The Number That Knocks: How Secret Risk Scores Decide Which Families Get Investigated — and Why Washington Just Paid for More
Investigation by Project Milk Carton
The Number That Knocks: How Secret Risk Scores Decide Which Families Get Investigated — and Why Washington Just Paid for More
By PMC Intelligence Desk | July 20, 2026
PROLOGUE: SEALED AT MIDNIGHT
At 11:59 p.m. Eastern Time on Monday, July 13, 2026, a window closed in Washington.
It closed quietly, the way federal windows do — no press conference, no floor vote, no hearing. It was the application deadline for funding opportunity HHS-2026-ACF-ACYF-CA-0037, a $6 million program run by the Administration for Children and Families inside the Department of Health and Human Services. The applications that arrived before midnight are sealed inside the federal grants system now. The public does not get to see who applied. What the public does get to see — because it is printed in the funding notice itself — is what the money is for.
The notice offers up to ten awards of $400,000 to $600,000 each, open only to state, territorial, and tribal child welfare agencies, to pilot predictive analytics in child welfare. In the agency's own words, it seeks tools to help make decisions "during hotline intake," systems to "flag high-risk cases during investigations," models for "generating risk scores, forecasting maltreatment," and analytics to "identify, target, and retain foster and adoptive placements." Awards are expected around September 30, 2026. Two more application cycles follow, in June 2027 and June 2028.
Read cold, that is a technology grant. Read against the ten-year public record — which is what this investigation does — it is something else: a federal offer to scale, statewide, a category of software whose documented history includes a tool in Illinois that assigned more than 4,100 children a 90-percent-or-greater probability of death or serious injury while missing the children who actually died; a Los Angeles pilot that generated 3,829 false alarms; an Oregon tool terminated weeks after an Associated Press investigation into its racial patterns; a New York City model with 279 variables that families, their lawyers, and even caseworkers are never told exists; and a Pittsburgh-area tool that has been under an open federal disability-discrimination inquiry since 2022 — an inquiry that has produced, in three and a half years, no public findings at all.
The machines score families. Nobody scores the machines. And as of midnight on July 13, the applications to build more of them are in.
This is the sixth module in Project Milk Carton's investigative series on the machinery of American child welfare. Like its predecessors, it runs on one rule, stated up front: every factual claim, number, name, case, document, and quotation in this article comes from primary and public sources— federal statutes and regulations, the funding notice itself, peer-reviewed research, federal court records, government reports, investigative reporting of record, and nonprofit financial filings, including filings held in Project Milk Carton's own public-data archive. Every figure is attributed. Where the analysis is ours, we say so. Where we could not verify something, we tell you that too, in a note at the end. Project Milk Carton produces transparency research from public data; we do not speculate, and we do not invent.
The investigation proceeds in ten movements. First, what these machines actually predict — because it is not what the agencies say. Second, what the machines eat — the data that goes in, and why every input is a proxy for being poor, disabled, or already surveilled. Third, the documented record: four jurisdictions, four failures. Fourth, the two academics who built most of this field and the conflict at the center of its evidence base. Fifth, the vendor — a $350-million-a-year nonprofit whose flagship safety tool failed its only independent statewide test. Sixth, the money: how the federal government has been paying for these systems all along, through plumbing no legislature ever voted on. Seventh, the disability question, and the federal inquiry that went silent. Eighth, the funding notice itself, read line by line — because ACF's own document is the most damning exhibit in this file. Ninth, the accountability gap: no statute, no courtroom, no audit, no one whose job this is. Tenth, what would fix it — four changes, none of which require new science.
Then we will come back to the envelope, because the story does not end at midnight. It starts there.
PART I — THE MACHINE: WHAT THE SCORE ACTUALLY PREDICTS
The substitution at the heart of the field
Begin with the single most important fact about child-welfare risk scoring, the one the brochures never lead with: the flagship models do not predict child abuse.
The Allegheny Family Screening Tool — the AFST, the most studied and most emulated algorithm in American child welfare, running on every maltreatment referral in Allegheny County, Pennsylvania since August 2016— produces a score from 1 to 20 for each family called into the county's child-abuse hotline. That score is not a measured probability that a child is being harmed. It is the model's estimate of the probability that the county will place the child in foster care within two years.
The distinction is not pedantic. It is the whole game.
