What we’re building on, and what we won’t claim.

A sentence that rests on a relayed source needs a first-hand read and review by someone who studies this before it should be relied on. Until then, treat those rows as leads.

1. F1

   Deeper conversation

   Relayed, not opened

   Kardas, Kumar & Epley (2022).

   Found

   Strangers assigned a deeper 20-minute conversation reported higher felt connection than those assigned a shallow one. People also tended to expect deep conversation to be more awkward than it turned out to be.

   The design does

   Offers optional “go a little deeper” prompts in the prep chat. Skipping is as prominent as using, and a line names the source.

   We won’t say

   Deep talk makes friends. It raises connection. Anything about later contact: the studies did not establish it.

2. F2

   Brain signals

   Self-listed, under review

   A manuscript under review; its details are withheld until it is published.

   Found

   We hold one self-listed source on brain signals and conversation. It is a lead and nothing more, and classification figures from small samples are not accuracy you could expect in use.

   The design does

   Nothing. No brain, voice or text-to-brain inference appears anywhere in the product.

   We won’t say

   We can detect connection. Any accuracy figure.

3. F3

   The liking gap, and who sees what

   Relayed, not opened

   Boothby, Cooney, Sandstrom & Clark (2018); Vazire (2010).

   Found

   People tend to underestimate how much a conversation partner liked them. Self and others each see some traits more accurately than the other does.

   The design does

   A reflection with three voices (what you said, what a participant reported, what the chat showed), each line correctable. A sealed, independent “see again” answer; one-sided interest is never revealed.

   We won’t say

   Made-up reassurance. Someone likes you before it is mutual. That four near-strangers see you as accurately as established friends do.

4. F4

   Time together

   Relayed, not opened

   Hall (2019); Pearce, Launay & Dunbar (2015).

   Found

   Time together, leisure and everyday talk are associated with closeness. Group formats can change the pace of bonding without determining where it ends up.

   The design does

   Makes the second plan the high point of the loop, shows a plain count of meetings, and lets people choose an activity format.

   We won’t say

   No hours-to-friendship figure, You are three meetings away, or streak counter.

5. F5

   Group size

   Our own reasoning

   No source behind it.

   Found

   Nothing we hold establishes five as the best group size.

   The design does

   Treats five as a starting choice: one table, four people to reflect on, and one no-show doesn’t cancel the evening. Group size is a variable the pilot can vary.

   We won’t say

   Research shows five. Any justification from friendship-group-size theory. Anonymous, in a group of five.

6. F6

   The event matters a lot

   Our own reasoning

   Consistent with Pearce et al. (2015), but our inference.

   Found

   Venue, host, access and who turns up can swamp any effect of who is in the group.

   The design does

   Shows duration, access, host, cost and how to cancel on the ticket. Asks “did you enjoy the activity” and “would you see them again” as separate questions, and keeps a record of format, venue, host and attendance.

   We won’t say

   That the product evaluates hosts, or pairs people with hosts.

7. F7

   An optional friendship check-in

   First-hand read (extract)

   Kaufman et al., Friendship Network Satisfaction (online 2021; figures below are from pages 5, 6, 11, 13 and 16); Kaufman (2020), UCLA dissertation.

   Found

   A 14-item scale rating agreement from 0 to 5 about your whole friendship network over the past year. High internal consistency (alpha .96). Two facets, closeness and socializing, add little beyond the total. Two online samples, about 2,100 and 2,000 people. No test-retest data, no norms, and no testing of whether it works the same way for different groups.

   The design does

   Offers it as an optional whole-network check-in, as published, outside the main flow, shown only to the person. Scoring and permission to use it are marked pending confirmation from the scale’s authors.

   We won’t say

   Anything that ranks a person against others, or says what a total should be. Your score went up. A shortened validated version. A person, pair or event is never rated with it.

8. F8

   Satisfied with friends is not the same as not lonely

   First-hand read (extract)

   Kaufman et al.; Kaufman (2020).

   Found

   Satisfaction with friends and loneliness were only weakly related (a correlation of about .22 in size, in the expected direction, page 14). The authors report that loneliness varies widely at every level of friendship satisfaction. A schematic follows in the check-in section.

   The design does

   Puts one line beside the optional check-in: being satisfied with friends is not the same as not being lonely.

