Domain Atlas / Benefits navigation & public-facing chat
MyFriendBen benefits screener
MyFriendBen is an advisory multi-state benefits screener: an anonymous survey of about six minutes returns a report of programs a household is likely eligible for, with estimated dollar values, and submits nothing, routing people instead to separate government application channels that make every determination independently. All of its eligibility math is computed by PolicyEngine, a separate nonprofit whose open-source codebase encodes federal and state statute; MyFriendBen deliberately built no proprietary rules logic, and the same substrate sits under each of its state deployments, so an encoding error would misestimate benefits in every state simultaneously and a correction would propagate to every state simultaneously. That substrate's scale is indicated by a $300,000 NSF POSE Phase I award announced on August 18, 2025. The shared-substrate structure is documented in both organizations' technical materials; no incident of a cross-state encoding error appears in the public record. The accuracy figure attached to it, 'over 90%', is a builder, infrastructure-organization and operator claim with no published methodology and no independent audit.[4]
What happened
MyFriendBen is a nonprofit benefits screener. A household answers an anonymous survey of roughly six minutes — household size, income, expenses, current benefits, county or zip — with no name and no identity documents, and receives a report of programs it is likely eligible for, each with an estimated dollar value and a time to apply. The tool submits nothing. It routes people to separate government application channels, and the operator documentation is explicit that its results are estimates and do not guarantee eligibility. It was prototyped in early 2022 by a Denver philanthropy after feedback sessions with about 100 Colorado families, modeled on an existing city benefits screener, piloted with 2-1-1 Colorado call staff in the autumn of 2023, and was operating publicly statewide through 2023; the independent journalism that anchors most of its verified figures was published on January 5, 2024, which is the coverage date and not the launch date. In October 2024 the Bill & Melinda Gates Foundation gave $2.4M to scale it to additional states, with a stated goal of unlocking $100M in benefits over two years. It spun out as an independent nonprofit around 2025, and by mid-2026 was live in Colorado, North Carolina, Massachusetts, Illinois and Washington, with a Texas accelerator announced on June 30, 2026. One infrastructure-side post also lists New York among its states; the organization's own six-state list omits it, so New York's current live status is unclear in the record.
The structure that makes this case an atlas object is underneath the screeners rather than inside any one of them. All eligibility math — SNAP, Medicaid, SSI, TANF, WIC, school meals, and federal and state tax credits — is computed by PolicyEngine, a separate small nonprofit whose open-source codebase encodes federal and state statute as computable rules. MyFriendBen deliberately built no proprietary rules logic. Its own backend is a white-label multi-tenant layer holding per-state program catalogs, feature flags and a materiality threshold that admits a program only if it is worth at least $300 a year, and independent local anchors operate the front doors: Code the Dream with NC 211 in North Carolina, Benefit Illinois's Illinois Benefit Hub, the MASSCAP community-action network in Massachusetts, Child Poverty Action Lab in Texas. Each configures a divergent catalog over identical math. The consequence is a cross-jurisdiction common mode: six or more nominally independent state screeners share one eligibility computation, so an encoding error would misestimate benefits in every state simultaneously, and a fix would propagate to every state simultaneously too. That codebase is maintained by an organization whose scale is indicated by a $300,000 National Science Foundation Phase I award for open-source ecosystems, announced on August 18, 2025. The correlated structure is documented and real. The harm mode is hypothetical: no incident of a cross-state encoding error appears anywhere in the verified record, and nothing here asserts one.
What the record does not contain is any independent measurement of whether the estimates are right. An accuracy rate "over 90%" appears in the infrastructure organization's post, in the North Carolina operator's launch release and in the builder's materials, always without a published methodology and never with an external audit; it is usable only as an optimistic upper bound. There is no government oversight body for this tool at all — it is a nonprofit product outside any statutory scheme. Effective oversight is a board, its philanthropic funders, a youth advisory board of people with lived benefit-system experience, family co-design sessions that demonstrably redirected the roadmap in 2022, and one Urban Institute evaluation. That evaluation is commissioned, is described by a funder's own post as currently under way, and is visible only through the deployment's own citations; its interim figures from 301 users — 64% discovering a benefit for the first time, 91% planning to apply within three months, 93-95% reporting satisfaction — measure discovery, intent and satisfaction, not calculation accuracy and not verified enrollment. No oversight body has ever audited or forced correction of the eligibility calculations themselves.