Child maltreatment is largely unobserved — no dataset records what actually happens inside every home. Foster-care placement, by contrast, is perfectly recorded, because it isa decision the agency itself makes. A model trained on placement is not learning what harms children. It is learning what the agency has historically done — which families it has historically investigated, which children it has historically taken — and then recommending the agency do it again. The American Civil Liberties Union said this plainly in its 2021 report Family Surveillance by Algorithm: "any tool built from a jurisdiction's historical data runs the risk of perpetuating and exacerbating, rather than ameliorating, the biases that are embedded in that data." Most tools in use, the ACLU found, are designed to predict removal — not abuse.
The pattern repeats wherever these systems appear. Douglas County, Colorado's tool — the Douglas County Decision Aid, built by the same team as the AFST — predicts a child's likelihood of being removed from home by CPS within two years of a referral. Allegheny's separate "Hello Baby" model, applied to newborns county-wide, is designed to identify infants at greatest risk of removal from the home by their third birthday. In each case the agency's press language says risk of harm; the model's target variable says risk of us.
Once you see the substitution, the feedback loop is obvious. The agency's past decisions become the model's training data. The model's scores shape the agency's future decisions. The future decisions become the next round of training data. A family that resembles the families the agency investigated in 2014 gets flagged in 2026 — not because anyone observed danger, but because the machine recognized the agency's own reflection.
Who sees the number — and who never does
Now trace the score's path through a single case, because its visibility is engineered with the same care as its mathematics. In Allegheny County, the number is generated at the moment a hotline call comes in, and it is displayed to the call screener deciding whether to open an investigation. From that point forward, the research record on these systems shows, the score lives entirely inside the agency. The family that was scored is not told a score exists. The family's lawyer, if the case reaches a courtroom, is not shown it. The dependency judge who will decide whether the children come home rules on a case that was selected for investigation, in part, by a number that appears nowhere in the file before the bench. In some implementations, researchers found, legal professionals are excluded from the loop deliberately.
Investigation data visualization
So, the machine's influence enters at the front door — deciding which families get investigated at all — and then vanishes from every record a family or a court could examine. The decision it shaped is fully visible. The shaping is not. Keep that architecture in mind through everything that follows, because it explains the otherwise inexplicable: how tools this consequential could run for a decade, in multiple states, without a single legislative vote, court challenge on the merits, or published independent audit. It is difficult to contest a number you have never been told exists.
PART II — THE INPUTS: A POVERTY DETECTOR BY DESIGN
What the machine eats
If the target variable is the first deception, the feature list is the second.
The AFST draws on Medicaid claims, public-benefits history, behavioral-health records, drug and alcohol treatment histories, jail and juvenile-probation records, birth records, and prior contact with the child-welfare system. It has at times drawn on Supplemental Security Income data and on recorded diagnoses for mental, behavioral, and neurodevelopmental disorders, including schizophrenia and mood disorders.
Read that list again and notice what every item has in comm on: the county only holds these records for people poor enough to touch public systems.
A middle-class parent who sees a therapist through an employer health plan is invisible to the model. A poor parent who sees a therapist through Medicaid is a data point. A parent with a private disability policy does not exist in the feature space; a parent on SSI does. The model does not measure parenting. It measures surveillance density— how much of your life happens to be recorded in the government's files — and surveillance density in America is a nearly perfect proxy for poverty, and a strong proxy for disability and race.
The Human Rights Data Analysis Group and the ACLU took this from theory to measurement in a 2023 paper presented at the ACM Conference on Fairness, Accountability, and Transparency. Working at the level of the AFST's individual predictors, they found that the tool's "ever involved in juvenile probation" variable affected 27 percent of referrals involving households with at least one Black member — versus 9 percent of non-Black households. That is not a model discovering risk. That is a model rediscovering the carceral history of a neighborhood and scoring it as parenting.
New York City runs the least accountable version of all. In May 2025, the investigative outlet The Markup revealed that the city's Administration for Children's Services has, since 2018, used a risk model with279 variable— including neighborhood, mother's age, number of siblings, and mental-health history — to flag families for heightened scrutiny by its Emergency Response Office. Families are never told the algorithm flagged them. Neither are their lawyers. Neither, in many instances, are the caseworkers acting on the flags. ACS's own internal audit conceded the system contains "implicit and systemic biases" — and concluded it was still better than what came before. This, in a city where Black families are reported to child services at roughly seven times the rate of white families and are about thirteen times more likely to have a child removed. A 279-variable secret score in that system is not a decision aid. It is a shadow docket.
Why the inputs meet the mission
Now put the inputs beside the American definition of "neglect," because the two facts detonate on contact.
Roughly63 percent of removals nationally cite neglect. Of the542,900substantiated child-maltreatment victims in federal fiscal year 2023, about three in four experienced neglects rather than physical or sexual abuse. In Indiana in 2024,87 percent of children entering foster care had parents who were not even accused of physical or sexual abuse. And "neglect," in day-to-day American practice, is frequently a description of poverty: an empty refrigerator, a shut-off notice, an unstable apartment, an untreated illness.
