Between Us: The Psychology and Mathematics Behind My Relationship Game

Abstract

Two people in the same relationship can live in surprisingly different versions of it. Between Us asks each partner, separately, which kind of support they prefer in four everyday moments and how understood, accepted and cared for they have felt, and asks each to predict the other’s answers. Comparing the two sets shows where preferences align, how accurately each partner anticipates the other, and whether either over- or underestimates how supported the other feels. These comparisons sit beside each person’s relationship-specific attachment profile, measured with the published ECR-RS questionnaire (Fraley et al., 2011). The game items are new and not validated; all scoring is deterministic. The ten couple archetypes, built from each partner’s two perspectives, are a storytelling device, not a psychological typology, and the combined report is computed on request and never stored.

Keywords: adult attachment, ECR-RS, perceived partner responsiveness, interpersonal perception, projection

How to read this article

Every statement in the game and in this article carries one of five evidence tiers, shown below. Each formula is followed by a short summary in plain words, and a single worked example (two players I call Riya and Sam) runs through Sections 4 to 6, so every number can be checked by hand.

1. Motivation

Partners are each other’s most practised observers, but practice isn’t accuracy. People often assume a partner wants what they themselves would want, a tendency research on partner perception separates from genuine accuracy (Kenny & Acitelli, 2001). A questionnaire answered by one person can only describe that person. Answered by both, with each also predicting the other, it produces small, checkable comparisons: did you both choose comfort after a bad day, did you each know what the other would choose, does your partner feel as understood as you think?

2. Scientific background

Attachment, in two dimensions

Adult attachment research describes how people experience closeness with important others along two continuous dimensions (Brennan, Clark & Shaver, 1998). Attachment anxiety is worry about whether the other person will be there and will care; attachment avoidance is discomfort with opening up to and depending on them. People who are secure tend to score low on both. Both are positions on a scale, so I never convert scores into style labels such as “anxious type”.

Relationship-specific attachment

People can feel differently in different relationships. The ECR-RS uses one short set of statements to assess attachment with one specific person. In its development sample of 21,838 adults in dating or marital relationships, the two romantic-partner scales were internally consistent (Cronbach’s α = .91 for anxiety, .87 for avoidance; values above .80 are conventionally considered good), and how anxious or avoidant someone felt with a partner was only modestly related to how they felt with parents or friends (Fraley et al., 2011, Study 1).

Intimacy as an interpersonal process

In Reis and Shaver’s (1988) model, intimacy grows when one person shares something personal and experiences the partner’s reply as understanding, validating and caring. That experience is called perceived partner responsiveness. Laurenceau, Feldman Barrett and Pietromonaco (1998) tested the model with a diary: after every social interaction of ten minutes or longer, participants completed a version of the Rochester Interaction Record (Reis & Wheeler, 1991), for one week (Study 1, 69 participants) or two weeks (Study 2, 89 participants), on 5-point scales. Sharing about oneself and the partner’s sharing both predicted how intimate an interaction felt, and feeling responded to explained part of those links, a pattern called partial mediation (tier P). In Study 2, which measured responsiveness as feeling understood, cared for and accepted, sharing emotions predicted intimacy while sharing facts did not (tier P).

What I take, and what I don’t

I take the constructs: feeling understood, accepted and cared for, and the role of disclosure. I do not take the method. The original studies sampled real interactions as they happened, with U.S. college students, from one person’s perspective; the authors note their data were correlational and did not include both partners. I ask about hypothetical scenarios and a two-week look back, and I compare both partners. My items are therefore inspired by this research, not validated by it.

3. Measuring attachment with the ECR-RS

Instructions, as in the original: “Please answer the following questions about your dating or marital partner.” Each statement is rated from 1 (strongly disagree) to 7 (strongly agree). The nine statements below are shown in the order of the paper’s Table 1, word for word. A tenth statement in that table, “I don’t fully trust this person,” related to both scales at once and was dropped by the authors; I do not use it. Item 8 keeps the original “him or her”; changing it would alter a validated item, so I note the limitation instead.

