Domain Atlas / Child welfare & family services
Los Angeles County Project AURA
In a retrospective test against historical outcomes, Los Angeles County's Project AURA — a proprietary risk model built by SAS — correctly flagged 171 of the highest-risk children but produced 3,829 false positives, a false-positive rate of about 95.6% that DCFS's own public-affairs director confirmed on the record, and the county shelved the tool in 2017 without ever using it on a live case.[4]
What happened
Los Angeles County's Department of Children and Family Services contracted the analytics firm SAS around 2013 to build Project AURA (Approach to Understanding Risk Assessment), a predictive risk model that scored children in abuse and neglect referrals on a 1-to-1,000 scale from a mix of administrative data — prior referrals, law-enforcement involvement, mental-health records, and alcohol and substance-abuse history. The model was trained and tested retrospectively against 2011 and 2012 child deaths, near-fatalities, and "critical incidents." In that retrospective test, at a high-risk cut, AURA correctly identified 171 of the worst-outcome children but also produced 3,829 false positives — a false-positive rate DCFS's own public-affairs director confirmed on the record as "no less than 95.6%." The algorithm was proprietary to SAS, and users faulted it for focusing narrowly on the caregiver while missing the fuller family story. AURA was never used on a single live case; it was only run against past reports. In May 2017 the LA County Office of Child Protection reported to the Board of Supervisors that DCFS "is no longer pursuing Project AURA" — a status disclosed nearly two years after the false-positive results were already known. The tool's development had unfolded amid intense scrutiny of DCFS following the May 2013 death of eight-year-old Gabriel Fernandez in Palmdale. LA County later launched a separate, in-house tool — the Risk Stratification Pilot — in 2021, which, unlike AURA, is non-proprietary, draws only on the county's own records, and deliberately excludes race, ethnicity, and geography.
The sociotechnical reading
Among the Atlas's never-deployed systems, AURA is the one arithmetic stopped. In system-map terms the governing control fired before the model was ever wired to an operator: a retrospective test against historical outcomes exposed a base-rate problem no accuracy figure could hide — 171 true flags against 3,829 false ones — and the county declined to deploy. That is the cheapest and most reversible control on the whole map: a permission gate willing to run the evaluation before go-live and to answer "no." Two things complicate the clean story. The model was proprietary, so the county was evaluating a box it could not open, and the evidence that stopped AURA was an outcome statistic rather than an inspectable mechanism — which is why the vendor-inspectability question sits next to the permission gate here. And the gate held slowly: the false-positive rate sat known for nearly two years, surviving renewed interest, before the tool was formally shelved. The lesson for anyone rehearsing here is that the base-rate arithmetic that makes a rare-outcome screener saturate is knowable in advance, and refusing deployment is a governance event in its own right — not a failure to ship. It is the counterpoint to every case in this Atlas whose controls arrived only after the harm.
The concepts used in this reading are defined in the Field Guide; the governance responses live in the Practice Library. The model organization for this case can be stress-tested in the PAN Lab.