Domain Atlas / Child welfare & family services
The vendor's ledger: Family-Match, the eharmony-derived adoption matcher the states kept coming back to
Family-Match, a proprietary two-sided 'relational fit' adoption matching algorithm built for Adoption-Share by former eharmony researchers, produced 1 known adoption in Virginia's two-year test (an official's statement to the AP; VDSS said the tool 'had not proven effective'; the pilot's end is undated in the record) and 2 adoptions in Georgia's year-long pilot ended October 2022; caseworkers in Florida, Georgia, and Virginia said it often led them to unwilling families, and Virginia social workers were perplexed that the algorithm seemed to match all the children with the same group of parents. In Florida the vendor's own quarterly report claimed 603 placements yielding 431 adoptions over five years — figures partner agencies could not verify: FamiliesFirst Network's records showed 76 Family-Match placements with no documented adoption plus 3 failed trial placements since 2019, and Children's Network of Southwest Florida counted 22 matches and 8 adoptions in five years while making hundreds of matches and hundreds of adoptions without the tool over the same period.[2]
What happened
Adoption-Share Inc. — a 501(c)(3) founded in 2013 in Brunswick, Georgia by Thea Ramirez, a former social worker who had earlier run a network for crisis pregnancy centers and framed adoption as abortion reduction — commissioned its Family-Match algorithm from Gian Gonzaga, who managed eharmony's matching algorithms, and his wife Heather Setrakian (eharmony says it is 'not affiliated' and calls them 'simply former employees'). The design is two-sided, a matching operator rather than a risk scorer: adults seeking to adopt register on the platform and complete compatibility surveys, foster parents or social workers input each child's information, the algorithm generates a 'relational fit' score over the two pools, and caseworkers see a ranked list of top prospective parents per child, whom they then vet for a trial placement and eventual legal adoption. Virginia announced the first-in-nation test in November 2017 — a pilot ordered by then-Gov. Terry McAuliffe at the urging of a campaign donor he had appointed the state's 'adoption champion' — and matching began in 2018 in both Virginia and Florida, where the Selfless Love Foundation (founded by the then-CEO of the Patron tequila company and his wife) financed a free statewide rollout across Florida's privatized community-based care lead agencies, reaching all of them within a year. Georgia's pilot launched December 1, 2021 in 3 of the 14 regions of the state's Division of Family and Children Services (DFCS), with roughly 300 children statewide awaiting adoption. Political sponsorship traveled with the product: Ramirez was spotlighted at the First Lady's June 30, 2020 White House foster-care roundtable with Health and Human Services (HHS) Secretary Azar (AP's reporting places the event in the Situation Room), presented at the American Enterprise Institute, and the vendor pitched Delaware, Missouri, New York City, and US HHS.
The independently observed results were thin. Virginia's two-year test produced one known adoption — an official's statement to the AP, not a published evaluation; the AP dates no end for the pilot (circa 2020) — and VDSS concluded the tool 'had not proven effective'; social workers 'did not find the tool particularly useful,' and were perplexed that the algorithm seemed to match all the children with the same group of parents. Georgia's pilot produced 2 adoptions in its year and was ended in October 2022. Caseworkers in Florida, Georgia, and Virginia told the AP the tool often led them to unwilling families — the phantom-match complaint that dominates the operator record — and veteran Florida adoption worker Connie Going, whose viral adoption story Ramirez cites as her inspiration, said the tool gives waiting parents false hope and makes her job harder. Florida is the vendor's showcase, and its numbers are the vendor's own: a quarterly report obtained by the AP claimed 603 placements yielding 431 adoptions over five years. Partner agencies could not corroborate: FamiliesFirst Network in Pensacola carried 76 Family-Match placements with no documented adoption plus 3 failed trial placements since 2019 — unexplained by the agency's own attorney — and Children's Network of Southwest Florida counted 22 matches and 8 adoptions in five years while making 'hundreds of matches and hundreds of adoptions' without the tool over the same period. The AP investigation (November 6, 2023, by Sally Ho and Garance Burke), built on public-records requests across four states, concluded the tool 'produced limited results in the states where it has been used' and that Ramirez had 'overstated the capabilities of the proprietary algorithm to government officials.'
