Over the past 50 years, we've optimized society for convenience.
Double Cup asks what it would look like to optimize it for belonging.
The report tests whether neighborhood coffee shops can become measurable community infrastructure — belonging that healthcare can act on, and account for.
Take two cups. One is an invitation.
In twelve-step recovery, one suggested practice is to grab two cups at the coffee pot — one for yourself, one for someone who may need it. It is the Double Cup Rule: the second cup is not really about coffee, it is a small act of service.
Double Cup brings that small act into everyday neighborhoods. The report asks what healthcare could do with the belonging that follows.
Health plans have become increasingly capable of identifying social needs. They still lack reliable, measurable places to respond.
A Connector shop is both a recognizable place to belong and a measurable destination for an otherwise open-loop referral.
Repeated small acts of service accumulate. Do enough of them in one place, often enough, and that place begins to pull people toward it: those who want to give know where to go, those who could use an invitation know where they will find one, and belonging has somewhere to begin.
That is what turns an ordinary coffee shop into an intentional destination — the missing where a referral needs.
Today's system is built to identify social risk, document it, and generate referrals. What remains largely undefined is a scalable, measurable destination — where connection itself becomes part of the intervention.
This report explores a single design question:
Can belonging become a measurable intervention?
Rather than beginning with a solution, it follows the path from problem to evidence, from evidence to design, and finally to a pilot model that can be measured in practice.
ⓘ Regulatory measures, published research, and program requirements are real and fully cited.
Member-level figures, charts, and projected outcomes are illustrative. They demonstrate the analyses a Double Cup pilot would perform and the plausible shape of its results — not actual program performance.
Illustrative figures are clearly identified throughout.
Healthcare has become increasingly accountable for identifying social needs. It still lacks equally scalable places to intervene.
Under HEDIS SNS-E, plans must deliver an intervention within 30 days of a positive social-needs screen, while a new Medicare Advantage behavioral-health Star measure now evaluates depression screening and follow-up care.
Yet social determinants are documented in only 0.5–2.4% of encounters, and referral pathways often end with little visibility into what happens next. Healthcare is increasingly expected to improve social outcomes without a measurable, community-based destination. Double Cup proposes one.
Connection is already understood as protective. This section assembles what the science settles, marks off what it leaves open, and narrows to the one problem worth solving: whether belonging can be made observable, repeatable, and measurable.
Across decades of population research, people with stronger social ties live longer and stay mentally and physically healthier. The effect is not marginal — pooled across millions of participants, the mortality impact of isolation rivals well-established risk factors like obesity and physical inactivity.
Disconnection is not a stable state — it is a loop. Withdrawal reduces positive interactions, which worsens mood and coping, which drives further withdrawal. Each turn lowers the odds of the next connection.
Isolation isn't static. Because each turn lowers the odds of the next positive interaction, mood and coping erode while withdrawal grows — so the loop tightens rather than loosens over time.
What research establishes is that connection matters and isolation harms. What it does not establish is a practical, everyday way to manufacture connection and see that it happened.
Health plans already fund in-person engagement, community navigation, and peer support. Double Cup takes the one domain none of them are scoped to — social isolation — and pairs it with a privacy-preserving evidence layer from day one.
Tap any row to expand
Minnesota's move to make peer services reimbursable proved two things at once: the demand is real, and reach without a verifiable record is fragile. In 2024 the state halted Medicaid payments to a leading Minneapolis peer provider, Kyros, over billing-integrity concerns, and it wound down. The lesson isn't that peer-style reach fails — it's that the evidence has to be built in, never reconstructed after the fact. That's why Double Cup's evidence layer is consented, aggregate, and EVV-native from day one — belonging is the product; the evidence is a privacy-preserving byproduct.
The question is no longer whether connection matters.
The question is whether it can be intentionally designed, repeated, and measured.
How belonging becomes observable, repeatable, and measurable.
One input — buy two cups, offer one — accumulates into belonging when the act is easy to initiate and easy to repeat.
The intervention is intentionally small. Its value is the slope: each repetition compounds on the last.
Three parts, organized around the member. Each is limited alone; together they form the Double Cup ecosystem.
Read it as a care ecosystem: People, Places, and Partners aligned on one shared foundation.
