Domain Atlas / Housing & homelessness services
LA's coordinated-entry triage revision: the fix that needed fixing
Work with this case in the PAN Lab ↗
The Los Angeles Coordinated Entry System replaced the VI-SPDAT survey for single adults with the Los Angeles Housing Assessment Tool, a 19-item self-report score whose weights were derived by a regression on 71,747 historical assessments linked to county records; where the CESTTRR research estimated the VI-SPDAT scored near chance (AUC 0.54) with racial false-negative gaps up to 8.5 percentage points, the equity-adjusted successor was deliberately traded down in overall accuracy (AUC 0.60, from an accuracy-only 0.64) to close those gaps to under one percentage point, and every such figure is a pre-deployment estimate on 2015 to 2018 held-out data rather than an observed post-launch outcome.[2]
What happened
The Los Angeles Housing Assessment Tool (LA HAT) is the triage instrument the Los Angeles Coordinated Entry System (CES) adopted to replace the VI-SPDAT survey for single adults — the largest such system in the country. It is a 19-item self-report questionnaire whose per-question point weights were derived by a regression model: the December 2024 press account describes ordinary least squares, while the project's technical appendix describes a minimally penalized Lasso linear model with positive coefficients. The weights were derived from 71,747 historical VI-SPDAT assessments (54,543 unique people, July 2015 to October 2018) linked to Los Angeles County integrated InfoHub records from seven named county agencies, against a community-defined two-year adverse-event composite outcome — emergency or inpatient visits, crisis-stabilization episodes, justice involvement, substance-use-disorder diagnoses, or death. It was built by the Coordinated Entry System Triage Tool Research and Refinement (CESTTRR) project (March 2020 to May 2023), a community-based effort led by USC's Center for AI in Society with the California Policy Lab at UCLA, at the prompting of LAHSA's Ad Hoc Committee on Black People Experiencing Homelessness, and governed at design time by a Community Advisory Board and Core Planning Group.
The research reason for the swap was stark. The incumbent VI-SPDAT scored barely above chance at identifying vulnerability (an area-under-curve of 0.54 against a 0.50 coin flip) and was racially biased: its generalized false-negative rate was 54% for white participants, with gaps up to 8.5 percentage points for Black, Latinx, and Native Hawaiian or Pacific Islander participants, so vulnerable clients of color were systematically under-scored. The revised tool was deliberately made less accurate to make it fairer. An accuracy-only version reached 0.64; the fielded equity-adjusted tool sits at 0.60, closing the race and ethnicity false-negative gaps from 5.9 to 0.7 percentage points for Black clients and 3.2 to 0.2 for Latinx clients, with 57% of its top decile experiencing an adverse outcome versus 49% under the VI-SPDAT. The report is candid that the instrument sits well below the 0.831 that a hypothetical data-rich model could reach, and below a best-possible generalized false-negative rate of 42%. The item count was cut from 35 to 19: the algorithm selected 9 items and roughly 10 were reinstated by the community boards, which the report says did not meaningfully change accuracy or equity. Two design choices matter for what follows. The deployed tool is a self-report questionnaire, administrable by pencil and paper and then keyed into a Homeless Management Information System, with no live administrative-data feed — the county-data linkage was used only to derive the weights. And a separate USC-built housing-matching optimization algorithm was deliberately not deployed: Los Angeles chose to swap only the scoring instrument and keep human matching discretion.
The distinctive event is the transition, not the tool. In 2024 the CES Policy Council approved the LA HAT for permanent-supportive-housing (PSH) consideration alongside the VI-SPDAT, with existing VI-SPDAT scores kept valid to reduce the disruption of the change. LAHSA fully launched implementation in 2025: 94 training sessions run with the partner A.C.T.I.O.N. to Healing, 3,139 staff trained past a 3,000 goal, access rolled out across 688 eligible HMIS programs, and the tool translated into the county's nine threshold languages. During this dual-tool phase the PSH-consideration thresholds were 8-plus on the VI-SPDAT and 17-plus on the LA HAT. By LAHSA's account, initial quantitative data and provider feedback showed participants were more likely to obtain an eligible score under the VI-SPDAT, so direct-service providers opted to administer the VI-SPDAT over the LA HAT — a trend LAHSA says "perpetuated the racial bias of the VI-SPDAT in the System" and slowed the new tool's uptake. The correction was governance, not a better model or a retraining memo. On April 22, 2026 the CES Policy Council lowered the LA HAT threshold from 17-plus to 12-plus, ruled that where both scores exist the most recent LA HAT score governs, and ordered programs with LA HAT access to discontinue new VI-SPDAT completions; the VI-SPDAT was deactivated for those programs on May 1, 2026 and removed system-wide for new completions on June 30, 2026, with the phase-out of legacy VI-SPDAT scores to be decided in Fall 2026. The current CES PSH Prioritization and Matching Guidance (revised May 27, 2026) states plainly that "the VI-SPDAT is being phased out of the System due to its inherent bias," while noting that participants eligible under either tool on April 22, 2026 are not made retroactively ineligible.