Feed that system a model whose inputs are Medicaid enrollment, SSI receipt, public-benefits history, and behavioral-health diagnoses obtained through public insurance — and whose target is removal. You have not built a child-safety instrument. You have built a poverty detector wired to a foster-care actuator. The rest of this article is about what happened in the places that switched one on.
PART III — THE RECORD: FOUR JURISDICTIONS, FOUR FAILURES
Illinois: 100 percent risk of death, and the children who actually died
In May 2016, the Illinois Department of Children and Family Services signed a$366,000 sole-source contract— no competitive bid — for a predictive tool called Rapid Safety Feedback, marketed by the Florida nonprofit Eckerd (about which much more in Part V) with its for-profit partner MindShare Technology. The system mined DCFS case files and assigned children scores from 1 to 100 representing their probability of death or serious injury.
What it produced, per reporting by the Chicago Tribune and the trade press that covers child welfare, belongs in a textbook — as a warning label. The system assigned more than 4,100 Illinois children a 90 percent or greater probability of death or serious injury. It rated369 children under the age of nine at 100 percent— a literal certainty of death or serious harm. Caseworkers drowned in alarms that meant nothing. And while the machine was crying wolf four thousand times over, it failed to flag the children who actually died.
Investigation data visualization
DCFS Director B.J. Walker shut the program down in December 2017, and her explanation deserves to be carved above the door of every agency that filed an application before this July's midnight deadline: predictive analytics wasn't predicting any of the bad cases.
One more detail from Illinois, because it previews Part V. The no-bid contract had been awarded under then-DCFS Director George Sheldon— who came to Illinois from Florida's Department of Children and Families, the state where Eckerd held a lead-agency contract worth roughly$73 million, and who resigned in June 2017 amid controversy over the steering of contracts to former associates.
Los Angeles: 171 hits, 3,829 false alarms
Los Angeles County's first attempt was called Project AURA, built by the software firm SAS and tested against historical cases from 2015 to 2017. In testing, it correctly identified171 children at highest risk of harm. It also flagged3,829 false positives— families the system marked for the highest level of concern who did not belong there. That is roughly22 false alarms for every child correctly identified. The county's Department of Children and Family Services quietly abandoned AURA in 2017.
Then it came back. In August 2021, DCFS launched a new "Risk Stratification Model" pilot — built with the academic network described in Part IV — in its Belvedere, Lancaster, and Santa Fe Springs offices, where it operates today.
Oregon: terminated weeks after the AP came calling
Oregon's Department of Human Services built its Safety at Screening Tool explicitly on the Allegheny model. It was terminated at the end of June 2022— weeks after an Associated Press investigation by reporters Sally Ho and Garance Burke exposed the racial patterns in the Allegheny tool Oregon had copied. Oregon reverted to Structured Decision Making, an older actuarial framework with a property the algorithms lack: its factors are printed on a form a human being can read, question, and challenge.
Pittsburgh: the tool that stayed
And in Allegheny County itself — the birthplace, the showcase, the model jurisdiction cited in the federal funding notice's intellectual lineage — the tool remains in daily use, ten years on, with no authorizing statute, no parental right to see or contest a score, no disclosure in dependency court, and an open federal civil-rights inquiry that has gone silent. Parts IV and VII take Pittsburgh apart in detail.
Hold the scoreboard in your head as you go: one tool killed by its own director for not predicting the bad cases. One abandoned after a 22-to-1 false-alarm ratio. One terminated under investigative scrutiny. One under open federal inquiry. That is the county-level track record the federal government now proposes — in writing — to scale to entire states.
PART IV — THE BUILDERS: WHEN THE INVENTORS GRADE THE INVENTION
Two academics, four jurisdictions
Most of American child-welfare risk scoring traces to two people. Rhema Vaithianathan, a health economist at Auckland University of Technology and director of the Centre for Social Data Analytics, and Emily Putnam-Hornstein, formerly of the University of Southern California and now at the University of North Carolina at Chapel Hill, where her work is tied to the Children's Data Network.
Together or through their teams, they built the Allegheny Family Screening Tool. They built Allegheny's Hello Baby newborn model. Vaithia nathan's team-built Douglas County, Colorado's Decision Aid. The Children's Data Network led development of Los Angeles County's Risk Stratification Model — the one now running in three DCFS offices. Officials in Pennsylvania, California, and Colorado opened their integrated government data systems to the two developers, who then selected which data points would go into the models that score those states' families.
