Predict-Align-Prevent

What this models

This example runs on the Predict-Align-Prevent-class place-based risk terrain mapping network: 8 components and 14 pathways between them. Every context in the Lab is a stylized model, never a reconstruction of any actual deployment, and each assumption behind it carries a provenance label.

Evidence base: 4 assumed · 6 published baseline. In the Lab, the shaded evidence band behind each headline readout draws its width from the least-established class below.

• assumed

  This models the place-based prevention-mapping pattern documented in the Predict-Align-Prevent case file, not a reconstruction of the actual models. The topology's defining feature is a documented absence rather than a component: no intake screener, investigator or caseworker node exists, because the record states that no frontline worker ever receives a score about a specific family and that frontline child-protection decision-making is untouched by the model. Every other child-welfare network in this catalogue has a per-case object between the model and a person; this one has a map. The current protocol's household-level counterfactual service-array stage softens that purity claim for the protocol era - it is framed as voluntary service matching for households in high-risk cells rather than family risk scoring - and it has no enrolled jurisdiction, so the 2016 to 2023 deployments this org draws hold the claim cleanly.

• baseline

  Two operator classes are drawn because the record documents two bodies with different roles and different documented actions: state agency program staff who receive the maps and decide what they mean, and family-resource-centre-led community coalitions in Manchester, the Winnipesaukee region and Coos County who converted map-identified need into targeted outreach and funded requests-for-proposals with stipends. The pathway between them is drawn at the strongest intensity because two federal-reporting sources fix the sequence and the consequence: mapping was completed for all three service areas before those sites received implementation resources, and the coalitions then issued funded solicitations against the areas the process identified, with no intermediate step documented that re-derives the need independently. That is the deployment's single best-evidenced causal channel.

• baseline

  The two model inputs run at the highest intensity for different documented reasons. The environmental feed carries 200 or more engineered exposure features - counts, Euclidean distances and average distance to the five nearest neighbours for each risk and protective factor - against a single event count per cell, so the predictor mass is overwhelmingly environmental. The record feed re-reads years of geocoded history on every refresh: 6,500 accepted cases across four years in the Richmond engagement, a year of state substantiations plus city police data in the origin study. The record feed is marked privacy-sensitive and the environmental feed is not, because the sensitive payload here is address-level children's case data rather than the place-level point data.

• baseline

  The model-side check is drawn live rather than dry, which is the opposite of the usual finding in this domain, and the evidence is specific. The builders ran leave-one-group-out cross-validation across 149 neighbourhoods, read three sub-models against each other in an ensemble, and compared the result against a kernel-density baseline; the current protocol retains spatial k-fold cross-validation against that same baseline. The check produced an adverse published result stated verbatim - the meta-model generalizes well across neighbourhoods of varying poverty rates but does not generalize well across neighbourhoods of varying race, despite race and income being excluded from the feature set - and no external body is documented responding to it. Richmond performance figures are builder-generated from a commissioned report; only the origin-study figures carry independent corroboration, through the 2021 systematic review.

• assumed

  The write-back pathway from community outreach into the geocoded record carries this Atlas's own structural reading and not a claim any cited source makes: concentrating institutional presence in the top-tier cells can raise detection and reporting in those cells, so new geocoded events accrue where the surface pointed and the next refresh reads them as signal. The framework's documented acknowledgment is different and narrower - stigmatization of individuals, households and neighbourhoods through high-risk labels, mitigated in its account because the intervention consists of optional supportive services rather than nonoptional punitive measures - and the framing about aligning resources before there is a suspicion of maltreatment is the organisation's own site language, not language from the cited sources. Because no cited source quantifies whether reporting rose in any targeted area, the pathway is carried at the lowest live intensity rather than asserted at strength.

