OSINT · Publicly Available Sources INVESTIGATION ORACLE

Scored Before They're Heard: How Federal Child-Welfare Dollars Built a Secret Risk-Rating System Between Parents and Their Children

September 08, 2026 OPUS · Claude Opus Project Milk Carton

Scored Before They're Heard: How Federal Child-Welfare Dollars Built a Secret Risk-Rating System Between Parents and Their Children

In at least eleven states and a growing list of counties, a parent accused of neglect is assigned a number before a single fact of the allegation is verified — a predictive risk score, computed in seconds from Medicaid claims, jail records, public-benefits enrollment and prior hotline calls, displayed to a hotline screener as they decide whether to send an investigator to the door. The parent is never told the score exists. Their attorney cannot obtain it in discovery. No juvenile court in the country has subjected one to a Daubert reliability hearing. Independent researchers who have gotten inside these models — at Carnegie Mellon, the Human Rights Data Analysis Group, and most recently in the Journal of Public Child Welfare — have found they flag Black children and disabled parents at multiples of the base rate, and the Justice Department's Civil Rights Division opened a disability-discrimination inquiry into the most famous one, Allegheny County's Family Screening Tool, in 2023. And in June 2026, with that inquiry still unresolved and no public findings ever issued, the Administration for Children and Families sent $6 million to ten more jurisdictions to build their own.

The setup: the score arrives before the story

The mechanism is deceptively mundane. A mandated reporter — a teacher, an ER nurse, a landlord, a relative — calls a state child-abuse hotline. Before the screener finishes intaking the narrative, a predictive risk model queries an integrated data warehouse for every adult and child named on the call, and returns a single number.

In Allegheny County, Pennsylvania, that number is the Allegheny Family Screening Tool score: an integer from 1 to 20 estimating the probability that the child will be placed in out-of-home care within two years. It has been running on live calls since August 2016. A score at or above the threshold triggers a "mandatory screen-in" — the referral goes to investigation regardless of the screener's own read of the allegation. The threshold was initially set at 18; the county later lowered it to 16. The Human Rights Data Analysis Group documented that a January 2019 recalibration was projected to roughly double the share of referrals hitting the high-risk protocol, from about 11.5 percent to about 22 percent; observed prevalence came in around 24 percent. That is a policy change of enormous consequence — thousands of additional families investigated — enacted by adjusting a cutoff, with no rulemaking, no hearing, and no vote.

New York City's Administration for Children's Services runs a quieter version. The Markup revealed in May 2025 that ACS has used a predictive model since 2018 that scores families on 279 variables, trained on cases from 2013 and 2014 that ended in a child being severely harmed. High scorers are routed to the agency's Office of Child Safety for heightened scrutiny. Parents are not told. Their lawyers are not told. According to The Markup's reporting, in many instances the assigned caseworkers were not told either.

What the model is actually learning

The central defect is not the math. It is the label.

None of these models are trained on substantiated harm to a child, because that outcome is rare, poorly measured, and inconsistently defined across jurisdictions. They are trained on prior system contact — a proxy that is abundant, cleanly recorded, and thoroughly contaminated. Allegheny's tool predicts future placement, which is itself a discretionary agency decision. NYC's predicts a rare severe-harm outcome using a decade-old snapshot of who the agency happened to investigate.

The input variables come from the same well. The AFST draws on the county's data warehouse: Medicaid claims, behavioral-health and substance-use treatment records, jail and probation records, public-benefits enrollment, birth records, and prior CYF referrals. Every one of those is a measure of poverty and public-sector dependence, not of parenting. A family with private insurance, a private therapist, and money for a lawyer generates almost no signal. A family on Medicaid with a mother in outpatient behavioral health and a father with an old probation case generates a dense one. Virginia Eubanks named this in Automating Inequality in 2018: the model does not predict child abuse. It predicts visibility to government.

The disparate results follow mechanically. Carnegie Mellon researchers analyzing AFST scores from August 2016 through May 2018 — findings first reported by the Associated Press in April 2022 — found the tool recommended mandatory investigation for 32.5 percent of Black children reported as neglected, versus 20.8 percent of white children. The same researchers found human screeners disagreed with the algorithm's score about one-third of the time.

The disability finding is starker still. A study published in the Journal of Public Child Welfare on June 17, 2026 — "Evaluating disability bias in the Allegheny Family Screening Tool" — concluded that parents and children with disabilities were significantly more likely to be penalized by every element of the scoring examined: socioeconomic indicators, healthcare utilization patterns, public-benefits involvement, and criminal-legal contact. The AFST contains no explicit disability variable. It doesn't need one. Disability is fully encoded in the proxies. A disabled parent uses more Medicaid services, receives more benefits, and has more documented agency contact — and the model reads all of that as risk.