   We won’t say

   Reduces loneliness. A low number means lonely. Nothing clinical.

9. F9

   Twenty reports are not twenty facts

   Our own reasoning

   No source behind it.

   Found

   Five people give ten pairs and twenty directed reports, and they are not independent of each other. A handful of groups can only describe what happened.

   The design does

   Shows follow-ups (happened, planned, declined, no reply) with their denominators, and treats no answer as not a no.

   We won’t say

   Twenty data points. Any p-value or success rate drawn from the demo’s example people.

1. Enjoyed the planReported that evening.
2. Felt understoodReported, privately.
3. Wants to see someone againSealed, and shown to nobody.
4. MutualBoth said so independently.
5. Second meetingBoth people report it happened.

• We don’t say proves, causes or boosts. We say “reported” and “associated with”.
• We don’t say the product builds friendship, reduces loneliness or improves well-being.
• We don’t show a percentile, a norm, a compatibility figure, or a shortened validated instrument.
• We don’t say match. The words here are groups, plans and events.
• We don’t say anonymous about a group of five. We say “a participant reported”.
• We don’t invent warm feedback. A person’s words appear only if a source permitted them, with a way to correct them.
• We don’t promise hours, say conflict builds bonds, or detect friendship from brain, voice or text.
• We don’t say our edge is unique. Others overlap with us, and our edge is a hypothesis.
• We don’t imply a sponsoring institution, a named expert’s involvement, a live pilot, or real users. Everyone shown is fictional, so the example shows how the loop works and says nothing about privacy, access control or benefit.
• We don’t treat pilot thresholds as benchmarks. They are bars we chose.

• Chat readingOff until all five say yes, and anyone can turn it off again. With it off, every other part of the plan works the same.Shown in the demo
• Private noteA note only you read. It stays with you unless you choose to carry it to the second plan.Shown in the demo
• Sealed answersWho you’d see again is shown to nobody. Only a mutual yes opens anything, and a one-sided answer stays hidden.Shown in the demo
• ReportsGo to a person at the product, never to the group, and never change a sealed answer.Shown in the demo
• Research useOff by default. No ethics review exists yet, so no data from real people can be used for research.Designed, not built

1. Which friendship instruments are cleared for use in a demo, and who confirms how they are totalled?
2. Which conversation signals, if any, may be used with consent in an app, and which claims stay off limits?
3. What is a research-defensible next step after an optional self-report?
4. Does any study we hold measure later contact after a first conversation? None we have read so far does.
5. There is a missing bridge between a measure of connection after one conversation and a measure of a whole friendship network. Nothing we hold connects them.
6. Group size and event format could be pilot variables. What design would tell them apart from the effect of a particular host and venue?
7. Events and hosts are the largest unknown. What a good host and venue cost per group, and how much of any result they explain, is not something we can say.
8. No ethics review has been done, and none exists today.

Chosen, not benchmarks. With a handful of groups, results would be descriptive only.

• Boothby, E. J., Cooney, G., Sandstrom, G. M., & Clark, M. S. (2018). The liking gap in conversations. Psychological Science.Relayed, not opened
• Hall, J. A. (2019). How many hours does it take to make a friend? Journal of Social and Personal Relationships.Relayed, not opened
• Kardas, M., Kumar, A., & Epley, N. (2022). Overly shallow? Miscalibrated expectations create a barrier to deeper conversation. Journal of Personality and Social Psychology.Relayed, not opened
• Kaufman, V. A. (2020). Doctoral dissertation, UCLA (friendship network satisfaction and well-being).First-hand read (extract)
• Kaufman, V. A., et al. Friendship Network Satisfaction: a multifaceted construct. Journal of Social and Personal Relationships (published online 2021; issue year to confirm).First-hand read (extract)
• Miao, G., et al. Manuscript under review; details withheld until it is published.Self-listed, under review
• Pearce, E., Launay, J., & Dunbar, R. I. M. (2015). The ice-breaker effect: singing mediates fast social bonding. Royal Society Open Science.Relayed, not opened
• Vazire, S. (2010). Who knows what about a person? The self–other knowledge asymmetry (SOKA) model. Journal of Personality and Social Psychology.Relayed, not opened