The impact figures are similarly soft, and they contradict each other. A vendor press release in October 2024 reported 20,000+ Coloradans supported, $800M in benefits identified and $12M delivered. A builder case study reported 55,000+ households helped to apply for $33M over two years. A funder post in 2026 restated Colorado as 100,000+ households and $52M+ accessed. A joint release in June 2026 reported 115,000+ households and $58M nationally, while the organization's own navigator page in the same period reported 65,000+ families and $1.2B+ identified. The denominators slide between identified, applied for, unlocked and delivered, and that slippage is itself the datum. Exactly one set of figures comes from independent journalism: in all of 2023 in Colorado, about 5,500 households were screened, about $30M in benefits was identified, and about $5M was estimated to have actually been obtained — a roughly one-in-six ratio between what was found and what was received, with the median user reporting an income just over $8,200 a year and a household size of two.
Two further structures matter. Screens are stored anonymously in the backend database, together with eligibility results and, where a feature-flagged conversational assistant is enabled on results pages, conversation history. That store never syncs with any government system, so a screening error leaves no trace in any official record and no correction pathway flows back from what a household actually received. Full human discretion exists at every exit — a household decides whether to apply, a navigator can second-guess the output, and a government caseworker makes every determination independently — but the un-overridable failure is the silent false negative: an eligible household missed by the screen and told it qualifies for nothing simply never applies, and because the screening is anonymous and disconnected from agency records, no caseworker, auditor or appeal ever sees it. Meanwhile the documented feedback loop runs outward. Aggregate screening data and a Colorado Child Tax Credit Calculator derived from the same code, credited with identifying $177M, supported advocacy around Colorado HB24-1311 and HB25-1335, and the builder credits that ecosystem with unlocking an $810M state child tax credit — a rule the substrate then had to encode. The independent research frame for the whole category, from the Aspen Institute's Financial Security Program, is that no systematic evaluation exists of whether benefits screeners increase benefits access at all.
The sociotechnical reading
Most of this atlas's correlated-failure cases are correlated inside one institution: one fleet, one vendor contract, one agency retiring many engines onto a shared one. This case is the cross-organization version, and that difference is the whole lesson. Six nonprofits in six states, with separate boards, separate funders, separate program catalogs and separate front doors, look from the outside like six independent attempts at the same problem. Underneath they are one computation. Independence at the governance layer buys nothing when the eligibility math is a shared dependency, and none of the six holds a staging gate, a rollback right or a notice period over a change to it. The protective half of the same structure is real and should be said plainly: because the substrate is one open-source codebase, a correction also reaches every state at once, which is more than most federated deployments can manage. Concentration is not a defect here so much as a trade, and the trade has never been priced, because nobody has measured the error rate on either side of it.
The second lesson is about where an advisory system's errors actually go. This tool has no authority: it cannot determine, cannot deny, cannot write into an eligibility system, and every real decision is made later by a caseworker who never sees the screen. That is genuinely protective, and it is also why the error channel is so hard to see. Errors propagate through belief — what a household concludes, and what a navigator working from the report takes to be true in a job where learning one program's rules the other way takes months. The system's own design then closes every route by which such an error could be found. Screening is anonymous, so there is no household to follow up; the store never syncs with government records, so there is no determination to compare against; nothing is submitted, so there is no application to audit; and the failure that matters most, the eligible household shown nothing, produces no artifact at all. Discretion exists on the apply side and not on the discourage side. The one number in the record that bears on the gap — about $30M identified against about $5M obtained in a single state-year — was produced by a newspaper, not by the system.
Third, this is a transparency asymmetry worth naming, because it cuts against the usual story. The entire stack is open source and inspectable, published under a public licence, which is the opposite of the proprietary risk-score opacity most of this atlas documents. And the headline accuracy claim covering it is unaudited, its method unpublished, repeated verbatim by the builder, the infrastructure organization and a state operator. Inspectability is not verification: anyone may read the encodings, and nobody has checked them against the statutes they compile. The same asymmetry runs through the impact reporting, where mutually inconsistent totals with sliding denominators sit beside a single independently reported figure. The honest boundary here: none of this measures anything about the households. What this map models is an institutional propagation loop — a shared computational substrate, three operator groups at three different employers, and a record store that is, from the front line's side, write-only by design. No harm to any household has been documented, and the reason that is not reassuring is that the system is built so that such harm would leave nothing behind to document.
The concepts used in this reading are defined in the Field Guide; the governance responses live in the Practice Library.