The money: a 50 percent match, no questions asked

This is not a story about counties buying software with their own money. It is a story about a federal reimbursement stream with no substantive quality condition attached.

Two federal pipes fund these systems, and neither requires an accuracy threshold or a bias audit:

Title IV-E administrative match. Section 474(a)(3) of the Social Security Act and 45 CFR § 1356.60(c) make federal financial participation available at 50 percent for administrative expenditures "necessary for the proper and efficient administration" of a state's Title IV-E plan. Section 474(a)(3)(E) fixes that rate. Longstanding ACF policy treats data collection system initiation, implementation, and operation as an allowable administrative cost at the 50 percent rate. A predictive model sitting inside the hotline workflow is, on paper, a data collection function. Half the bill goes to Washington.

CCWIS. The Comprehensive Child Welfare Information System rule, finalized June 2, 2016 (proposed August 11, 2015) and codified at 45 CFR Part 1355, replaced the old SACWIS framework. CCWIS-eligible costs also draw 50 percent Title IV-E FFP. Approval runs through the Advance Planning Document process under 45 CFR Part 95, Subpart F — a state submits a planning, implementation, annual, or operational APD, and ACF's Children's Bureau approves the federal cost allocation.

The APD is the choke point that was never used as one. It is where a federal reviewer reads exactly what a state proposes to build, what data it will ingest, and what it will cost — and where a bias audit, a calibration report, a false-positive ceiling, or a disclosure requirement could be made a condition of the match. None of those has ever been required.

The scale of the underlying spend is now documented by HHS itself. A June 2026 ACF/ASPE assessment found that after roughly a decade and more than $2.2 billion in federal CCWIS claims, implementation shows "minimal progress": fewer than a third of CCWIS projects are operational, 15 states have not declared any new CCWIS project, and transitional legacy systems account for 85 percent of total claims — $1.9 billion. Two billion dollars of federal child-welfare IT money has flowed with so little outcome accountability that HHS's own analysts cannot say what it bought. That is the compliance culture into which predictive risk modeling was inserted.

The 2026 escalation

On November 13, 2025, President Trump signed the executive order "Fostering the Future for American Children and Families," announced alongside First Lady Melania Trump. Among its directives to HHS: expand states' use of technological solutions in child welfare, explicitly including predictive analytics and tools powered by artificial intelligence.

ACF moved fast. On March 5, 2026, it published an issue brief, Modernizing Child Welfare Technologies and Tools: Opportunities for Predictive Risk Modeling to Improve Child Safety and Outcomes, which instructs states to identify a use case, assess their administrative data, establish governance, and — the operative sentence — "determine the appropriate Title IV-E funding and technology pathway." ACF simultaneously clarified that eligible PRM costs can draw Title IV-E support through CCWIS or administrative cost-allocation rules. Translation: the 50 percent match is open for algorithms, and here is the paperwork.

Then came the funding. Notice of Funding Opportunity HHS-2026-ACF-ACYF-CA-0037, "Predictive Analytics in Child Welfare Demonstration Grants," applications due July 13, 2026: $400,000–$600,000 per award, up to $6 million, three-year project periods, roughly ten awards. ACF received 21 applications from 20 jurisdictions and funded ten: Indiana, Missouri, Nebraska, New Jersey, North Carolina, Ohio, Oklahoma, Texas, the Muscogee (Creek) Nation, and the District of Columbia. The funded work includes machine learning, natural language processing, automated case summaries, text analytics, cross-agency data integration, and predictive models embedded directly within CCWIS — meaning the score stops being a bolt-on and becomes native to the case record.

In Congress, Senator Todd Young (R-Ind.) has introduced the Using Data to Help Protect Children and Family Act, which would put $10 million behind an HHS pilot with five states or tribes. It sits in the HELP Committee.

Nothing in the ACF brief, the NOFO, or the executive order requires a published false-positive rate, a calibration curve by race or disability, an external audit, or notice to the families scored.