This is not, by itself, an indictment. Somebody has to build things. The indictment is what happened next.
The conflict, stated plainly
The federal government funded a national study of predictive risk modeling in child welfare —co-authored by Vaithia nathan and Putnam-Hornstein— which concluded that their own Allegheny approach could serve as a model for other jurisdictions. The federal government paid the tool's inventors to evaluate whether the tool should spread. It spread.
The same pattern runs through the evidence base that agencies now cite in the tools' defense:
Investigation data visualization
The most-cited finding favorable to the AFST — that it reduced unconditional racial disparities in screening by 31 percent, and score-conditional disparities by 48 percent — comes from a paper co-authored by Putnam-Hornstein and Vaithia nathan. Two of its authors are the tool's builders.
Allegheny County's standard answer to critics is a Stanford evaluation finding no adverse consequences — an evaluation the county itself commissioned.
The most damaging independent finding comes from researchers at Carnegie Mellon University, who examined every referral in the county from August 2016 to May 2018. Their result: had call screeners followed the AFST's recommendations, the Black–white disparity in screened-in cases would have been 20 percent. The actual disparity was 9 percent— because the human screeners overrode the algorithm, disagreeing with its scores roughly a third of the time. Sit with that: the humans were the de-biasing mechanism. The machine was the bias.
And the Human Rights Data Analysis Group and ACLU found that screen-in rates in Allegheny did not drop after the AFST's 2016 introduction — meaning the tool's central promise, that it would help the county investigate fewer families more accurately, is not visible in the county's own numbers a decade later.
We include the favorable findings, name their authors, and let the reader weigh them — that is what a transparency organization does. But an evidence base in which the strongest supporting studies are written by the inventors and commissioned by the customer, while the strongest critical studies come from independent researchers, is not an evidence base. It is a brochure with footnotes.
PART V — THE VENDOR: THE $350 MILLION NONPROFIT AND THE TOOL THAT DIDN'T WORK
Eckerd, by its own filings
The commercial engine of the field's first wave was Eckerd— legally Eckerd Youth Alternatives, Inc., doing business as Eckerd Connects, EIN 59-2551416, of Clearwater, Florida — which marketed Rapid Safety Feedback nationally with its for-profit partner MindShare Technology.
Eckerd is not a church basement. Its own Form 990 filings — verified against the nonprofit financial records in Project Milk Carton's public-data archive — show revenue of $351.9 million in FY2021(3,480 employees), $312.3 million in FY2022(3,364 employees), and $164.2 million in FY2023, with assets of $94.5 million. Aggregate officer and director compensation exceeded$3.3 million in its peak recent year. The two-year revenue drop — a 53 percent collapse from FY2021 to FY2023 — coincides with the loss of Florida lead-agency child-welfare contracts, the same kind of contract that once made Eckerd a roughly $73 million vendor to the State of Florida.
The evidence base, tested once
Rapid Safety Feedback's origin story, as marketed, was Hillsborough County, Florida: after RSF launched there, Eckerd said, no child receiving its in-home services died of abuse. That claim is a count of deaths inside Eckerd's own client population — not a controlled comparison of any kind. It is the sort of statistic that sounds like science and functions as advertising.
The tool's one independent statewide test came in 2022, when a peer-reviewed study in Child Abuse & Neglect examined a full statewide implementation of Eckerd Rapid Safety Feedback and found no positive effect on repeat maltreatment. None. By then, variants of RSF had already spread to Ohio, Indiana, Maine, Louisiana, Tennessee, Connecticut, Oklahoma, and Alaska— and had already been purchased, sole-source, by Illinois, where it scored 369 small children at a 100 percent probability of death or serious injury while missing the children who died.
One more line belongs in this file, carefully stated. In February 2024, Florida's Second District Court of Appeal decided McGee v. Eckerd Youth Alternatives, Inc. d/b/a Eckerd Connects (No. 2D22-4080) — an appeal brought by the personal representative of the estate of Kwon McGee, a deceased minor. We have verified the case's caption, court, and date from the docket; we have not yet obtained the opinion's full text, so we make no claim here about the underlying facts or the outcome. But the existence of wrongful-death litigation against the company, involving a deceased child, sits in the public record alongside a marketing pitch built on the absence of child deaths — and readers deserve to know both facts exist.
PART VI — THE MONEY: PLUMBING NOBODY VOTED ON
Two pipes, no line item
There is no line item in federal law called "predictive analytics." There does not need to be. The tools have been paid for all along through two ordinary reimbursement channels — and that ordinariness is precisely what makes them invisible.