What the records requests exposed underneath the outcome dispute was an ownership inversion. Virginia officials said that once families' data was entered, 'Adoption Share owned the data'; assistant director Traci Jones said 'We did not have access to the algorithm even after it was requested.' In Georgia, Family-Match gathers whether foster youth have been sexually abused, the gender of their abuser, whether they have a criminal record, and whether they 'identify as LGBTQIA' (lesbian, gay, bisexual, transgender, queer, intersex, asexual) — data typically restricted to tightly secured child protective services case files, held in a vendor-owned store. Two Florida agencies used the tool informally without any contract and would not say how children's data was secured; officials in three states said they were not sure how the tool scored families on the sensitive variables powering it, and Ramirez refused to provide details of the algorithm's inner workings. The ledger itself was fed by design: the April 2023 'confidential' user guide instructed workers not to delete cases matched outside the tool but to document them 'so that Adoption-Share could refine its algorithm and follow up with the families'; Miami's Citrus Family Care Network said vendor staff asked social workers to have parents register in the tool even when it played no role in the adoption; Georgia's spokesperson confirmed Family-Match could claim credit for pairings already in its system. Vendor-embedded staff — a 'Family-Match Director' and family coordinators — trained localities, ran the family recruitment funnel, and shaped record-keeping practice from inside operations. 'We're using, essentially, kids as guinea pigs for these tools,' Brown University's Suresh Venkatasubramanian, a former Biden OSTP official, told the AP. 'They are the crash test dummies.'
The lifecycle across four states is the case's second story, and every turn of it is a procurement act. Georgia ended its pilot because the tool didn't work as intended; Ramirez then met with the governor's office, lobbied a statehouse committee calling the tool 'an incredible feat,' and by July 2023 Georgia's Department of Human Services (DHS) signed a new agreement — this time for free. Virginia dropped the matching pilot, then by 2022 awarded Adoption-Share an even larger contract for the Faster Families Highway recruitment portal, which the nonprofit says 'leverages' the Family-Match application: the state's own December 2023 report to the General Assembly shows $212,546 spent in SFY2023 ($188,000 budgeted) and $246,100 planned for SFY2024, all 120 local departments enrolled by December 31, 2022 — and VDSS and the vendor only beginning to draft policies and procedures, including how to remove families and response-time expectations, in February 2023, after statewide enrollment. Florida's philanthropy-funded rollout converted to a $350,000 DCF contract in October 2023, plus a second Florida Department of Health contract to build an algorithm for placing medically complex children; DOH did not respond to the AP's repeated requests for comment. Only Tennessee walked away before deploying: after more than two years of setup, its DCS reviewers formally questioned why Family-Match needed certain sensitive data points and how they influenced the match score — the prospective-parent questionnaire asked household income, self-ratings on 'conventional' and 'uncreative,' and agreement with a statement about seeking God's help — and the state scrapped the rollout. Everything after November 2023 is vendor-reported: the February 2025 annual report claims a fall-2023 Georgia expansion to 7 regions and 77 counties, a fall-2024 five-year renewable agreement, 600+ cumulative Florida adoptions, a SOC 2 Type 2 audit, and a tenth consecutive institutional review board (IRB) approval, while IRS filings show revenue climbing every year since the investigation to a record $1,710,394 in FY2025. No state has ever published an evaluation of the matching pilot, and no follow-up investigation, inspector-general audit, or legislative response is documented through July 2026.
The sociotechnical reading
Every other child-welfare scorer in this Atlas ranks one population against a decision threshold, and its drama is what an operator does with the score. This case's core object is different in kind: a cross-population edge. Family-Match's output is a proposed pairing between a scored child and a scored family, and nothing of value exists until that proposed edge is realized through caseworker vetting, family consent, and judicial finalization. Read as a system, that changes what can fail. Matching operators have failure modes scalar risk scorers cannot have: candidate-set degeneracy (Virginia's workers watching the algorithm match all the children with the same small group of parents), phantom edges (proposals to families who turn out unwilling — the dominant caseworker complaint in three states), and a claimed-versus-realized edge gap, which this record documents at every scale: 603 claimed placements against 431 claimed adoptions against 8 verified at one agency against 76 tool-attributed placements with no documented adoption at another. And because value lives on realized edges, the binding constraint is pool composition, not score accuracy — nationally, recruitment-based placements with previously unknown families are on the order of 5,000 of 50,000 annual foster adoptions, and the vendor's own funnel controls which families enter the pool at all. A sharper matcher over a structurally thin pool proposes the same scarce families more confidently.