The named players and the incentive structure

The AFST was built under a December 2014 partnership between Allegheny County DHS and AUT Enterprises Ltd., the commercial arm of Auckland University of Technology, led by Professor Rhema Vaithianathan of AUT's Centre for Social Data Analytics, with Emily Putnam-Hornstein (then USC, now UNC-Chapel Hill). The same two researchers have since advised or built models for multiple U.S. jurisdictions and have described deploying real-time risk models at three organizational levels — hotline screeners, caseworkers and supervisors, and follow-up review. Putnam-Hornstein is also a lead voice in the FREOPP whitepaper Recognize: Modernizing Child Protection Through Predictive Risk Modeling, which supplied intellectual cover for the 2025–26 federal push. The circle is small: the people who build the tools write the policy literature recommending the tools, and then evaluate the tools.

The county's own validation is thinner than it appears. Allegheny commissioned an independent impact evaluation from Stanford (PI Jeremy Goldhaber-Fiebert), completed March 2019, comparing ~31,000 children referred pre-implementation (Jan. 1, 2015–July 31, 2016) with ~34,000 post (from Dec. 1, 2016). It reported a moderate accuracy gain and reduced racial disparity in case openings. But read the outcome definition: "accuracy" for a screened-in report meant that the agency took further action — opened a case, connected to an existing case, or received another referral within 60 days. The measured improvement was roughly 358 to 381 children per month, about 24 additional children. The model is scored on whether it successfully predicted what the agency did next. An algorithm that increases investigations, and is graded on whether investigations occurred, will always validate.

On the vendor side, Eckerd Connects (legally Eckerd Youth Alternatives, Inc., EIN 59-2551416, Clearwater, Florida) marketed Eckerd Rapid Safety Feedback nationally with its for-profit partner MindShare Technology. RSF was adopted statewide in Florida and picked up by Oklahoma, Maine, and Connecticut. Eckerd's Form 990 filings show revenue of roughly $351.9 million (FY2021), $312.3 million (FY2022), and $164.2 million (FY2023) — a 53 percent collapse coinciding with the loss of Florida lead-agency contracts; a 2021 Florida Inspector General report found the organization exceeded the state cap on executive salaries. SAS Institute built Los Angeles County's model. The market is a mix of universities monetizing methods, nonprofits with nine-figure government revenue, and enterprise analytics vendors — all of whom assert trade-secret protection over model weights and validation results. FOIA requests filed through MuckRock to Colorado and Virginia seeking exactly that — data sources, model weights, validation testing, disparate-impact studies — illustrate the standoff; HHS's own ASPE/MITRE guidance, Considerations in Contracting Vendors for Predictive Analytics, warns agencies about precisely this loss of control, and is advisory only.

The documented failures

Illinois. DCFS paid $366,000 for Eckerd/MindShare's Rapid Safety Feedback, which scored children 1–100 on the risk of death or serious injury within two years. A joint report by the Office of Executive Inspector General and the DCFS Inspector General concluded that then-Director George Sheldon and DCFS committed mismanagement by structuring the no-bid arrangement as a "grant." The Chicago Tribune found 4,100 Illinois children assigned a 90 percent or greater probability of death or injury — an obvious calibration failure, since nothing close to 4,100 children died — while the system failed to flag several children who subsequently died after maltreatment reports. Director Beverly "B.J." Walker ended it in December 2017: "Predictive analytics [wasn't] predicting any of the bad cases."

Los Angeles County. DCFS and SAS built Project AURA (Approach to Understanding Risk Assessment) and tested it retrospectively on past reports. It correctly identified 171 highest-risk children — and produced 3,829 false positives, a 95.6 percent false-positive rate. The county quietly dropped it. AURA was never run on a live case; had it been, roughly 22 families would have been subjected to heightened intervention for every one correctly identified.

Oregon. The state's Safety at Screening Tool, modeled on Allegheny's and launched in 2018, was dropped in June 2022, weeks after the AP investigation, and replaced with the Structured Decision Making model. Oregon had restricted inputs to internal child-welfare data and applied a deliberate "fairness correction" — and abandoned it anyway. Senator Ron Wyden, who had pressed Oregon DHS on racial bias, put it plainly: "Making decisions about what should happen to children and families is far too important a task to give untested algorithms."

Three jurisdictions ran these systems and shut them down after seeing the internals. The federal response in 2026 was to fund ten more.

The courtroom black box

The legal architecture is the most damning part, because it is not a gap that emerged accidentally — it is a gap nobody has closed.

Lauren and Andrew Hackney of Pittsburgh, both with developmental disabilities, brought their 8-month-old daughter to a children's hospital after struggling to feed her. Hospital staff alerted Allegheny County DHS; the child was severely dehydrated and malnourished. The county removed her. More than a year later she remained in foster care. The Hackneys could not challenge whatever score their family received, because the county would not disclose it — and neither the county nor the tool's authors have explained which variables were used to measure them as parents.