Pipe one: Title IV-E administrative match. Section 474 of the Social Security Act reimburses states at roughly50 percent for the administrative costs of operating their child-welfare programs. Risk scoring is, accounting speaking, administration. Building a model, running it, and paying the analysts who maintain it are administrative costs. Half the bill goes to Washington, and no one ever has to ask Congress for an algorithm budget.
Pipe two: CCWIS. The Comprehensive Child Welfare Information System rule — 45 C.F.R. Part 1355, effective 2016 — offers states a50 percent federal match on both the development and the ongoing operation of a modern child-welfare data system, against a far lower rate for legacy systems. That differential is not a detail; it is the engine. A modern integrated child-welfare data system is a $100-million-plusundertaking, and few states can absorb a nine-figure build without the enhanced federal rate — which means the CCWIS match functions less like a subsidy and more like a steering wheel. States build what Washington will pay half of.
And what CCWIS pays half of is the plumbing: the integration of Medicaid, behavioral health, courts, and benefits data into a single query able spine. Every dataset described in Part II — the Medicaid claims, the SSI records, the probation files, the diagnoses — becomes reachable from one place because the federal government paid to connect them. The connection happens under federal supervision, too: states seeking the match file Advance Planning Documents with the Children's Bureau describing what they intend to build, and the Bureau reviews and approves them, document by document. And once that spine exists — federally subsidized, federally reviewed — the marginal cost of bolting a predictive model on top is trivial. The expensive part of a risk-scoring system was never the model. It was the surveillance infrastructure underneath it, and that part was bought with federal money, on federal sign-off, years before anyone said the word "algorithm" out loud.
Follow the logic of that arrangement to its end. The Children's Bureau reviews and approves the architecture of these data systems, document by document, through the APD process — while maintaining that it has no role in reviewing the models built on top of them. The federal government pays to construct the data warehouse, inspects the warehouse's blueprints, and then averts its eyes from what gets built inside. That is not oversight. That is laundering — and until this summer, it was at least passive laundering.
Investigation data visualization
On June 2, 2026, it became active. That is Part VIII. But first, the question the federal government has been refusing to answer since 2022.
PART VII — THE DISABILITY QUESTION, AND THE INQUIRY THAT WENT QUIET
The complaints
In the fall of 2022, civil-rights complaints were filed with the U.S. Justice Department's Civil Rights Division alleging that the Allegheny Family Screening Tool discriminates against parents with disabilities, in violation of Title II of the Americans with Disabilities Act. The complaints did not arrive out of nowhere: Associated Press reporting documented that DOJ attorneys had themselves cited the AP's investigation of the tool in urging complainants to file formally.
The alleged mechanism is not exotic. The National Council on Disability has documented that parents with disabilities receive public benefits — SNAP, Medicaid, SSI — at high rates. A model that ingests benefit receipt and behavioral-health records will penalize disability whether or not anyone typed the word "disability" into the code. The currently deployed version of the tool —AFST V3— includes features reflecting recorded diagnoses of behavioral and mental-health disorders that are recognized disabilities under the ADA.
The science caught up on June 17, 2026
For years that mechanism was an inference. On June 17, 2026, it became a finding. The Journal of Public Child Welfare published "Evaluating disability bias in the Allegheny Family Screening Tool," a peer-reviewed study that tested the question empirically using the American Community Survey, the Medical Expenditure Panel Survey, and the National Survey on Drug Use and Health. Its conclusion: parents and children with disabilities were significantly more likely to be penalized by every element of the AFST scoring examined— socioeconomic indicators, healthcare utilization, public-benefits involvement, and criminal-legal contact. In the authors' words: "Despite containing no explicit disability measures, the AFST effectively encodes disability status through proxy variables."
The silence
And the federal inquiry? We re-searched the public record for this investigation, in July 2026. As of publication: no public findings. No letter of findings. No settlement. No enforcement action. The most recent substantive reporting on the inquiry dates to January 2023. Three and a half years after the complaints were filed, the tool remains in daily use on every family reported to Allegheny County's hotline — and the Civil Rights Division has said nothing in public.
Three and a half years of silence is itself a finding. It means that the one federal watchdog that ever stirred on this subject went quiet — during the same season in which a different arm of the same federal government, the Administration for Children and Families, decided the technology deserved $6 million and a national scaling program.
Which brings us to the document at the center of this investigation.
