Skip to content
ParamergeParamerge
Stephen Lieberman
Through Paramerge
Contact

PAN Lab example

Santa Clara County Homelessness Prevention System

A measured lever on an unmeasured target: a homelessness-prevention screen

A points questionnaire scores a household's risk of losing its home, and a caseworker offers the highest-scoring ones a few thousand dollars - back rent, a deposit, a car repair - never a denial. Modeled on Santa Clara County's Homelessness Prevention System. A randomized trial says the help works: assisted households were far less likely to become homeless, at about 2.47 dollars of community benefit per dollar spent. So what is left to govern? The thing the effect size cannot tell you: whether you are reaching the right people. Most at-risk applicants would have stayed housed anyway - so the money spent on them looks exactly like success. Watch the target, not just the lever, as the same program spreads to ten new places at once.

Stylized model of a documented deploymentHousing & homelessness services

Open this example in PAN Lab v0.1 to apply pressures and levers and watch what the system does.

What this models

This example runs on the Santa-Clara-Prevention-class screened flexible-assistance lever network: 4 components and 11 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: 2 assumed · 4 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 screening-plus-intervention pattern documented in the Santa Clara County Homelessness Prevention System case file - not a reconstruction of the actual program. The atlas-relevant object is the topology: a transparent points-based intake questionnaire (the trial randomized within a middle band of scores 8 to 13, not a machine-learning model) that prioritises a real, randomized-trial-calibrated intervention lever - flexible emergency financial assistance delivered with high caseworker discretion. It is deliberately distinct from the library's Los Angeles County Homelessness Prevention Unit, which is a supervised machine-learning ranker whose causal trial is still pending and whose dominant failure mode is the missed high-risk person, and from the separate 2016 Silicon Valley Triage Tool for prioritising already-homeless individuals. Any reading that treats this screen as an opaque model, or the lever as an eligibility gate, misreads it.

  • baseline

    The lever is measured and the effect is good. A registered randomized controlled trial (n=1,263; 514 treatment, 749 control; July 2019 to December 2020) run by the University of Notre Dame's evaluation lab, with outcomes tracked through the county Homeless Management Information System (HMIS), reported assisted households 81 percent less likely to become homeless within six months and 73 percent within twelve (Notre Dame / LEO releases); the peer-reviewed article's abstract states a 3.8 percentage-point reduction from a 4.1 percent base rate; and CalMatters reports the underlying rates as 0.9 percent treated versus 4.1 percent control, which it pairs with a 78 percent relative reduction. These are separately sourced figures and are not fused here. The researchers' conservative benefit-cost estimate is 2.47 dollars per net dollar. Program-scale figures (31,000-plus households, roughly 7,000 dollars average, 90-plus percent housed at two years) are Destination: Home's own reporting.

  • baseline

    The defining structural property is that a measured lever is not a measured target. Becoming homeless is statistically rare even among at-risk applicants - about 96 percent of the trial control group never became homeless - so screening precision, not lever strength, is the binding constraint, and the study co-author cautions on the record that precious resources could go to people who would have stayed housed anyway. The deadweight is invisible because those people do stay housed, so mis-targeting looks exactly like success. The two load-bearing absences are drawn as inactive checks: a record-against-record reconciliation of the screen's targeting (the scores it logs to the records) against the measured Homeless Management Information System (HMIS) outcomes (does it reach the true would-be-homeless, or deadweight?), and an independent peer evaluation for a standing per-site evaluation of whether the lever's strength survives transfer to each heterogeneous replication jurisdiction.

  • baseline

    High human discretion is a design feature, not an override pathology. Caseworkers determine assistance type and amount case-by-case, non-financial services accompany the money, and households can return for repeat help; the lottery replaced first-come-first-served only during the evaluation window because demand exceeded funds. The correction edge is therefore drawn strong, and the deference channel from the score is drawn moderate rather than strong - the score routes an offer of help, never a denial. Protecting that discretion (deskilling-arrest) is what keeps a benefit from hardening into a list you are either on or invisible to; the closed Homeless Management Information System (HMIS) outcome loop is the positive measurement property that made the trial possible.

  • baseline

    The system is entering a live topology transfer. As of February 2026 the same lever is being re-instantiated across about ten structurally heterogeneous US jurisdictions under the 77 million dollar Right at Home initiative (minimum 5 million dollars per site over three years, implementation by January 2027, a goal of keeping 10,000-plus households housed), with the same evaluation lab as the common national evidence partner assessing each site 2026 to 2031. The reinforcing operator-to-operator edge carries the lever design and the Results for America replication toolkit propagating to the new sites; the risk it models is the playbook outrunning the per-site evidence, and connection-auth (authorise each site's adoption with its own evaluation attached) and peer-governance (govern the multi-site network) are its governors. The 77 million dollar figure is the announced amount; CalMatters reports nearly 80 million dollars raised.

  • assumed

    Served people are not in the dynamics; whether any household does or does not become homeless is documented in the case file and measured outside a diagram like this one, never computed here. The privacy surface is the self-reported special-category intake (domestic-violence history, disability) entered to obtain help and the sensitive county Homeless Management Information System (HMIS) outcome records; the evaluation additionally drew on consumer-reference address-change data. The trial enrolled a middle-band (8 to 13) risk slice during a window overlapping COVID-era eviction moratoria and rental assistance, which may depress base rates, so effects at other risk levels and outside that window are unmeasured, and the program-scale and 'first randomized controlled trial (RCT)' framings are Destination: Home's own claims, not independently audited. A score, an assistance decision, or an outcome on this map is an institutional signal, never a person.