Two caveats frame the whole account, and the developers were among the first to state them. Every equity and accuracy figure above is a pre-deployment estimate on 2015 to 2018 held-out historical data, not an observed post-launch outcome; no post-deployment outcome evaluation has been published, though LAHSA's implementation milestones name a first-phase implementation evaluation commissioned from Arc4Justice for 2026. And LAHSA's claim that the dual-tool thresholds produced disparate eligibility rates cites initial quantitative data without releasing the underlying numbers, so the direction is documented but the magnitude is not. The allocation environment gives the scoring real weight: as of late 2024 reporting, for every available slot for permanent supportive housing in Los Angeles County about four more are needed, roughly 75,000 people are unhoused (up from about 53,000 in 2018) and about 17,000 are waiting for PSH, and Black people are under 10% of the county population but over 30% of the unhoused. Community members with lived experience shaped and also criticized the effort: advisory-board member Reba Stevens argued that "everybody is vulnerable" and that scoring cannot fix the underlying housing shortage, while lead researcher Eric Rice acknowledged that he had helped make "a system that is inadequate... fair, or more fair," not a solved one. This case is the successor-and-transition event; the Atlas's separate VI-SPDAT entry documents the legacy instrument itself.
The sociotechnical reading
Two of the Atlas's housing cases stand on either side of this one. The VI-SPDAT case is the discredited instrument this system retired. The Allegheny Housing Assessment is another county's answer to the same problem — a warehouse-data score that got more accurate yet did not close the racial gap downstream. This case is the thing between them that neither captures: the handover itself. And the handover has its own failure mode, one that has nothing to do with whether the new score is any good.
Los Angeles did the hard part right. It commissioned three years of community-based research, linked tens of thousands of records to build a fairer instrument, and made a deliberate, negotiated choice to trade some accuracy for equity — accepting a lower area-under-curve to close the false-negative gaps that had systematically under-scored people of color. It even declined to deploy an available matching algorithm, keeping human discretion in the loop. By every measure of tool design, the successor was more careful and fairer than what it replaced. And then, during the transition, the retired bias walked back in — not through the model, but through the gap between two live channels. For a stretch, both instruments ran in parallel, and they qualified people at different thresholds: a client was likelier to score high enough for housing consideration under the OLD, biased tool. That single fact handed frontline workers a choice they had never had before — which instrument to administer — and, facing a scarce queue where getting a client over the line is the whole job, they reached for the easier one. By the system's own account, providers opted for the VI-SPDAT, and the bias the replacement was built to remove kept ordering the list.
On the system map this is a shape no other Atlas cell has: not one model whose output flows to a decision, but two competing scoring channels feeding one operator who chooses the channel. The governance question it raises is not "is the new score fair?" — by its builders' pre-deployment measures, it was — but "while two instruments run side by side, who reconciles them, and who notices that the easier one is the biased one?" The answer, for a costly stretch, was no one: nothing reconciled the two channels' comparative eligibility output, and no standing oversight watched which tool was being administered and to whom, so the arbitrage ran until initial quantitative data surfaced it. What resolved it is telling. Not a sharper model — the model was not the problem. The council recut the new tool's threshold to align the two channels, forced the biased channel's deactivation, and ruled the new score supersedes the old. The instruments that fit this shape are the ones for a transition, not a tool: reconcile the two channels while they run in parallel, so the arbitrage shows up as a disagreement; shut off the retired channel once you have decided to retire it, so there is no easier-qualifying escape hatch; and make each rank legible about which instrument produced it, because a score from a biased tool and a score from a corrected one are not the same evidence on a shared list. The distinct lesson the Atlas draws here: replacing a biased judge is a governance problem in the handover, not just a modeling problem in the tool — when an old instrument and its fairer successor run at mismatched eligibility thresholds, the operator's instrument choice becomes an arbitrage, and if the discredited tool is the easier qualifier, the retired bias re-enters through the transition no matter how good the replacement is. A better successor is still just one of two live channels; the safeguard is reconciling them, deactivating the one you retired, and keeping the ranks legible. The honest boundary throughout: served people experiencing homelessness are not modeled in the paired Lab; the harm surface is institutional — a biased channel capturing decisions during a transition, a persisted legacy score on the list, an unwatched eligibility gradient — and every equity and accuracy figure is a pre-deployment estimate on historical data, the LA HAT bias-reduction numbers the builders' predicted values rather than observed outcomes, with the first post-deployment evaluation still to come.
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.