PART VIII — THE CHECK: READING HHS-2026-ACF-ACYF-CA-0037 LINE BY LINE
What the government announced
On June 2, 2026, the Administration for Children and Families announced $6 million in competitive funding to help state, territorial, and tribal child-welfare agencies pilot predictive analytics. ACF Assistant Secretary Alex J. Adams framed it this way: "Child welfare caseworkers are tasked with making high-stakes decisions about child safety, often under significant pressure and without complete information... Promising use cases have been tested at county levels, and we want to support scaling these interventions at the state level to improve child welfare outcomes."
Read that sentence against Parts III through VII. The "promising use cases tested at county levels" are Allegheny — under an open, silent DOJ civil-rights inquiry — and Los Angeles, whose first attempt was abandoned at 22 false alarms per hit. The county-level track record the Assistant Secretary invokes includes a kill in Illinois, an abandonment in Los Angeles, a termination in Oregon, and an unresolved federal disability complaint in Pennsylvania. The initiative builds on an ACF roundtable convened in December 2024 and a March 5, 2026, ACF brief titled Modernizing Child Welfare Technologies and Tools.
We obtained the full funding notice — the NOFO, as modified June 12, 2026 — from the federal grants system and read it line by line. The mechanics first, because they matter to who is bound by what: the program runs under Assistance Listing 93.670, Child Abuse and Neglect Discretionary Activities. The awards are cooperative agreements, not grants — a legal form in which the federal agency stays substantially involved in the funded work. The project period is three years, with no cost sharing required of the states: Washington pays the whole freight. Eligibility is restricted to state, territorial, and tribal child-welfare agencies — the counties that ran the original experiments cannot apply; only governments big enough to scale them can. And the NOFO situates itself in a policy lineage, citing alignment with three executive orders: EO 13930, "Strengthening the Child Welfare System for America's Children"; EO 13960, "Promoting the Use of Trustworthy Artificial Intelligence in the Federal Government"; and EO 14359, "Fostering the Future."
That last citation deserves one beat of attention. The federal government's own trustworthy-AI framework is being invoked as the banner over a program whose subject tools are, as this file documents, unauthorized by any deployment statute, invisible to the families they score, and unexamined by any required independent audit. The word "trustworthy" is doing the work that an audit requirement would otherwise have to do.
Investigation data visualization
Five findings from the government's own document:
1. It solicits the exact uses that already failed
The NOFO's "areas of particular interest" include predictive analytics "to help make decisions consistently and accurately during hotline intake"; systems to "flag high-risk cases during investigations"; models "generating risk scores, forecasting maltreatment"; identifying candidates for Title IV-E prevention services; and analytics "to identify, target, and retain foster and adoptive placements." Hotline-intake screening is precisely the Allegheny use case — the one under federal inquiry. It is precisely the Oregon use case — terminated. Flagging high-risk cases is precisely the Illinois use case — killed by the director for not predicting the bad cases. The wish list is the graveyard.
2. The frame is foster-home supply, not family preservation
The NOFO's rationale states the aim of achieving "a1:1 ratio of available foster homes to children in need," and ACF gives priority consideration to applicants who join, or pledge to join, its "A Home For Every Child" initiative. Read plainly: the federal frame is placement-supply matching — sorting children toward available beds faster — not keeping families together. A machine tuned to that objective is a machine tuned to remove.
3. The bias-audit contradiction is written into the funding terms
The tools' central documented failure mode — per the FAccT 2023 HRDAG/ACLU findings, the Carnegie Mellon override study, and the June 17, 2026, disability-bias study — is demographic bias. Now read the NOFO's terms. Its required monitoring covers model recalibration, thresholds, and clinical overrides. It nowhere requires an independent bias audit. And its funding restrictions state that grantees "may not use funds from this NOFO for any diversity, equity, inclusion, and accessibility (DEI and DEIA) activities" — explicitly including "DEI- or DEIA-related research." We state this precisely, as a documented tension in the notice's own text: the NOFO does require engaging families and community stakeholders in design and oversight "with a focus on transparency," and it does not ban states from auditing their tools with other money. But a funding program whose subject's best-documented defect is demographic bias, whose terms do not require a bias audit, and whose restrictions arguably foreclose spending its own funds on bias-focused research, has designed itself not to find the problem.
4. The federal hands are on the wheel this time
These are cooperative agreements, not arm's-length grants: the NOFO states the Children's Bureau will be "reviewing and approving the planning stages," coordinating implementation, and "providing close oversight." Remember Part VI — the same federal office that has long insisted it has no role in reviewing predictive models will now contractually approve these ones. The laundering argument is no longer an inference. It is a signature line.