• baseline

  Review here is structural rather than algorithmic, and the review step is wired to the model accordingly. The operational deployment sat inside a federal Children's Bureau cooperative agreement that carried continuous-quality-improvement requirements and produced cross-site documentation in March 2024; the mapping products entered that reporting. Its documented scope is how the maps were used and where implementation resources went, and every institutional response to the maps in the federal record is a resource-routing action. The agreement ran September 2018 to September 2023 and is the only documented operational planner use: the origin city remained a published study plus local advocacy with no evidence of operational allocation, and the Arkansas, Washington and Oregon engagements are vendor-claimed with no independent confirmation. The current program is a protocol published on 30 December 2025 with a participating jurisdiction under recruitment as of the paper's September 2025 status statement, and no newer enrollment evidence as of mid-2026 - recruitment-stage research rather than an operating system.

• baseline

  The record-to-outside pathway is drawn because the modelling was performed by parties outside the jurisdiction - a Texas nonprofit and a private consultancy - and the jurisdiction's geocoded child-welfare records had to reach them: 6,500 accepted cases in the Richmond engagement, and negotiated access to stewarded state databases in New Hampshire, with the engagement funded by a private foundation. It is drawn as a boundary crossing rather than as an in-network operator group precisely because the jurisdiction does not govern what happens on the far side, and this Lab names pathways across the boundary without measuring anything beyond them. It is carried at moderate intensity rather than high because the current protocol makes jurisdiction-level data-governance agreements and identifier masking explicit requirements of the transfer.

• assumed

  The machine-write pathway into the jurisdiction's own records is drawn empty because the record documents it as absent rather than because the shape looked bare: unlike score-and-screen systems there is no per-case disposition writing back to a family's record, and the delivered products are maps and reports. Two further pathways are omitted entirely rather than drawn dry. There is no cross-model coupling to a second system, because the framework moved between jurisdictions as sequential consultancy engagements with published open-source code rather than as a live multi-site coupling. There is no peer-review check inside either group of staff, because the record documents neither peer challenge nor case conferencing over the maps; what scrutiny the field received arrived from a national prevention nonprofit's non-endorsing research review and from civil-liberties reporting, both outside this deployment's boundary.

• baseline

  Demand and capacity are derived, not defaulted. Standing workload is set low because the operators drawn here work at a planning cadence the record fixes precisely: a ten-month planning period between the September 2018 award and the September 2019 implementation plan, mapping completed once per service area before resources flowed, and a five-year agreement - not a referral stream. The non-model counterfactual is set high because the record documents a working alternative running beside the maps: the three target communities were selected partly through the Centers for Disease Control and Prevention (CDC) Social Vulnerability Index, an established federal targeting instrument, and no independent outcome evaluation of any deployment exists to establish that the modeled targeting outperformed it. The leading national prevention nonprofit's research review of these techniques states explicitly that it is not intended as an endorsement of any of them.

• assumed

  Served families, children and neighbourhood residents are not in these dynamics; this Lab reads institutional propagation only. The case's real harm surface is place-level and demographic - neighbourhood stigmatization from a high-risk label, differential intensity of institutional attention in mapped places, and the builders' own documented failure to generalize across neighbourhoods of differing racial composition - and it is recorded in the case file and the external equity register, never computed here. Benefit and exposure land on the same cells, because the voluntary services and the concentrated attention arrive together. All impact claims - lead-poisoning training books, vaccine administration, nutrition-program enrollments, culture change - are vendor-reported, and no deployment has an independent outcome evaluation.

