Housing & homelessness services

A warehouse-data risk score replaced a self-report homelessness survey and got far more accurate — yet a peer-reviewed evaluation found it did not close the racial gap in who actually gets scarce housing.

A self-report homelessness triage questionnaire became the U.S. coordinated-entry standard across dozens of states without validation — then its own creators retired it in 2020 on equity grounds, after commissioned research found it systematically under-scored people of color.

A predictive model that reaches single adults before they lose housing — offering cash and help, never a denial — where the error that matters is the majority it misses, and even the headline 71% result is associational until a 2027 trial.

A vendor 'single view' that flags residents likely to become homeless months ahead — where the model's real reach is set by how many alerts one officer can act on, and the marquee figures are vendor pre/post numbers a randomised trial was still testing.

A city-built, caseworker-facing model that flags shelter clients at risk of chronic homelessness about six months out — consent-based, open-source and explainable by design, yet resting on a testing-phase accuracy figure no one re-validated after deployment and a shelter-only data frame that cannot see the people who never come through a public shelter door.

A generative benefits-navigation copilot whose evaluated accuracy gains land hardest on the newest staff and the toughest questions — where the built-in citation check is easiest to skip — in a vendor-and-academic-co-authored 2025 Los Angeles County pilot.

A pan-London service that probabilistically links three separately governed record systems — street-outreach contacts, charity casework, and borough statutory applications — into a single rough-sleeping journey per person, read across all 33 London local authorities. It makes no individual-level decisions: its documented weakness is a self-reported 91% recall, meaning roughly 9 in 100 true cross-system matches are missed, an undercount its own DPIA concedes propagates into the figures every borough sees at once. The atlas's memory-consolidation case: the lesson is what happens when many independent readers consolidate onto one shared layer whose known error is never reconciled.

San Jose mounted cameras on a sedan - its "camera car" - and drove one district, training vendor AI to detect potholes and trash — and, in the same pilot, RVs, lived-in vehicles, and homeless encampments. Described by city officials and housing advocates as the first US experiment training AI to recognize tents and lived-in vehicles, it was accurate on infrastructure (97% on potholes) and badly inaccurate on habitation (12.5% on lived-in vehicles). No detection ever drove an operation: the response loop was never wired. The atlas's declared-rule-versus-feasible-flow case — the data policy forbade active law-enforcement use while preserving a police request path to the footage, and a vendor read license plates the policy said it would not, before the city removed every habitation class and declined to recommend deployment.

The largest homeless-services system in the U.S. replaced a discredited triage survey with a fairer, regression-derived score — then watched the retired bias walk back in during the transition. Because the old and new instruments ran in parallel at mismatched eligibility thresholds, a client was likelier to qualify for housing under the OLD biased tool, so frontline workers kept administering it. The atlas's judge-replacement-transition case: the lesson is that swapping a biased score is a governance problem in the handover, not just a modeling problem in the tool.

A homelessness-prevention program whose targeting runs on a transparent points-based intake questionnaire, not a machine-learning model, and whose intervention - flexible emergency financial assistance - is one of the few in the Atlas backed by a published randomized controlled trial. The domain's positive counterexample: the lever is measured and effective, and the open question is not whether to trust the model but whether the screen reaches the right people on a base rate where most at-risk applicants would have stayed housed anyway - a question the same program is now testing across ten new jurisdictions at once.

A Calgary shelter and University of Calgary engineers built deliberately simple, interpretable screening for chronic shelter use — explicit rules like "81+ stays in 90 days" that flag risk months earlier than the official definitions — and then, instead of a score, gave frontline staff an interface that shows the raw client history and lets them read it. They studied their own staff for two and a half years and found the deference calibrated (whether a score of, say, 30% really does mean thirty cases in a hundred) to the stakes: staff read everything for high-stakes barring decisions, treating the data as "the canvas," but leaned on it more for lower-stakes triage. The atlas's directly observed deference case: the lesson is that vigilance can be designed into a tool's form — and that the residual risk then migrates to the low-stakes tail of that deference and to the records the staff themselves author.

A tenant-screening vendor turned an applicant's credit and eviction file into a 200-800 number and an accept or decline recommendation, while the leasing desk that signed the lease wrote that it could not override the outcome — and a federal settlement responded not by fixing the model but by switching its output off for housing-voucher applicants for five years, with a civil-rights organization holding the only key back in.

Competing landlords fed RealPage's revenue-management pricing engine their nonpublic lease data — executed rents, renewal rates, terms, occupancy — and the engine recommended daily rents back to all of them. The 2024 federal antitrust complaint alleges roughly 80% of the commercial revenue-management software market and data agreements reaching over 16 million units, including units of landlords who were not customers. The loop is the case: accepted recommendations became executed-lease records in the shared dataset, which shaped the next day's recommendations for every participating rival, while a price-movement band was alleged to damp decreases harder than increases and an automatic setting executed with no human review at all. The remedy is topological rather than accuracy-based — nobody audited whether the prices were right. The proposed decree stales the shared memory to a twelve-month floor, bars sub-state geographic modelling, re-symmetrises the band, turns automatic acceptance off by default, bans vendor-hosted meetings of competitors, and appoints a monitor for three years. Nothing is adjudicated: no admissions, no fine, and final entry was still pending in mid-2026.

New York City routes families toward or away from full homelessness-prevention services on fifteen questions and a point total — and it is the rare targeting instrument bolted into a statutory public self-reporting loop, so the city has had to describe the tool in a public register four years running, including a mid-cycle amendment disclosing its 2023 revision, while the agency's own research office peer-review-published both the screener's accuracy and its distortions: caseworkers' totals piling up at exactly the eligibility cutoff, and supervisor-approved overrides that misjudged risk in both directions.

A rental application was denied on a flag that said only that a disqualifying record existed — the record was a pending shoplifting charge later withdrawn, the on-site agent had been configured out of seeing it, and the company holding it never completed the file disclosure the law provides. Eight years of litigation ended with a court-certified map of who decides what in automated tenant screening, and with the vendor owing nothing.