Items 1–4 describe closeness, so agreeing with them means less avoidance. Their answers are flipped before averaging: an answer $x_\ell$ on item $\ell$ counts as $8-x_\ell$, turning 7 into 1 and 1 into 7. Avoidance is the mean of items 1–6 after that flip; anxiety is the mean of items 7–9:

In plain words. Both scores run from 1 to 7. Higher anxiety means more worry about being cared for; higher avoidance means less comfort leaning on this person.

A note on direction: the paper’s text says items 5 and 6 “were reverse keyed”, describing them relative to a factor whose high end is closeness (items 1–4). Because the paper reports avoidance means where higher means more avoidance (2.47 with a partner, on the 1–7 scale), items 1–4 are the ones flipped here. This matches the ECR-R items they were drawn from (Fraley, Waller & Brennan, 2000).

Scores are stored as exact integer sums, versioned, timestamped and tied to one relationship. Retaking replaces the earlier answers; older answers are not kept.

Descriptive bands (a product convention)

To choose wording, the solo report places each score in one of five bands anchored to the response scale rather than to population norms. The three-way split in the right-hand column is also what the combined report’s closeness compass uses to describe each person; in integer sums it is anxiety ≤ 10 or ≥ 14, and avoidance ≤ 21 or ≥ 27. Because anxiety averages three answers and avoidance six, the means that can actually occur are steps of 1/3 and 1/6, so the middle band holds 3.67 to 4.33.

These cut-offs are not empirically established. “Leaning lower” means the person, on average, disagreed with the statements; it is not a percentile. For context only, the development sample averaged 3.25 (SD 1.98) for anxiety and 2.47 (SD 1.31) for avoidance with a partner, in a web sample that was 81.5% women. Two limitations noted by the authors apply here: all three anxiety items point the same way, so they may partly pick up a general tendency to agree; and, like most self-report measures, the scale tells less secure people apart better than very secure ones.

4. My new game: “Do you speak the same emotional language?”

Newly developed, exploratory, not validated. Each partner, alone, chooses in four moments what they would prefer and what they think their partner would prefer; then rates three statements about the last two weeks on a five-point scale shown as faces, and predicts how the partner would rate the same statements about them. Every rating must be answered. Choices are categories: the options have no order or numerical distance. Option order is shuffled for each player, and kept identical between their own answer and their guess, to reduce order effects.

The three statements are “I feel that my partner understands my feelings”, “I feel accepted when I share something personal with my partner” and “I feel cared for when I turn to my partner for support”. They are single new items adapted from the constructs in Laurenceau et al. (1998), whose Study 2 asked, after each interaction, how much the partner understood you, how much you felt cared for, and whether the other person saw you as acceptable.

A worked example

Riya and Sam each played alone. Their choices and guesses were:

On the three statements, Riya rated her own last two weeks 4, 3 and 5. Sam rated his 4, 5 and 3. Riya guessed Sam would say 5, 5 and 5.

5. Evaluation methodology

Four measures come out of the game. The first two use the four moments; the last two use the three statements. None of them is combined into an overall score, because there is no honest way to weigh them against each other.

5.1 Preference alignment

In plain words. In how many of the four moments did the two of you want the same thing?

Each moment adds 1 when both partners chose the same option. Riya and Sam matched on the vulnerable moment and the moment of closeness, so $m=2$. Alignment describes similarity, not compatibility: couples who want different things can understand each other perfectly well.

5.2 Directional understanding

In plain words. How many of your partner’s choices did you predict? It is counted separately in each direction, because reading someone well and being easy to read are different things.

Riya guessed Sam right in three moments ($k_{A\to B}=3$); Sam guessed Riya right in two ($k_{B\to A}=2$). The two numbers are never averaged. Their difference, the asymmetry$\;\Delta=k_{A\to B}-k_{B\to A}$, is 1 here; it reflects these four questions, not love or effort.