What this example does not show

• BINDING FRAMING: the surveillance-amplification loop this network draws - map-directed attention concentrating institutional presence in top-tier cells, which can raise detection and reporting there, enriching the records the next refresh trains on - is this Atlas's OWN structural analysis, and not an admission by the framework's builders. What the cited protocol actually acknowledges is narrower: stigmatization of individuals, households and neighbourhoods through high-risk labels, mitigated in its account because the intervention consists of providing optional supportive services rather than imposing nonoptional punitive measures. The often-quoted framing about aligning resources before there is a suspicion of maltreatment is the organisation's own site language and appears in none of the cited sources. No cited source quantifies whether reporting rose in any targeted area, so the loop is carried as structural and unmeasured.
• Served families, children and neighbourhood residents are not modeled here. This Lab reads institutional propagation only, and estimates no differential harm to served people. The documented harm surface of this case is place-level and demographic - neighbourhood stigmatization from a high-risk label, differential intensity of institutional attention in mapped places, and the builders' own finding that the model does not generalize across neighbourhoods of differing racial composition - and it is recorded in the case file and measured outside any diagram like this one.
• Evidence status differs sharply by figure. The origin-study numbers - the top tenth of cells capturing 52 percent of the following year's cases against 43 percent for a conventional hotspot model - are independently corroborated verbatim by a 2021 systematic review. Every other performance figure, including the roughly 70 percent against roughly 35 percent tier comparison and the error measure behind it, is builder-generated in commissioned reports. The impact claims - lead-poisoning training books, vaccine administration, nutrition-program enrollments, culture change - are vendor-reported, with no independent confirmation at all for two of the four claimed state engagements.
• Operational scope is narrower than the framework's reach suggests. The origin city was a published study plus local advocacy; no evidence shows child protection there allocating resources by the maps. The only documented operational planner use is the three-region federal cooperative agreement that ran September 2018 to September 2023. The current program is a protocol published on 30 December 2025 whose participating jurisdiction was under recruitment as of the paper's own September 2025 status statement, with no newer enrollment evidence as of mid-2026 - recruitment-stage research rather than an operating system.
• Sector reception is contested, and stays contested on this board. The leading national prevention nonprofit's research review of these techniques states explicitly that it is not intended as an endorsement of any of them, and 2021 reporting collected named critics of geographic hot-spot identification in child welfare. Nothing here resolves that dispute; the network shows the pathways, not a verdict on the approach.

Sources and evidence

What this example rests on, claim by claim. Every entry resolves to the same ledger the Evidence Registry publishes.

• In Richmond, Virginia, Predict-Align-Prevent's open-source place-based model (built with Urban Spatial; report byline Ken Steif, Matthew D. Harris, and Sydney Goldstein, 2019) reported its highest risk tier capturing about 70% of held-out child-maltreatment events against about 35% for a kernel-density baseline, while covering about 10% of city land area holding roughly 48,500 residents including about 8,200 children — builder-generated figures from a commissioned report — and the builders' own published fairness audit stated verbatim that the meta-model 'generalizes well across neighborhoods of varying poverty rates, but does not generalize well across neighborhoods of varying race,' despite race and income being excluded from the feature set.

  empirical
  • Vendor Steif, Harris and Goldstein, Predicting child maltreatment risk in Richmond, VA: An open source framework (Urban Spatial with Predict-Align-Prevent, 2019) https://urbanspatial.github.io/PredictingChildMaltreatmentInRichmondVA/
  • Vendor Urban Spatial, An open source geospatial risk predictive framework for child maltreatment (portfolio summary, 2019) https://urbanspatialanalysis.com/portfolio/an-open-source-geospatial-risk-predictive-framework-for-child-maltreatment/

• New Hampshire's 2018-2023 federal Community Collaborations cooperative agreement is the only documented operational planner use of Predict-Align-Prevent's maps: state narratives confirm the mapping (funded by Casey Family Programs, using address-level inputs accessed through stewarded databases) was completed for all three service areas before sites received implementation resources, and the March 2024 ACF/OPRE grantee profile documents that Community Implementation Teams used PAP-identified areas of need to target family outreach and issued funded requests for proposals with stipends — documented resource-routing behavior change, with no independent outcome evaluation of maltreatment effects in any deployment.