The second inversion is who sits where. In the usual topology the state owns the records and a vendor supplies a tool; here the vendor is inside the operational loop and the state is outside its own evidence. Adoption-Share embedded a Family-Match Director and coordinators who trained localities, ran the family funnel, and instructed caseworkers' record-keeping — including the April 2023 user guide's instruction to document matches made outside the tool inside the system. Meanwhile the outcome ledger belonged to the vendor: 'Adoption Share owned the data,' access to the algorithm was refused even on request, and agencies 'couldn't explain Family-Match's self-reported data.' The Lab draws this as structure: a second operator class for the vendor's embedded staff, wired into the caseworker floor as an instruction channel and into the matcher as the funnel that composes the family pool; a credit-claiming write into the vendor store drawn beside the ordinary child-side write; two record stores holding competing accounts of the same placements, the vendor's ledger and the agency's own record; and the state's procurement tier reading that ledger at full strength, while the reconciliation between the two stores and the disclosure of the scoring method to that same tier both sit at zero. The party being evaluated supplies the evaluation, and the two pathways that would break that loop are the ones the record documents as never performed. This is the mirror image of the domain's concealed-allocator sibling: there, operators cannot see the model's judgment; here, the deciders cannot see the model's evidence.
Third, the operator seam this domain usually argues about is not the problem — which is itself the lesson. Discretion was formally maximal and actually exercised: the tool is advisory, caseworkers vet every candidate, families must consent, judges finalize, recommendations were routinely discarded, and no documented case exists of the score overriding a caseworker's judgment. The Lab draws that discretion and, beside it, its documented limit: the caseworkers' influence over what the matcher can produce is drawn light, while the vendor's recruitment funnel — which composes the family pool the ranking runs over — is drawn at full strength. And the discretion did not save the system, because the harm never ran through deference: it ran through wasted search effort on phantom matches, false hope for waiting families, and — decisively — an evidence loop the operators' corrections never reached. A caseworker can discard a bad proposal; she cannot discard the ledger row that proposal became, and the ledger rows, summed by the vendor, were the evidence at every procurement table. The drop-and-resume lifecycle — Georgia ending a two-adoption pilot and re-signing for free after lobbying; Virginia dropping a one-adoption pilot and buying a larger rebranded portal, drafting removal policies only after statewide enrollment; Florida converting philanthropy funding to a state contract and adding a second deployment — is what procurement looks like when the party being evaluated curates the evaluation evidence, when political sponsorship (a White House roundtable, a donor-championed pilot, a statehouse lobbying visit) substitutes for evaluation, and when free pricing removes even the budgetary occasion for scrutiny. The one state that gated before deploying — Tennessee, which formally asked why the sensitive fields were needed and how they moved the score — is the one state with nothing to unwind.
So the instruments that fit this cell govern the evidence, not the judgment. The contract is the surface most of them attach to: a vendor gate that prices data ownership, algorithm inspectability, and independently verifiable outcome reporting into the relationship (the record shows two agencies running the tool with no contract at all — no gate even nominally existed); write-side gates and provenance labels that separate tool-proposed, externally-made, and vendor-solicited entries before they can be summed into an annual report; a correction budget for the reconciliation nobody ever funded — walking claimed matches back to realized adoptions; and a cadence that makes the schema interrogation and the re-signing evaluation recurring events rather than one state's one-off. The distinct lesson the Atlas draws here: in a two-sided matching deployment, govern the ledger before the algorithm — a matching system's success metric is a count of edges, edge counts live in a store, and whoever owns that store owns the deployment's truth. The honest boundary throughout: served children and families are not modeled in the paired Lab, which reads institutional propagation only; every favorable figure is a vendor claim from the vendor's own store and is labeled as such; the pilot tallies are officials' statements, not evaluations; the drop-and-resume turns are procurement authority actions, never adjudicated failures; and the adjacent matching-theory literature — whose authors have collaborated with this vendor since at least 2018 — is a mechanism-class prior about caseworker-driven search, never efficacy evidence for this tool.
The concepts used in this reading are defined in the Field Guide; the governance responses live in the Practice Library.