A correct guess can come from knowing someone or from assuming they are like you. To tell these apart I split the correct guesses by whether the partner’s choice differed from the guesser’s own:

I also count how often each person guessed their own answer for the other, a simple measure of projection:

In plain words. $k^{\text{diff}}$ counts the guesses that projection could never produce, because they name something the guesser would not have chosen. $p$ counts how often someone assumed “they’ll want what I want”.

Two of Riya’s three hits (the bad day and the unspoken feeling) were preferences unlike her own, so$\;k^{\text{diff}}_{A\to B}=2$. Sam guessed his own choice in all four moments ($p_B=4$): both of his hits came where the two already agreed ($k^{\text{diff}}_{B\to A}=0$). The combined report shows these counts for both partners in its second chapter (Section 6.1).

5.3 Perceived responsiveness

In plain words. How understood, accepted and cared for you have felt by your partner over the last two weeks, in your own words.

Each person rates the three statements in $T$ from 1 to 5; $s_i[t]$ is person $i$’s rating of statement $t$. Their perceived responsiveness is the mean of their three ratings:

Riya rated 4, 3 and 5, so $R_A=4.00$. An average can hide a mixed picture, so I also compute the spread,$\;\max_t s_i[t]-\min_t s_i[t]$. The report always shows all three ratings, so a mixed picture stays visible. Partners see each other’s ratings only after both agree to the combined report (Section 11).

The three items were written for the game. Validated multi-item measures of perceived responsiveness exist, such as the Perceived Responsiveness and Insensitivity scale (Crasta et al., 2021); mine are shorter and unvalidated, so $R_i$ is a descriptive summary, not a scale score.

5.4 Responsiveness calibration

In plain words. How well can you tell how supported your partner has felt? It compares your guesses about their three ratings with the ratings they actually gave.

For each statement, the error is the guess minus the real answer:

A positive error means you pictured your partner feeling better supported than they said; a negative error means worse. Two summaries are then taken:

In plain words. Bias is the average error with its sign kept. It answers “do my guesses lean too rosy or too gloomy?” The mean absolute error (MAE) is the average size of the errors with the sign removed. It answers “how far off am I, in either direction?”

Both are needed because they fail in different ways. A guesser who is two points too high on one statement and two points too low on another has a bias of 0, which looks perfect, but an MAE of 1.33, which shows they are not. A guesser who is exactly one point too high on everything has an MAE of 1 but a bias of +1, which reveals a consistent tilt. This mirrors the distinction in partner-perception research between being accurate and being biased in a consistent direction (Kenny & Acitelli, 2001).

Riya guessed 5, 5 and 5; Sam said 4, 5 and 3. Her errors are +1, 0 and +2, so her bias is +1.00 and her MAE is 1.00: she pictures Sam feeling somewhat better supported than he reports. Figure 6 lets you change her guesses.

For the label, the game avoids rounding by using integer sums: the total error $E=\sum_t e_t$ and the total size $\sum_t |e_t|$, with $n=3$. The rules are applied in this order:

Riya’s total error is $E=3=n$, so her guesses are classed as over. In the fourth chapter both partners see each person's three ratings beside the other's guesses, with the bias and MAE available in a collapsed panel; the class itself is used only inside the interpretation layer (Section 6.2).

5.5 Why the wording is cautious

With four questions, a correct guess can be luck. If guesses were independent and each landed with probability$\;p$, the chance that one guesser gets at least three of four right (a count$k_{i\to j}\ge3$) would be

In plain words. A stranger guessing at random would score three or four about 5% of the time. But if one option is popular enough that guessing it is right half the time, a stranger who always picks it scores three or four almost a third of the time.

So exact counts are always shown, and the report never presents a high count as proof of anything.

6. The combined report

In plain words. Knowing someone, feeling understood by them, and wanting the same things are three different experiences. The combined report measures each of them for both partners, then tells the story of where they meet.

