Domain Atlas / Housing & homelessness services
VI-SPDAT
Work with this case in the PAN Lab ↗
The VI-SPDAT was the dominant U.S. homelessness triage assessment for roughly a decade, adopted in at least 39 states and the District of Columbia by 2015, before its own creators announced its phase-out in December 2020 on equity grounds; a 2019 commissioned racial-equity evaluation across four Continuums of Care found race predicted 11 of 16 subscales and that people of color received statistically significantly lower prioritization scores.[3]
What happened
The Vulnerability Index — Service Prioritization Decision Assistance Tool (VI-SPDAT) was released in 2013 as a joint instrument: OrgCode Consulting's SPDAT prescreen merged with Community Solutions' Vulnerability Index (the tool that grew out of the 100,000 Homes Campaign). Updated versions followed in 2015 and 2020, with separate forms for single adults, families, and youth. A frontline homeless-services worker administers the survey by interview, asking yes/no items across four domains — history of housing and homelessness; risks; socialization and daily functioning; wellness — and sums them into a score (0 to 17 for single adults). Fixed thresholds route the score to a housing tier: roughly 8 and above steers toward permanent supportive housing, 4 to 7 toward short-term rapid re-housing, and 0 to 3 toward minimal services or self-resolution. Higher vulnerability means more "Yes" answers means a higher score. It is worth stating plainly what the tool is and is not: it is a fixed-weight actuarial questionnaire scored by hand or in a Homeless Management Information System (HMIS), not a machine-learning system. It belongs in an atlas of algorithmic allocation tools, but calling it "artificial intelligence" would be inaccurate.
By 2015 the VI-SPDAT was implemented in at least 39 states, the District of Columbia, and internationally, becoming the industry-leading coordinated-entry assessment as HUD pushed Continuums of Care toward standardized prioritization. Because it was open-source and free, many communities deployed it without OrgCode training. OrgCode had designed it as a "Decision Assistance Tool, not a Decision Making Tool," but the central concern in its December 2020 phase-out statement was that communities were relying solely on the score to allocate resources rather than using it as one input, and that the tool "was never designed using a racial or gender equity lens." Several strands of evidence had accumulated by then. A 2019 racial-equity evaluation by C4 Innovations (commissioned by Building Changes), covering four Continuums of Care — King and Pierce Counties in Washington, Multnomah County in Oregon, and the Blue Ridge area in Virginia — found that race predicted 11 of the 16 subscales and that people of color received statistically significantly lower prioritization scores, with subscales tilted toward vulnerabilities White clients were more likely to endorse; being White was a protective factor for single adults. (A secondary restatement of the same data reported White clients were more than 60% more likely to receive a high prioritization score than BIPOC clients.) Cronley's 2020 intersectional analysis of a large community sample found that White women scored consistently higher on vulnerability than Black women and all men, and that being White directly and significantly predicted higher scores — evidence of a bias that could mask trauma among Black women and delay their housing. A 2024 evaluation by the Central Valley Health Policy Institute at California State University, Fresno, analyzed 1,369 single-adult assessments in the Fresno-Madera Continuum of Care (71% BIPOC, 29% White) and found BIPOC clients more often answered "No" to Risk and Wellness items, lowering their scores and overrepresenting them in the less-intensive rapid re-housing tier; importantly, the study's aggregate chi-square tests of the association between race and recommended housing assignment in that single sample were not statistically significant, so the clear finding is at the item-endorsement level rather than a proven end-to-end assignment gap. Separately, the instrument showed poor test-retest reliability (most participants scored higher on re-administration) and poor inter-rater reliability (scores varied by interviewer and site). Its predictive validity for housing outcomes was genuinely mixed rather than uniformly poor: positive for the youth version (Rice et al., 2018), null for single adults in one study (Brown et al., 2018), and positive in another multi-community sample (Petry et al.).
OrgCode announced the phase-out in December 2020, ended all support for the tool at the close of 2022, and stated it would not develop a replacement, relinquishing that role to "experts in racial and gender equity." The residual tool and local variants remain in use in some Continuums of Care into 2025 while communities build successors: the Fresno-Madera CoC commissioned CESMAT (2024), and in Rochester, New York, Partners Ending Homelessness switched to the Homelessness Assessment Tool (HAT) on June 2, 2025 — a switchover that required everyone already on the local Prioritization List to be re-assessed with the new tool. There is no authoritative national count of how many VI-SPDAT assessments were ever administered: because the tool was open-source and decentralized, any total is an estimate from adoption breadth against a system that shelters an estimated 1.4 million people a year, not a measured figure. Advocates and OrgCode alike note that the tool "revolutionized" homeless-services delivery by replacing pure caseworker discretion with a standardized process — which is why its withdrawal reopened a governance gap between biased-tool prioritization and unchecked human discretion, both of which carry equity risk.
The sociotechnical reading
This case shares an outline with the Eckerd RSF story — an instrument that scaled without an independent validation ever firing, copied jurisdiction to jurisdiction so its blind spots travelled with it. But VI-SPDAT teaches something the Eckerd case does not, on two distinct axes. First, the failure is a design-versus-use collapse, not a false efficacy claim. The tool was explicitly labeled "Decision Assistance, not Decision Making"; what broke was the boundary itself. Nothing in the coordinated-entry workflow required a human to disagree with the score, so an aid meant to be one input became, in the builder's own words, the sole decision rule across a whole sector. On the system map that is a specific, legible shape: the model-to-model check edge sits empty (no validation, reliability, or equity test ran as the questionnaire became the standard), and the operator-to-operator override that should have caught a bad rank sits latent (present by design, collapsed in practice). Second, the bias here is unusually legible and lives in the input, not in a hidden weight. This is a deficit-based questionnaire on which people of color — and Black women in particular — more often answer "No," scoring as less vulnerable and routing to thinner help. Because the tilt is in how vulnerability is elicited and which items count, "improve the model" is the lever that moves it least; the productive levers act on the elicitation, the labeling of the score as a rank rather than a placement, and the missing checks around it.
The governance reading follows from those two features. The map's clearest instruction is a conformity or validation gate — the check that never fired before an instrument became the thing every community trusts — backed by an audit-and-monitoring cadence, because the only scrutiny this tool received came after deployment, from academics and advocates rather than any governing body, and because a single fixed questionnaire adopted across dozens of states is a monoculture whose errors are correlated by construction. Two loops make the harm persistent rather than one-shot: the biased score is written to the shared HMIS, orders the community prioritization list, and re-infects every referral read off that list; and poor test-retest reliability means re-administration itself shifts priority. That is why clearing stale scores off the list is a real lever — but a content-aware one: Rochester's 2025 switchover re-assessed its whole list with a new instrument rather than merely swapping tools, whereas deleting records unread would only strip the benign entries that were diluting the biased ones. The uncomfortable coda, and the part that distinguishes this case from every other in the Atlas, is that withdrawal did not resolve the dilemma. Retiring a biased standard hands the decision back toward the unchecked caseworker discretion the standardized process had partly displaced — so the honest governance question is not "keep the tool or drop it?" but "what fills the space a standard occupied?" The lesson is that the presence of a standard is not the presence of a safeguard, and removing one is not the same as building the other. The honest boundary throughout: none of this measures harm to the unhoused people being ranked. The service-rate disparity and the scoring bias are documented outside any diagram like this one, and a one-shot housing allocation is a rationing decision, not an error that spreads.
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.