained below.
Turn the result into a conversation
- Start with a facet: notice where you chose alike and where you differed, rather than judging the relationship from the overall number.
- Explain your perspective: ask, “What mattered most to you?” and “Which trade-off or assumption shaped your choice?” The same answer can come from different reasons; different answers do not by themselves demonstrate conflict.
- Explore missing context: ask, “What extra information might change your answer?” Compare your interpretations without treating one option as the correct response.
- Choose a useful next conversation: discuss how you would handle a relevant real-life situation together. The test has not observed your actual behavior, so avoid using the score to label either partner or decide whether to marry, stay together or separate.
Your responses reflect the questions, available options and circumstances of answering. They are not a permanent description of either person. Similarity is not proof of a healthy relationship, and difference is not proof of an unhealthy one.
One overall score, sixteen facets
The focused module contains 24 questions on decision choices, plan evaluation, judgments and interpretations. The broader module contains 60 questions on problem framing, information use, uncertainty, planning, follow-through, flexibility, negotiation, conflict repair, support, autonomy, resource fairness and shared direction.
Both partners see identical facts and answer independently. Selecting the same substantive option counts as a match; different options count as different responses. Options are not ranked as better or worse, and their position carries no scoring value.
How we calculate matching
- Within a question: matching choices count as 1 and different choices as 0. If an anchor is repeated, its observations are averaged first so the question still receives one question weight.
- Within a facet: we average its question-level values and multiply by 100. A facet requires at least four unique questions with substantive answers from both partners, across at least three sessions.
- Across the questionnaire: the overall score averages all sixteen facet scores before rounding. Each facet contributes 1/16: the four focused facets together contribute 25%, and the twelve broader facets contribute 75%. This is an explicit design choice, not a learned measure of importance.
The overall score is shown only when every facet meets the coverage rule. We do not silently drop a facet and redistribute its weight. “Need more context” and skipped answers are reported separately and excluded from the matching denominator. Two people both requesting context are not treated as agreeing on a substantive decision.
For example, if the four focused facets are all 100% and the twelve broader facets all 0%, the overall match is 25%. An 80% score describes the resulting average across your qualifying facet scores; it does not mean you are similar on 80% of every aspect of life.
The separate focused summary remains available. Scores are raw agreement, not percentiles or comparisons with other couples. A common answer can produce a match without showing unusual similarity. No score bands label a relationship good or bad.
Four sessions, balanced forms
Each of four sessions contains 26 scenarios: eight focused and eighteen broader. The sessions open on days 0, 3, 7 and 12 after pairing. Across the sessions, each person sees 104 presentations drawn from 84 unique questions. Repeated anchors account for the difference. Brief mood, stress and interruption check-ins are separate and do not affect matching.
Form order is randomly assigned to each couple, and both partners receive the same order. Existing couples retain their original question version; newly paired couples receive version 0.7. Different versions are never pooled into one couple’s report. Read the version 0.7 question library.
How we reviewed the questions with AI
Disclosure updated 23 September 2026 · current questionnaire version 0.7. This is our AI-only question-development and review method. Its purpose is to find confusing wording, missing assumptions and overlapping response options. It does not establish human psychometric validity, compatibility, demonstrated problem-solving ability or relationship outcomes.
1. Construct coherent fictional couples
The completed v0.6 synthetic review used 16 new fictional adults in eight couples, divided into two batches. Profiles varied in age, education, work, upbringing, resources and relationship experience. Each included a life history, dated memories, shared relationship facts, private interpretations, behavioral tendencies with exceptions, and a fixed session state. Reading, numeracy and emotional habits were described functionally; IQ and EQ were not measured. Appearance and demographic labels were not treated as causes of personality.
Profile writers had not seen the questions or earlier findings. A separate coherence audit identified six construction defects, which were repaired before answering; the original versions were retained. This was a purposive Canadian, English-capable fictional cohort, not a representative sample.