Robin Frank, a veteran Pittsburgh family-law attorney, filed a Justice Department complaint on behalf of a client with an intellectual disability. She has described a judge demanding to know a family's score and the county resisting — on the stated ground that it did not want the number influencing the proceeding. That reasoning is self-refuting: the number already influenced the proceeding, at the only moment that mattered, when it helped decide the case would exist at all.

Follow the evidentiary chain. The score is not offered as evidence; the investigation it caused is. What reaches the judge is a caseworker's affidavit, a home visit, a safety plan — all of it downstream of an unreviewable machine output. Because the score is never introduced, Daubert and its state analogues are never triggered. There is no expert to cross-examine, no error rate to contest, no peer-reviewed validation to attack, no known false-positive rate to put in front of a factfinder. As the American Bar Association's own children's-rights practice materials state flatly: there is no child-welfare-specific law governing how attorneys and courts should interact with these predictions. The Sixth Amendment confrontation right does not attach in civil dependency proceedings, and Fourteenth Amendment procedural due process has not been tested against a hidden pre-investigation score.

The accountability gap

Every entity nominally positioned to check this has declined.

ACF approves the federal cost allocation through the APD process and requires nothing. Its March 2026 brief invokes governance, validation, transparency, and "human in the loop" — all as aspirations for grantees, none as conditions of the 50 percent match. DOJ's Civil Rights Division opened its ADA inquiry into Allegheny County after complaints filed in fall 2022 and AP reporting in January 2023, and has issued no public findings, no settlement, and no formal closure in the years since. The AFST remains in operation. State legislatures have largely not acted; Washington's SB 5356 algorithmic-accountability bill would prohibit discrimination via algorithm, but child-welfare-specific disclosure mandates remain essentially nonexistent. Vendors assert trade secrecy over weights and validation. Peer review is structurally compromised where the builders are also the evaluators and the policy advocates.

The single meaningful exception proves the rule. Colorado's HB 24-1046 required the state's Child Protection Ombudsman to commission a genuinely independent third-party audit of the state's Family Safety and Risk Assessment tools, including their differential impacts by race and ethnicity. The CPO contracted ICF, which conducted a mixed-method statewide review, and the report was delivered to the General Assembly on March 2, 2026. It found real problems. That is what oversight looks like — a legislature, an independent ombudsman, an outside auditor, a public report. One state has done it, for its actuarial tools, once.

Why it matters, and what would fix it

The base rates make the stakes concrete. In federal fiscal year 2023, CPS agencies received an estimated 4,399,000 referrals involving roughly 7,782,000 children, and substantiated 546,159 victims — a rate of 7.4 per 1,000. FY2024: 532,228 victims, 7.2 per 1,000. Roughly one in fourteen children referred is ultimately found to be a victim. Any screening tool operating at that base rate will generate false positives at overwhelming volume — as AURA demonstrated at 95.6 percent — and each false positive is a real investigation: a stranger in the home, children interviewed apart from parents, a case file that follows the family forever and raises the score on the next call. The feedback loop is the harm. A poor family investigated once is scored higher next time because they were investigated.

And it lands unequally by design. In New York City, 82 percent of ACS investigations in 2024 involved Black or Latino families, who are 48 percent of the population; Black families are reported at seven times the white rate and are thirteen times more likely to have a child removed. ACS's own internal audit conceded the training data likely carried "some implicit and systemic biases." A model trained on that record does not discover risk. It launders history into arithmetic and returns it with the authority of a number.

Five fixes are available immediately, and four of them require no new statute:

Condition the match. ACF can amend the APD approval standard tomorrow to require, as a condition of Title IV-E and CCWIS federal financial participation, a published external validation with calibration curves and false-positive rates disaggregated by race, disability status, and income.
Kill trade secrecy in the contract, not the courtroom. Federal cost allocation should be denied to any procurement in which the vendor retains proprietary claims over model weights, feature lists, or validation results used in a government decision about a family.
Mandatory notice and discovery. Any family investigated on a referral where a risk score was generated should receive the score, the input variables, and the threshold in writing — and it should be automatically discoverable in any dependency proceeding.
Publish the denominator. Not one state publishes how many children were removed from homes an algorithm ranked. That number should be reported annually to NCANDS.
Finish the DOJ investigation. Three years is long enough. Publish the findings or close it on the record.

Congress wrote a 50 percent check and never wrote a standard. Until the check is conditioned, "human in the loop" is a slogan describing a screener who sees the number first.