The report is worked out from both partners' saved answers each time it is opened, using only the measures in Sections 3 and 5. It is never stored. Each chapter shows its measures directly, in plain words, before any interpretation.

6.1 What each chapter measures

Chapter 2 also compares each person with someone who always guesses their own choice. That person would score exactly $m$, so

In plain words. counts how many more of the partner's choices someone predicted than that baseline. Riya: $G_A=3-2=1$; Sam: $G_B=2-2=0$. It is a rough, exploratory number, not a measure of empathy, and the report says so.

Prediction accuracy measures recognition of stated preferences in four hypothetical moments. It does not measure whether anyone actually gave support, was a good partner, or met the other's needs, and the report never describes it that way (Kenny & Acitelli, 2001; Maisel & Gable, 2009).

6.2 From measures to archetypes and stories

What I can say about it:

• It is deterministic: the same answers always produce the same report.
• It uses only the measures in Sections 3 and 5, and nothing about you that you didn't answer.
• Every statement is shown only when the numbers support it, and a missing answer never produces an inferred one. A person with a missing rating, possible only in an older version of the game, gets no archetype rather than a forced one.
• A partner's attachment answers are used only if that partner chose to share them.
• Both partners always get the same archetype; each reads their own side of it.
• Its categories and cut-offs are product conventions, not validated psychological boundaries.

Archetypes are product narratives, not psychological types. None of the cited papers validates them, and the fictional couples beside each archetype are editorial fun. The ten archetypes are: Certified Soulmates, Exquisite Sweethearts, Love Theorists, Opposite Ends, Chaotic Cuties, Plot Twisters, Love Improvisers, Emotional Scholars, Detour Darlings, Escape Artists. On screen, Certified Soulmates also takes turns as Power Couple, Golden Duo, Dream Team, Dynamic Duo.

6.3 Access

Everything unlocks for both partners together: the archetype with coins after both have signed in, or the whole report with a single payment; the four chapters after it together, with coins or a payment. Reports are recomputed from the saved answers each time they are opened, with the current scoring and content versions. An unlock records the versions it was bought under; any change to scoring or copy gets a new version number and an entry in the revision history below.

7. The solo report

Before a partner plays, each player reads a report about their own answers. It uses no partner data, so it cannot say anything about the partner except what the player guessed, and it always calls those guesses guesses.

The report opens with a reading of the two attachment scores, placed on a two-axis chart with the wording bands from Section 3, and interprets each of the four choices. It then draws on a library of cross-answer insights, each tied to an exact combination of answers, for example “time alone after a bad day” together with “being understood when you open up”. An insight is considered only when every answer it rests on is present; it is never inferred from a missing answer. If none applies, the report says so rather than inventing a pattern. How insights are chosen and ranked belongs to the same unpublished interpretation layer as Section 6.2. Identical answers always produce an identical report.

8. Attachment as context, not cause

Attachment scores never change a metric or an archetype. They are used to (1) show each person their own two positions, and their partner’s only if that partner opts in; (2) choose a framing in the solo report when the scores lean clearly one way and the answers contain matching evidence; (3) in the compass chapter, note when an answer sits alongside a score, for example higher avoidance and time alone chosen, with the words “it doesn’t explain it”; and (4) suggest a conversation about reassurance or closeness only when both partners have shared their profiles. Language describes fit, never cause: the report never says that someone’s anxiety makes their partner misunderstand them. Direct answers always override expectations: if someone with higher avoidance chooses talking-oriented options in at least two moments, avoidance-based framing is switched off. Laurenceau et al. (1998) also found that attachment style produced few and inconsistent differences in the intimacy process, a further reason to keep attachment as context only.