  empirical
  • Government Administration for Children and Families, Office of Planning, Research, and Evaluation, New Hampshire CWCC Grantee Profile: Community Collaborations to Strengthen and Preserve Families (2024) https://acf.gov/system/files/documents/opre/NH%20CWCC%20Grantee%20Profile_2024.03.28_5082.pdf
  • Government State of New Hampshire, Title V MCH Block Grant narrative: Other MCH Data Capacity Efforts (HRSA TVIS, 2020) https://mchb.tvisdata.hrsa.gov/Narratives/Other%20MCH%20Data%20Capacity%20Efforts/9510ebf8-ac54-4449-ab5c-eb868c982cbe

• The 2016 Fort Worth study (Daley et al., Child Abuse & Neglect), trained on 2013 state child-welfare substantiations and Fort Worth police data, reported the top 10% of grid cells capturing 52% of 2014 maltreatment cases against 43% for a conventional hotspot model — figures independently corroborated verbatim by the Marchment and Gill (2021) Crime Science systematic review, which also found it to be the single child-maltreatment application of risk terrain modeling in the reviewed literature; Fort Worth remained a published study plus local advocacy, with no evidence of operational resource allocation by the maps.

  empirical
  • Academic Daley, Bachmann, Bachmann, Pedigo, Bui and Coffman, Risk terrain modeling predicts child maltreatment (Child Abuse and Neglect, 2016) https://pubmed.ncbi.nlm.nih.gov/27780111/
  • Academic Marchment and Gill, Systematic review and meta-analysis of risk terrain modelling (RTM) as a spatial forecasting method (Crime Science, 2021) https://crimesciencejournal.biomedcentral.com/articles/10.1186/s40163-021-00149-6

Where this connects

Institutional pressures in this domain

• Workload surge — Demand outruns staffing; per-case attention shrinks and review becomes triage.
• Deadline pressure — Statutory or managerial timeliness rules reward fast approval of machine output over slow disagreement.
• Staff turnover — Experienced skepticism leaves; new staff calibrate their trust on the tool itself.
• Data & policy drift — The world, the intake process, and the rules change under a system trained on how things used to be — two mechanisms with different remedies: the statistical properties of what the system processes move (concept drift), or the mixture of inputs arriving in deployment differs from the mixture it was trained on (covariate shift).
• Compliance over substance — Paper controls (sign-offs, checklists) satisfy audits while the behavior they describe erodes.

All of them in context on the Child welfare & family services domain page.

Levers available here and the patterns behind them

• Gate vendor updates — Vendor quality gate
• Understand the system — Understand the system
• Vet connections — Connection authorization
• Store less data — Data minimization
• Mark AI-written records — Provenance labeling
• Assign a challenger — Structured dissent
• Check with a second model — Cross-model verification
• Review on schedule — Oversight cadence & retrospectives
• Keep skills sharp — Deskilling-arrest mandate
• Peer sharing rules — Peer-edge governance
• Pause AI on alarms — Deployment circuit-breaker
• Gate record entries — Human-in-the-loop write gating
• Upgrade model — Improve the model

Documented case histories

• The map, not the score: place-based risk terrain and the records it concentrates
• Allegheny Family Screening Tool
• Allegheny Hello Baby
• Douglas County Decision Aide
• The score nobody sees: New York City's concealed severe-harm QA algorithm
• The audit that reached the legislature before it reached the tools: Colorado's safety and risk instruments
• Eckerd Rapid Safety Feedback: origin and spread
• Illinois Rapid Safety Feedback
• The vendor's ledger: Family-Match, the eharmony-derived adoption matcher the states kept coming back to
• ProKid (Netherlands)
• Insight Bristol / Think Family Database
• Hackney / Xantura Early Help Profiling
• Sistema Alerta Niñez (Chile)
• The guardrail's blind side: DC's walled-off child-welfare chatbot that began writing into the case record
• US Birth Match
• Oregon Safety at Screening
• Los Angeles County Project AURA
• What Works for Children's Social Care ML pilots
• New Zealand MSD Predictive Risk Modelling
• Gladsaxe model