5. The legal authority is for studying, not scoring
The NOFO cites as its statutory authority the Child Abuse Prevention and Treatment Act —42 U.S.C. § 5106(b)(5), Congress's discretionary research-and-demonstration authority. That authority is real, and it means "no statute is involved" would be sloppy. So here is the precise claim, and it is worse than sloppy would-be Congress's research authority lets HHS fund demonstrations. *No federal or state statute authorizes deploying algorithmic risk scores against specific families, defines which government records may be ingested to score a parent, or grants any family the right to see or contest the number.* The gap between funding authority and deployment authority is where every family in this article lives. Washington has clear legal authority to pay for the machine. No legislature has ever authorized the machine to knock.
The first application window under this notice closed at 11:59 p.m. Eastern on July 13, 2026. Awards — up to ten of them, $400,000 to $600,000 each, three-year cooperative agreements — are expected around September 30, 2026, with further application cycles due June 30, 2027 and June 30, 2028.
PART IX — THE ACCOUNTABILITY GAP: NO ONE WHOSE JOB THIS IS
Stack every prior section and the shape of the hole becomes exact. At each place where American government normally installs a check, this field has none.
No legislature authorized the scoring. These tools were not created by statute; they were procured. CAPTA requires states to have screening procedures — it says nothing about algorithms. Title IV-E and CCWIS fund them without naming them. No elected body in Pennsylvania, New York, Colorado, or California has ever voted on whether an SSI record may be used to decide if a family gets investigated.
No court ever sees the score. Research on these systems finds they are used almost entirely internally by CPS staff; the scores generally never surface in dependency hearings, and in some implementations legal professionals are deliberately excluded from the loop. Set that against the constitutional baseline, because the baseline is not obscure. In Stanley v. Illinois (1972), the Supreme Court held that parents hold a fundamental liberty interest in the care and custody of their children. In Santosky v. Kramer (1982), it held that the state must prove parental unfitness by clear and convincing evidence before severing that bond — because the risk of error must fall on the government, not the family. In Troxel v. Granville (2000), it reaffirmed that fit parents' decisions carry constitutional weight. Half a century of doctrine, one consistent premise: when the state moves against a family, the family gets to see and test the state's case.
Now run an algorithmic referral through that doctrine. The evidence a parent would need to test includes a score they were never shown, produced by a model whose feature list the county has never fully published, computed from records — a Medicaid claim, an old probation entry, a diagnosis — that the parent has no mechanism to inspect or correct. There is no motion a parent can file against a number that appears in no filing. The confrontation right doesn't fail here; it simply never engages, because the thing to be confronted was designed to stay out of the record. Douglas County, Colorado — which says it will share scores with families who ask — is the outlier that proves the rule: disclosure is evidently possible, and nearly nowhere practiced.
No independent audit is required.Here is the comparison that ends the argument. The Family First Prevention Services Act built an entire federal apparatus — the Title IV-E Prevention Services Clearinghouse — to rate whether aparenting classis sufficiently evidence-based to merit federal reimbursement. A statistical model that decides whether the state investigates your family facesno such review at all. The evaluations that exist are, as Part IV showed, largely commissioned by the deploying agencies or authored by the builders. The federal government demands more proof from a home-visiting curriculum than from the algorithm that dispatches the visit.
Investigation data visualization
And the one watchdog that stirred went silent— while the funding agency wrote a check. That asymmetry is the entire regulatory posture of the United States toward child-welfare algorithms, as of this writing: the enforcement arm says nothing; the spending arm saysscale it.
PART X — WHAT WOULD FIX IT
None of this requires new science. It requires four decisions, each of which an elected body or a federal office could make this year.
1. Statutory authorization or nothing.No jurisdiction should score a family using government records without a law, passed by an elected legislature, specifying which records may be used and for what purpose. If the tool cannot survive a floor vote, it should not survive procurement.
2. Disclosure as a condition of use.If a score touches a screening or removal decision, the score, its inputs, and the model documentation must be disclosed to the parent and counsel, and be admissible for challenge in any dependency proceeding. A number that cannot survive cross-examination should not be generated by a government.
3. Ban the poverty and disability proxies.Medicaid enrollment, SSI receipt, SNAP, and behavioral-health diagnoses obtained through public insurance should be categorically excluded as model features. Their predictive power comes from surveillance density, not from risk — and the June 17, 2026 study shows exactly what they encode.
4. Condition the federal money.ACF holds the leverage and needs no new law to use it: make Title IV-E administrative match and CCWIS reimbursement contingent on independent, published, pre-deployment and annual bias audits — the same evidentiary standard the Prevention Services Clearinghouse already applies to a parenting curriculum. The office that approves the planning stages can require the audit in the same signature.