What this example does not show

  • Served people - households at risk of homelessness, and whether they do or do not lose their housing - are not modeled here; the Lab reads institutional propagation only, and those outcomes are documented in the case file and measured outside any diagram like this one. The tool denies no one and adjudicates nothing; the harm surface it reads is misallocation and epistemic (deadweight targeting on a low base rate, deference to a screen, and a result assumed to transfer), never an individual determination.
  • The effect figures are separately sourced and must not be fused: 0.9 percent treated versus 4.1 percent control (a 78 percent relative reduction, CalMatters), 81 percent lower within six months and 73 percent within twelve (Notre Dame / LEO releases), and a 3.8 percentage-point reduction from a 4.1 percent base (the published article's abstract) come from different documents; the trial randomized a middle-band (scores 8 to 13) risk slice during a July 2019 to December 2020 window overlapping COVID-era eviction moratoria and rental assistance, so effects at other risk levels and outside that window are unmeasured; the 2.47 dollar benefit-cost ratio is the researchers' self-described conservative estimate; and program-scale figures (31,000-plus households, roughly 7,000 dollars average, 90-plus percent housed at two years, the 'first randomized controlled trial (RCT)' framing) are Destination: Home's own reporting, not independently audited.
  • A safe-looking baseline is a property of this model, not a safety promise for any real deployment. The binding constraint the program's own evaluator names - targeting precision on a low base rate - is not resolved by the effect size, and it is not yet re-measured at the ten replication sites, whose per-site evaluations run 2026 to 2031; the 77 million dollars is the announced Right at Home figure (CalMatters reports nearly 80 million dollars raised), and the goal is 10,000-plus households.

Sources and evidence

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

  • In a registered randomized controlled trial of 1,263 imminent-risk applicants (514 treatment, 749 control) run by the University of Notre Dame's evaluation lab, households offered flexible emergency financial assistance averaging about 2,000 dollars, typically one to two months of back rent, through Santa Clara County's homelessness-prevention system were reported 81 percent less likely to become homeless within six months and 73 percent within twelve months; the peer-reviewed article's abstract states the assistance reduced homelessness by 3.8 percentage points from a 4.1 percent base rate, and the researchers conservatively estimated 2.47 dollars in community benefits per net dollar spent.

    empirical
    • Academic Phillips and Sullivan, Do Homelessness Prevention Programs Prevent Homelessness? Evidence from a Randomized Controlled Trial (The Review of Economics and Statistics 107(5): 1187 to 1196, 2025) https://doi.org/10.1162/rest_a_01344
    • Reference University of Notre Dame News, Targeted Prevention Helps Stop Homelessness Before It Starts (2023) https://news.nd.edu/news/targeted-prevention-helps-stop-homelessness-before-it-starts/
    • Academic Phillips and Sullivan, Do Homelessness Prevention Programs Prevent Homelessness? Evidence from a Randomized Controlled Trial (AEA RCT Registry, AEARCTR-0008261, 2021) https://www.socialscienceregistry.org/trials/8261

  • Because becoming homeless is statistically rare even among at-risk applicants - about 96 percent of the trial's control group never became homeless without assistance - the program's own co-author cautions that prevention resources can flow to households that would have stayed housed anyway, making screening precision on a low base rate the binding constraint; as of February 2026 the model is being replicated across about ten heterogeneous US jurisdictions under a 77-million-dollar initiative, with the same evaluation lab as the common evidence partner assessing each site.

    empirical
    • Investigative Kendall, A New Homelessness Strategy Is Sweeping California (CalMatters, 2026) https://calmatters.org/housing/homelessness/2026/03/homelessness-prevention-pilot/
    • Academic Phillips and Sullivan, Do Homelessness Prevention Programs Prevent Homelessness? Evidence from a Randomized Controlled Trial (The Review of Economics and Statistics 107(5): 1187 to 1196, 2025) https://doi.org/10.1162/rest_a_01344
    • Advocacy Destination: Home, Destination: Home Launches Right at Home, a National Initiative to Stop Homelessness Before It Starts (2026) https://destinationhomesv.org/news/2026/02/24/destination-home-launches-right-at-home-a-national-initiative-to-stop-homelessness-before-it-starts/
    • Reference University of Notre Dame News, Notre Dame's LEO Joins National Initiative to Stop Homelessness Before It Starts, Serving as the Lead Evidence Partner (2026) https://news.nd.edu/news/notre-dames-leo-joins-national-initiative-to-stop-homelessness-before-it-starts-serving-as-the-lead-evidence-partner/

Where this connects

Institutional pressures in this domain

  • Workload surge — Demand outruns staffing; per-case attention shrinks and review becomes triage.
  • Austerity & recovery incentives — Cost-cutting and overpayment-recovery targets tilt the system toward denial and enforcement errors.
  • Vendor opacity — The deploying institution cannot inspect the model, data, or update pipeline it is accountable for.
  • 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 Housing & homelessness services domain page.

Levers available here and the patterns behind them

Documented case histories