Belonging is the intervention. The evidence loop demonstrates its impact — without turning relationships into data.
People experience belonging. Healthcare receives evidence, in aggregate — never the relationship itself.
The instruments the model is built to move — read against today's baseline.
Regulatory facts are real and cited. Member-level figures are modeled and illustrative, tagged as such.
Double Cup measures participation and outcomes — not relationships.
You don't apply to belong. You belong when you choose to take part.
Membership is self-declared and open, so the record can stay about participation and outcomes — never identity, and never the relationship itself.
People can take part without identifying themselves. Belonging doesn't require a login.
Health-plan links use consented referral tokens — not member identities.
Collect only what's needed to demonstrate impact — nothing more.
Never records conversations, friendships, or who met whom.
Healthcare sees participation and outcomes in aggregate — never individuals.
End · Ch.03 — The Model
Small by design. Private by design. Built to strengthen belonging.
How the model performs under real-world assumptions.
Every step from screening to a documented intervention loses people. The measure, and the money, lives at the bottom of the funnel, exactly where open-loop referrals break down. The lighter bars model what a ready intervention destination recovers.
Modeled and illustrative. Hover any stage for its counts.
Of 1,193 positive screens, 646 produce a referral, but only 257 reach a documented intervention. The referral-to-intervention step converts at 39.8% and is the binding constraint: end to end, 78.5% of positive screens (936 members) never close a loop the plan can claim credit for. Screening volume is not the problem, and adding more of it would widen the gap rather than narrow it. Separately, 527 members in the highest-risk quartile were never screened at all, so they are invisible to the measure before the funnel even begins.
Not every quality measure is in reach of a coffee shop. These are the ones where social connection is a plausible, literature-backed lever. The projected point gains are modeled and directional; the pilot exists to size them for real.
SNS-E screens food, housing, and transportation. Social isolation, which the U.S. Surgeon General called an epidemic, isn't its own screened domain yet, and its Z-codes (living alone, lack of social support) are documented on around 1% of encounters.
So plans have a recognized need, growing pressure to address it, and no off-the-shelf intervention built for it. That gap is the opportunity: Double Cup is a low-cost, community-based destination purpose-built for connection, that logs what it does.
Belonging is the intervention. The evidence loop confirms it happened and whether it helped — in aggregate, capturing participation and outcomes, never the conversation, the friendship, or who met whom.
This is the difference between a resource list and belonging you can point to: the plan learns whether the invitation was taken and whether things improved — in aggregate. The evidence layer records participation and outcomes, never conversations, friendships, or who met whom. The Club recognizes generosity and participation, never a diagnosis.
The evidence layer is optional infrastructure that lets healthcare evaluate its investment in belonging — never the primary product, and never a window into anyone's relationships. It exists to answer four questions, and to collect only what those answers require:
Three tiers of evidence, ordered by how much cooperation each needs and how little each reveals: what Double Cup confirms itself, what a consented referral token attributes, and what only the plan's own claims can measure — always in aggregate.
Every Connector check-in and Double Cup redemption is a consented participation event — confirmation the invitation was taken. It's the minimum needed to show the intervention happened, exportable to the plan in aggregate. No conversation, no relationship, is ever recorded.
The Connector's structured note attaches the specific driver. A Z-code needs a documented risk or unmet need, not merely the circumstance, which the note supplies. Double Cup's own domain is social isolation (Z60.x); the Z59 codes mark adjacent food, housing, or transport needs a Connector surfaces and refers onward.
So the plan's care team can document and bill the non-clinical outreach that routed the member here. Community Health Integration (CHI) and Principal Illness Navigation (PIN) are the durable path for community-based staff, and PIN explicitly covers behavioral-health conditions. (HCPCS G0136 was dropped from SNS-E for MY2026 after the CY2026 fee schedule redefined it.)
Standards-based, so it reads as interoperable data to a plan rather than a bespoke feed.
Two directions: where a member came from, and where the Connector sends them onward. Both are captured with shared identifiers, so nothing rests on memory or self-report.
Each channel issues its own signed short-code or link, recorded on enrollment and every redemption. Every participant carries who sent them.