9. Interpretation and limitations

Between Us can describe how your answers to these specific items compare, how accurately each of you predicted the other, how supported each of you reported feeling over two weeks, and your own attachment scores for this relationship right now. It cannot tell you whether your relationship is healthy, will last or is “compatible”, who loves more, whether anyone is a good or bad partner, whether you should stay together, why any difference or attachment score exists, or anything about anyone else. Predicting a partner’s stated preference is not the same as giving them support, and the archetype thresholds and names are conventions, not findings. It uses hypothetical scenarios, single items, no norms for the new items, a self-selected audience, and has unmeasured effects on the people who read it. Between Us is a research-inspired reflection game, not a clinical assessment. Its results describe patterns in your answers, not permanent truths about your relationship. If a relationship involves fear or control, real support matters more than a quiz.

10. Validation roadmap

• Cognitive interviews with diverse couples on clarity, distinct options and the guess prompt.
• Response distributions and dominant options, from research-consented data only.
• Test–retest reliability after two weeks (κ for choices, ICC for ratings).
• Chance-corrected accuracy against a “popular answer” baseline and randomly paired non-partners, replacing fixed thresholds in a new interpretation version.
• Construct validity against established multi-item measures of perceived responsiveness and relationship satisfaction.
• Dyadic accuracy-and-bias models, repeated administration, independently measured outcomes, cultural generalisability, and whether reading the report helps, harms or does nothing.

No game data is used for research today, and none is sold. If it is ever used for a research paper, it will be anonymised first, with all personal information removed, and used only for that research.

11. Privacy and ethics

Answers stay in the player’s browser until they choose to save. Playing, reading the solo report and sending an invite need no sign-in: saving creates a coinless guest account on that device. Opening the combined report requires signing in, and signing in moves the guest’s answers onto the account. The combined report is never stored; it is recomputed from both partners’ saved answers whenever one of them opens it.

The combined report shows each partner the other’s choices, guesses and ratings, so it appears only after both have finished and both have agreed to that (consent version 3; agreements given under the earlier, outcome-only wording must be given again). Attachment scores are shared only by a separate, revocable opt-in. Retaking replaces a person’s answers; older answers are not kept. An invite link is a key to join, never to read. Invite tokens travel in the link’s fragment so they never reach server logs, and only their hashes are stored. Declining needs no reason. Either partner can disconnect at any time, which closes the report for both.

Playing, inviting and joining need no sign-in; answers saved without signing in are deleted after 30 days without a visit, unless a payment or a signed-in partner's report depends on them. To open the report, a pair either pays once or both partners sign in, and opening the archetype then costs exactly the coins every new account starts with. The other four chapters unlock together for both partners. The record of an unlock holds the chapter, the price or payment and the versions, not the report. Players who see the full report can answer four short feedback questions; those answers are stored with the game, never shown to the partner, and used only to improve it. There is no advertising or third-party analytics; I keep only daily aggregate counts, including the public games-played total, with no answers attached. Share images contain the archetype, its tagline, the avatars, the two reading counts and which of the four moments the two choices met in, never the answers themselves, ratings or attachment scores.