EPILOGUE: AFTER MIDNIGHT
The envelope sealed at 11:59 p.m. on July 13, 2026 will open, in effect, around September 30, when the Administration for Children and Families announces up to ten awards. Ten state, territorial, or tribal agencies will then begin building, with federal money and Children's Bureau sign-off, statewide versions of the machines this file documents. Two more application windows follow — June 30, 2027, and June 30, 2028. This was the first round, not the last.
Which means the awards themselves are now the story. Project Milk Carton will be watching for the award announcements, will identify every recipient agency, will file for the Advance Planning Documents and cooperative-agreement terms that federal law makes public, and will report what each state proposes to feed its model and who, if anyone, is allowed to audit it. The applications were sealed. The awards will not be.
The 2027 window matters for a different reason: it is the one that can still be shaped. Every fix in Part X — statutory authorization, disclosure, proxy bans, audit conditions — could be written into the next notice by the same office that wrote this one, without a single act of Congress. The Children's Bureau will hold the pen again in June 2027. Between now and then, the public record this investigation assembles is the argument for what that pen should write.
Until then, the arrangement this investigation documents stands exactly as the record left it: a family in Pittsburgh, or Denver's southern suburbs, or a New York City borough can be investigated because a model they have never heard of noticed they were poor, disabled, and previously reported — and the federal government, having quietly paid to build the pipes, has now put $6 million behind running more water through them.
The score is still invisible. The knock is not.
If you enjoyed this work and feel encouraged by independent investigations like this, please consider becoming a paid subscriber to The Constitutional Republic on Substack or donate at
Project Milk Carton Inc. Your gift keeps our site alive, spotlighting high-risk children! or at https://t.co/PvtnrODkNz
Your support helps us continue:
investigative reporting
public records research
educational transparency projects
community outreach
and future investigations like this series
Most importantly, please share this article.
Independent journalism survives when ordinary people help spread information powerful institutions would prefer to remain ignored, misunderstood, or buried beneath complexity.
Because silence protects systems.
Informed communities protect children.
ALL paid subscriptions are tax deductible Project Milk Carton | 501(c)(3) | EIN: 33-1323547
A NOTE ON SOURCES — AND WHAT WE COULD NOT CONFIRM
This investigation is built on: the NOFO for HHS-2026-ACF-ACYF-CA-0037 (as modified June 12, 2026), retrieved from the federal grants system; the ACF announcement of June 2, 2026 and its March 5, 2026 brief Modernizing Child Welfare Technologies and Tools; Associated Press investigative reporting by Sally Ho and Garance Burke on the Allegheny Family Screening Tool and the DOJ inquiry; The Markup's May 2025 investigation of New York City's ACS model; the ACLU's Family Surveillance by Algorithm(2021) and the ACLU/HRDAG paper presented at ACM FAccT 2023; the Carnegie Mellon override study of Allegheny referrals (2016–2018); "Evaluating disability bias in the Allegheny Family Screening Tool," Journal of Public Child Welfare, June 17, 2026; the 2022Child Abuse & Neglect study of a statewide Eckerd Rapid Safety Feedback implementation; reporting of record on the Illinois, Los Angeles, and Oregon programs; the CCWIS final rule and 45 C.F.R. Part 1355; federal court dockets; and Eckerd Youth Alternatives, Inc. Form 990 filings (EIN 59-2551416) verified in Project Milk Carton's public-data archive.
In the same spirit, what we could not confirm, and how we handled it:(1) We found no public findings, settlement, or enforcement action from the DOJ Civil Rights Division inquiry; the most recent substantive reporting is from early 2023, and we describe the inquiry only as open and silent. (2) The strongest evidence favorable to the AFST — a 31–48 percent disparity-reduction finding and a Stanford evaluation — is presented in Part IV with its provenance: two of the favorable study's authors built the tool, and the county commissioned the 3) We cite McGee v. Eckerd only by caption, court, and date; we have not read the opinion and make no claim about its facts or outcome. (4) Figures for Eckerd's federal HHS awards ($35.2 million in 2019; $14.6 million in 2025) come from an earlier records pull we could not re-verify this session and are therefore omitted from the body of this article. (5) The number of agencies that applied before the July 13 deadline is not public; we describe the window as closed and the awards as pending, and claim nothing more.
Project Milk Carton is a 501(c)(3) transparency and public-data organization. This article is public-record research, not legal advice. Corrections: if any figure or citation here is wrong, we will say so, prominently.
Project Milk Carton
501(c)(3) Nonprofit • projectmilkcarton.org • t.me/ProjectMilkCarton