2. Calibrate the personas and qualify the interviewer
Before the final review, we compared three persona formats in 48 fresh runs, with 576 answers and 192 separate factual probes. Two reviewers selected full biographies for a narrow observed improvement in individualized behavioral detail. That comparison did not measure resemblance to real people.
The interviewer persona had a fictional professional background, career history and personal traits, alongside a research-informed protocol for neutral probing. Those invented credentials were not evidence of expertise. We evaluated interviewer behavior on known problems and clean cases: whether it noticed ambiguity, asked non-leading questions, respected uncertainty and stopped appropriately. Failed qualification attempts and a missed clarification in a transfer case remain in the record.
3. Lock independent answers before interviewing
Each candidate answered all 84 questions before any interview, producing 1,344 initial answers. The answering roles were kept separate from partner answers, scoring guidance and prior review findings. We preserved the original choices and recorded hashes of inputs and outputs. These controls support record integrity; role separation relied on instructions and separate contexts, not enforced computer access boundaries.
4. Conduct actual exchanges between AI roles
Fresh interviewer actors received their assigned questions and original choices, then asked neutral follow-up questions. The original candidate actors supplied their own replies; the coordinator did not write candidate replies for them. The final cycle included 194 interview cases and 197 actual candidate replies, including one later focused clarification. Every question was interviewed in both batches, with at least two distinct couples represented per question. No candidate requested a change to an original choice.
5. Review disagreements, revise and retest
Two separately operating AI reviewers examined all 84 questions and the interview evidence before reconciling their judgments. Concerns required evidence and consideration of alternative explanations. We distinguished a wording defect from a legitimate preference, an unknown event detail, a persona inconsistency or an interviewer-induced assumption. Each question received a disposition. A supported wording change required a new version and fresh testing; earlier failures were preserved.
The preceding v0.5 round identified an unintended assumption about packing and changing accommodation in the holiday scenario. Version 0.6 added: “Neither preference requires changing accommodation during the trip.” Fresh couples then completed the review described here.
One reviewer also noticed that a candidate described moving-day helpers as “arranged” when the question only said they had been asked. A supplemental neutral interview clarified that their agreement was unknown. Both reviewers retained the question, while preserving the missed initial probe and uncertainty about the candidate’s original interpretation.
6. Report the result with its limits
Result of the v0.6 synthetic cycle: both reviewers retained all 84 questions, with no remaining supported wording revisions or unresolved actionable item dispositions within this procedure. This does not show that every possible ambiguity has been found.
- Interviews covered 119 of 420 possible question-and-option combinations. For 49 questions, only one selected option was interviewed. Full question coverage is not full option coverage.
- Most cases had one reply. Seven initial answers requested more context; no explicit skips occurred. Skip behavior and deeper response processes remain untested in this cycle.
- Generated explanations are simulated self-reports. A later clarification cannot reconstruct the original reasoning, and an unchanged choice does not prove an assumption had no effect.
- Separate AI actors shared related AI infrastructure and potentially common biases. That candidate/interviewer exercise has not been replicated across model families; the separate four-family content audit below assesses item wording. Exact provider snapshots and sampling settings were unavailable in the retained execution records; we cannot promise identical regeneration.
- Demographic variety and detailed biographies do not establish human equivalence or population representativeness. No real couples or recruited human experts participated in this AI-only review.
Our versioned project records retain the method, prompts, profiles, original answers, exchanges, failed attempts, reviewer disagreements and revision decisions. Earlier work included authored fictional reading traces and cold interviews; those development exercises are distinct from the v0.6 counts above. Related AI persona and interviewing methods already exist; we make no claim to have invented the first such method.
Four model families, one recorded content-review process
Completed initial content audit: OpenAI, Anthropic Claude, Google Gemini and xAI Grok each reviewed all 84 questions in a fresh context, producing 336 item judgments. Each received the same frozen question text and review criteria without the other models’ findings. This was an item-content audit, distinct from the fictional-couple interviews above.
- OpenAI: 84 items reviewed; three proposed revisions.
- Claude: 84 items reviewed; six proposed revisions.