12. References

• Brennan, K. A., Clark, C. L., & Shaver, P. R. (1998). Self-report measurement of adult attachment: An integrative overview. In J. A. Simpson & W. S. Rholes (Eds.), Attachment theory and close relationships (pp. 46–76). Guilford Press.
• Fraley, R. C., Heffernan, M. E., Vicary, A. M., & Brumbaugh, C. C. (2011). The Experiences in Close Relationships—Relationship Structures questionnaire: A method for assessing attachment orientations across relationships. Psychological Assessment, 23(3), 615–625. doi:10.1037/a0022898
• Fraley, R. C., Waller, N. G., & Brennan, K. A. (2000). An item response theory analysis of self-report measures of adult attachment. Journal of Personality and Social Psychology, 78(2), 350–365. doi:10.1037/0022-3514.78.2.350
• Lefebvre et al. (2026). Journal of Marital and Family Therapy. doi:10.1111/jmft.70127
• Li, T., & Chan, D. K.-S. (2012). European Journal of Social Psychology. doi:10.1002/ejsp.1842
• Overall, N. C., Pietromonaco, P. R., & Simpson, J. A. (2022). Nature Reviews Psychology. doi:10.1038/s44159-021-00011-1
• Rodriguez, L. M., Coy, A. E., & Hadden, B. W. (2021). Journal of Social and Personal Relationships. doi:10.1177/0265407520975858
• Kenny, D. A., & Acitelli, L. K. (2001). Accuracy and bias in the perception of the partner in a close relationship. Journal of Personality and Social Psychology, 80(3), 439–448. doi:10.1037/0022-3514.80.3.439
• Crasta, D., Rogge, R. D., Maniaci, M. R., & Reis, H. T. (2021). Toward an optimized measure of perceived partner responsiveness: Development and validation of the Perceived Responsiveness and Insensitivity Scale. Psychological Assessment, 33(4), 338–355. doi:10.1037/pas0000986
• Laurenceau, J.-P., Barrett, L. F., & Pietromonaco, P. R. (1998). Intimacy as an interpersonal process: The importance of self-disclosure, partner disclosure, and perceived partner responsiveness in interpersonal exchanges. Journal of Personality and Social Psychology, 74(5), 1238–1251. doi:10.1037/0022-3514.74.5.1238
• Maisel, N. C., & Gable, S. L. (2009). The paradox of received social support: The importance of responsiveness. Psychological Science, 20(8), 928–932. doi:10.1111/j.1467-9280.2009.02388.x
• Reis, H. T., & Shaver, P. (1988). Intimacy as an interpersonal process. In S. Duck (Ed.), Handbook of personal relationships (pp. 367–389). Wiley.
• Reis, H. T., & Wheeler, L. (1991). Studying social interaction with the Rochester Interaction Record. In M. P. Zanna (Ed.), Advances in experimental social psychology (Vol. 24, pp. 269–318). Academic Press.

13. Attribution and reuse

Concept and original game design: Prajneya Kumar.

The original Between Us game concept, questionnaire design, interpretation framework and implementation are developed by Prajneya Kumar. If you adapt or discuss this work, please credit Prajneya Kumar and link to this methodology page. Copyright covers original expression (this text, my new items, templates, names and artwork), not ideas, methods or scientific findings; asking for credit creates no exclusive right over the concept of a couples questionnaire. Attachment theory, the intimacy process model, the cited findings and the ECR-RS items belong to their authors.

This article and my original item and template text are licensed under CC BY 4.0. The ECR-RS items and quoted third-party material are excluded. The software is licensed separately.

Kumar, P. (YEAR). Between Us: The psychology and mathematics
  behind my relationship game (Version 2.3) [Methodology].
  https://prajneyakumar.in/research/between-us (accessed YYYY-MM-DD)

Revision history

• Version 2.3: Sections 5 to 7 now describe what each chapter actually shows; the interpretation layer that turns measures into archetypes and stories (Section 6.2) is no longer published.
• Version 2.2: new access plan (play without signing in; open the report by paying once or by both signing in); share images show where the two threads met; each archetype lists three or four fictional couples; feedback questions at the end of the full report.
• Version 2.1: the five chapters are rewritten around each person, with plain-language explanations of every number; Chapter 3 adds a couple-level comfort-zone pattern and Chapter 5 becomes three conversation cards (Section 6.4); the close-call note is retired.
• Version 2.0: the combined report becomes five chapters opened by one of ten couple archetypes, built from perceived responsiveness and prediction accuracy (Section 6); the eighteen couple types and sixteen lenses are retired; reports are computed on request and never stored; partners see each other’s answers after a new consent; retakes overwrite; evidence tiers renamed to match the five-way split used in the game.
• Version 1.1: plain-language summaries, notation box, worked example and figures; eighteen couple types; the solo report; outcome-only sharing; ratings can no longer be skipped. Wording bands aligned with the shared report’s attachment leans.
• Version 1.0: first version, written alongside the game.

Dates are recorded in IST when published.