The Connector's "refer onward" opens a referral the receiving org acknowledges and closes, over a community information exchange. This confirms the handoff; the kept clinical visit is confirmed in Tier 3.
A consented match of plan member ID to Double Cup participant ID. This single link is what makes Tier 3 possible; without it, nothing connects to outcomes.
The only tier that can see a BH clinic visit or a utilization change. Claims never flow into Double Cup. Double Cup sends exposure only, and the plan matches it against its own claims in a governed environment.
Nothing clinical: only who was exposed and when.
Index date = a member's first Double Cup intervention. Exposed members are compared to a propensity-matched cohort the plan pulls, or to the stepped-wedge rollout of the pilot itself. Reporting stays aggregate and de-identified under a BAA / limited data set.
Read top to bottom, this is also a build order. Tiers 1 and 2 are process metrics a pilot can stand up on day one. Tier 3 is the outcome question, and standing up its consent flow and data-sharing agreement is the pilot's real job, which is why every outcome figure in this report is still labeled modeled.
The same data that flags who to screen also shows where the referable, high-need population concentrates, and therefore where a Connector shop earns its keep. Both views below are illustrative, modeled on a Minnesota-shaped cohort.
The highest average isolation signal sits in Plymouth and Eden Prairie, but the unclosed positive screens concentrate elsewhere: Mpls Core (169), St Paul (128), and Plymouth (119) together hold 44% of every loop the plan failed to close. Of the 936 unclosed members, 350 have no Connector shop within reach — that coverage gap, not raw risk, is what a siting decision should optimize against. Outstate Rural (94 unclosed) is the one cluster where a coffee-shop model likely will not reach; it needs a different intervention, and the report should say so rather than overclaim.
A Claude-powered API earns its place in the parts of this system that turn on language and judgment, and stays out of the parts that must be auditable or private. That discipline is the point.
Unstructured text into structure, real-time Connector guidance, referral reasoning, and plain-language synthesis. Language and judgment, always with a human in the loop.
Risk scoring, dormancy prediction, and outcome attribution. Calibrated, auditable, and owned separately, so a number a plan audits is never "the AI decided."
A Claude assistant inside the Connector's app that turns the "notice, invite, refer" training into live support during a real interaction. It works from the public local-resource directory and only what the Connector chooses to share in the moment, never the member's record.
Converts a Connector's free-text note into the Z-codes, service codes, and FHIR resources from Panel 04, with a human confirming. Zero-retention: nothing stored or trained on.
Reasons over a messy, non-uniform resource graph to rank the best clinic or service for an expressed need, with rationale, feeding the closed-loop referral in Tier 2.
Writes the coffee-shop, sponsor, and health-plan narrative on top of the aggregate dashboards, refreshed on demand. A human signs off before anything reaches a payer.
Connector types: "Regular came in again, third time this week. Said the apartment's been quiet since his brother moved out and he's not sleeping. Didn't want to talk long. I gave him a second cup and mentioned the Thursday group."
Kept clear of PHI by design. The intelligence layer works from the public resource directory, de-identified aggregates, and what a Connector shares in the moment. It never touches the identity crosswalk or the plan's claims; Tier 3 stays in a governed environment with no model access, and where any member text is processed it runs under a zero-retention BAA with a human confirming the output. The auditable risk and attribution models stay separate, so "the AI decided" never enters a payer conversation.
The Population Fit tool decides, offline, whether a plan should pilot at all. This is the other half: once a pilot runs, did it move anything real? The honest answer is we don't yet know — so this is the measurement we'd commit to before claiming it did.
Evidence for the problem is real and cited. Evidence for the solution is projected, not proven. Everything below is the plan we'd use to replace projection with measurement — including the results that would tell us to stop.
Each rung is measurable before the next; the plan-value rung is the only one a payer can claim credit for.
This is observational, so we assume we're wrong until a design says otherwise. Attribution runs as a matched-cohort or stepped-wedge comparison — never a naive before/after — and the analysis is pre-registered so we can't move the goalposts after seeing the data. We'd redesign or walk away if:
This is a design prototype, not a study. What's real is cited; what's modeled is labeled. The detail below states exactly which is which.
Double Cup OS is a working prototype. Turning it into a public pilot takes partners — plans, shops, sponsors, and collaborators who want to make belonging measurable.