- Gemini: 84 items reviewed; no proposed revisions.
- Grok Expert: 84 items reviewed; two proposed revisions.
The eleven findings concerned eight distinct questions; three models independently flagged the same undefined bus-route reference. Every finding required an exact quotation, a concrete alternative reading and a proposed minimal correction. We checked all item IDs and quotations against the frozen source.
Two fresh AI adjudicators assessed every finding with provider labels withheld. They agreed on eight findings and disagreed on three. The coordinator recorded reasons for each final disposition, preserving both original judgments. Version 0.7 corrects six questions: technician and route references, market-stall coordination wording, the lunch-budget trade-off, dinner timing and option overlap, and whom to address after an accidental breakage. Suggestions to impose an additional reading-time preference or require a settled project role were declined because the questions intentionally allow uncertainty and ask for an initial response. Complete revised items were submitted to all four families for fresh targeted review.
Completed corrections and retesting: four targeted rounds covered six, two, two and one questions per family, adding 44 judgments for a total of 380. All 24 proposed finding records across the initial audit and retests received explicit dispositions. Six questions changed; no supported correction remains open. The final record preserves a disagreement about whether deciding today permits choosing to wait. Completion does not require unanimous model approval. These rounds developed the wording; they were not untouched confirmatory tests.
What makes the method useful
- Systematic coverage: an explicit disposition for every question in each family’s initial review makes omissions visible.
- Traceable corrections: frozen inputs, original outputs, exact evidence quotations, adjudication and before-and-after records connect changes to their reasons.
- Multiple perspectives: separate model-family reviews expose both recurring concerns and disagreements. Agreement is evidence to inspect, not a vote that establishes truth.
- Repeatable procedures: recorded prompts and acceptance rules allow the procedure to be rerun. Provider changes and sampling mean identical outputs are not guaranteed.
- Preserved failures: unsuccessful interviewer tests and rejected suggestions remain in the record instead of disappearing from the final account.
Interviewer testing is a separate result
Earlier cross-model trials tested whether an AI interviewer followed neutral-probing and stopping rules. No family achieved two passing runs under the revised unrestricted qualification. A constrained workflow subsequently accepted two Claude assisted runs; Gemini and Grok Fast had structural failures. In two paid Grok Expert runs, the first was structurally clean and the second contained one invalid evidence reference, so that two-run qualification did not pass. Those results do not prevent a model from contributing a separately checked content review, and are not counted as successful interviews.
The full planned cross-model candidate/interviewer matrix has not been completed. The four-family audit does not establish that synthetic people reproduce human understanding, that scores measure real-world skills, or that the method outperforms human validation. Different model families may share blind spots. We make no historical-priority claim.
Read the complete audit, dispositions and evidence →
Published work supports exploring AI pretesting: a 2026 survey-pretesting working paper found useful defect detection with model- and prompt-dependent errors, while four preregistered psychometric simulation studies found both structural similarities and important differences from human responses. These studies inform our checks; they do not validate DyadStudy.
Research that informed the design
The project’s literature review considered work on situation perception, values, decision styles and dyadic processes. Findings from those studies informed the topics; they do not transfer validity to newly authored scenarios. The repository preserves its evidence map, source-reading limits, instrument permissions and historical research plans.
- Weidmann and colleagues: personality similarity in couples
- Rentzsch and colleagues: similarity in situation perception
- AERA, APA and NCME: interpretations and intended uses of scores
- COSMIN content-review methodology — principles adapted from health measurement, not a certification of this tool
Privacy and optional follow-ups
Each partner controls their own account and answers. A shared descriptive report requires both partners to finish and authorize sharing. Optional private follow-ups and calendar downloads are available after the fourth session with separate consent; they do not contribute to your match score or authorize research reuse. Read the privacy notice.
Any future formal research would be separate from this questionnaire and would first need independent ethics approval, its own permissions and consent. The current product makes no claim to establish compatibility or predict marriage success.
Start or return to your reflection →