Predict-Align-Prevent

This network follows the place-based prevention mapping in the Predict-Align-Prevent case file. It does not rebuild the actual models. It has no intake screener, investigator, or caseworker, because the sources say no frontline worker ever receives a score about a family. Frontline child-protection decisions are untouched by the model. Every other child-welfare network in the Lab puts something about a single case between the model and a person. This one puts a map there. The 2025 protocol adds a household-level stage that weakens this claim for future work. Its authors frame it as voluntary service matching for households in high-risk cells, not family risk scoring. The sources show no jurisdiction enrolled to run it, so the claim holds for the 2016 to 2023 deployments drawn here.

The network draws two groups of staff, because the sources describe two bodies with different roles and actions. State agency program staff receive the maps and decide what they mean. Coalitions led by family resource centers in Manchester, the Winnipesaukee region, and Coos County turned mapped need into targeted outreach and funded requests for proposals with stipends. Two federal reporting sources fix the order of events. Mapping was completed for all three service areas before those sites received implementation resources. The coalitions then issued funded requests for the areas the process identified. The sources describe no step in between that works out the need independently. This hand-off is the deployment's best-documented way the maps changed what people did.

The network draws two inputs into the model, and the sources give a different reason for each. The environmental data carries more than 200 measures per cell, against one count of past cases. The measures are counts, straight-line distances, and average distances to the five nearest examples of each risk and protective factor. The case records are read again, years of them, every time the model is rebuilt. Richmond used 6,500 accepted cases over four years. The 2016 Fort Worth study used a year of state substantiations plus city police data. Only the case records are marked as sensitive, because they are address-level records about children, while the environmental data describes places.

The network draws the builders' own test as a working check, unlike most checks in child-welfare networks, because the sources show it ran. The builders held out one neighborhood at a time across 149 neighborhoods, and read three sub-models against each other. They compared the result with a simpler map of past cases. The 2025 protocol keeps a similar test against the same kind of map. The test produced an adverse published result, although race and income were left out of the model's inputs. The model "generalizes well across neighborhoods of varying poverty rates, but does not generalize well across neighborhoods of varying race." The sources document no outside body responding to it. The Richmond performance figures are the builders' own, from a commissioned report. Only the Fort Worth figures were confirmed independently, by a 2021 systematic review.

The pathway from community outreach into the geocoded records is the case file's own reading, not a claim any cited source makes. Outreach concentrated in the top-tier cells can raise detection and reporting there. New records would then gather where the map pointed, and the next model would read them as signal. Predict-Align-Prevent acknowledges something narrower: high-risk labels can stigmatize individuals, households, and neighborhoods. In its account that risk is reduced because the intervention offers optional supportive services rather than punitive measures. The often-quoted phrase "before there is a suspicion of maltreatment" comes from the organization's own website, not from the cited sources. No cited source measures whether reporting rose in any targeted area, so the network shows this pathway as a possibility, not a measured effect.

Review in this deployment covered how the maps were used, not how the model worked, and the network draws it that way. The mapping took place inside a federal Children's Bureau cooperative agreement, a form of federal grant, with continuous quality-improvement requirements. The mapping data was used in the grant's reporting, and a cross-site account of its grantees was published in March 2024. The review covered how the maps were used and where implementation resources went. Every institutional response to the maps in the federal documents is a decision about where resources go. The agreement ran from September 2018 to September 2023 and is the only documented operational use by planners. Fort Worth remained a published study plus local advocacy, with no evidence of resources allocated by the maps. The Arkansas, Washington, and Oregon engagements are claims on Predict-Align-Prevent's own website, with no independent confirmation for Washington and Oregon. The current program is a protocol published on 30 December 2025. As of the paper's September 2025 status statement, a participating jurisdiction was under recruitment. The sources show no newer enrollment as of mid-2026, so it is research at the recruitment stage, not an operating system.

The network draws a pathway from the records to the outside modelers because parties outside the jurisdiction did the modeling: a Texas nonprofit and a private consultancy. They needed the jurisdiction's geocoded child-welfare records. In Richmond that meant 6,500 accepted cases. In New Hampshire, state staff accessed stewarded state databases for the mapping, which Casey Family Programs, a private foundation, funded. The modelers are drawn beyond the network's boundary, not as a group of staff inside it, because the jurisdiction does not govern what happens there. The Lab names pathways across that boundary and measures nothing beyond them. The 2025 protocol makes jurisdiction-level data-governance agreements and masking of personal identifiers explicit requirements of the transfer.

The network includes a pathway for the model writing into the jurisdiction's own records, although the sources say the deployments made no such write. Unlike systems that score cases, no per-case decision was written back to a family's record, and the products were maps and reports. Two other pathways are left out entirely. There is no link to a second model system, because the method moved between jurisdictions as consecutive consulting engagements with published open-source code. There is no peer-review check inside either group of staff, because the sources record no peer challenge or case conferences over the maps. The scrutiny the field received came from outside this deployment: a national prevention nonprofit's research review that declines to endorse the methods, and civil-liberties reporting.

The network assumes a light standing workload, because the planners work to a planning schedule the sources fix, not a stream of referrals. A ten-month planning period followed the September 2018 award, and the implementation plan was approved in September 2019. Mapping was completed once per service area before resources were released, within a five-year agreement. The network also assumes the work would go reasonably well without the model, because a working alternative ran beside the maps. The three communities were chosen partly using the CDC Social Vulnerability Index, an established federal targeting index. No independent outcome evaluation of any deployment shows the model's targeting did better. Prevent Child Abuse America's research review of these methods states that it is not intended as an endorsement of any of them.

The families, children, and neighborhood residents served are not part of this network. The Lab traces how mistakes pass between the model, staff, and records, not between groups of people. This case's documented harms are about places and groups. They include stigma for neighborhoods labeled high risk, uneven institutional attention in mapped places, and the model's documented failure to generalize across neighborhoods of differing racial composition. They are recorded in the case file and outside this network, and never computed here. The same cells receive both the benefit and the exposure, because the voluntary services and the concentrated attention arrive together. Every impact claim is reported by Predict-Align-Prevent itself: lead-poisoning training books, vaccine administration, and enrollments in a nutrition program. No deployment has an independent outcome evaluation.