Skip to content
ParamergeParamerge
Stephen Lieberman
Through Paramerge
Contact

PAN Lab example

LA's coordinated-entry triage revision

Two scores in one queue: the transition that let a retired bias back in

The largest homeless-services system in the country retires a triage survey that barely beat a coin flip and systematically under-scored people of color, and replaces it with a fairer score built on tens of thousands of linked county records. Modeled on Los Angeles's coordinated-entry triage revision. But the two instruments run side by side for a while, and they qualify people at different thresholds — a client is likelier to score high enough for housing under the OLD, biased tool. So frontline workers keep reaching for it. The retired bias walks back in through the transition. The real question is not whether the new score is good enough. It is whether anyone reconciles the two channels while they run in parallel, whether the easier-qualifying legacy channel gets shut off, and whether the readers can see which tool a rank came from.

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 LA-HAT-class coordinated-entry triage-revision system network: 6 components and 12 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: 3 assumed · 3 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 two-channel triage-transition pattern documented in the LAHSA coordinated-entry triage-revision case file - not a reconstruction of the actual tool. The atlas-relevant object is the topology: two scoring instruments (the corrected, equity-adjusted LA HAT and the discredited legacy Vulnerability Index-Service Prioritization Decision Assistance Tool (VI-SPDAT)) running in parallel during a dual-tool transition, with the frontline worker choosing which to administer. This is the successor-and-transition event, distinct from the library's separate VI-SPDAT case (the legacy instrument itself) and from the un-deployed USC housing-matching optimization algorithm. Any reading that treats this as a single-score cell misses the point: the engine here is the operator's channel choice between two instruments, not one score's accuracy.

  • baseline

    The defining structural property is instrument arbitrage under mismatched cross-tool thresholds. During the dual-tool phase the PSH-consideration thresholds were 8+ on the Vulnerability Index-Service Prioritization Decision Assistance Tool (VI-SPDAT) and 17+ on the LA HAT; LAHSA states that 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 worker retraining: on 22 April 2026 the CES Policy Council lowered the LA HAT threshold to 12+, ruled the LA HAT score supersedes a coexisting VI-SPDAT score, and forced deactivation of new VI-SPDAT completions (removed for LA HAT-access programs 1 May 2026 and system-wide 30 June 2026), with the legacy-score phase-out to be decided in Fall 2026.

  • baseline

    The two load-bearing absences are drawn as inactive checks. First, the independent model check: while the two instruments ran side by side there was no standing reconciliation of their comparative eligibility output, so the mismatched thresholds let the easier-qualifying legacy tool capture the decisions unseen until quantitative data surfaced it. Second, the independent peer evaluation: there was no standing oversight of provider instrument-choice or equity monitoring of who was scored on which tool, so the arbitrage ran until detected rather than being caught early. These are the case's distinctive safety shape - a genuinely careful, community-governed redesign (a 133-page technical report, a deliberately negotiated accuracy-equity trade-off, community-reinstated items, trauma-informed administration) whose transition failed on a surface no one was watching: the eligibility gradient between the old channel and the new one.

  • baseline

    The corrected instrument was deliberately made less accurate to make it fairer, and the report is candid about its ceiling. The equity-adjusted LA HAT sits at AUC 0.60, down from an accuracy-only 0.64, 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 (pre-deployment estimates on 2015 to 2018 held-out data); the report bounds it well below the 0.831 AUC of a hypothetical data-rich model and a best-possible generalized false-negative rate of 42 percent. So this is not a story about a bad model: the successor was more careful and fairer than what it replaced, and it still failed to take hold - because the transition, not the model, was where the bias re-entered. That is exactly why a sharper or more accurate successor (improve-model) is the red herring here.

  • assumed

    The privacy surface is the self-report intake. The deployed instrument is a self-report questionnaire administered by a case manager and keyed into the Homeless Management Information System (HMIS) (or the domestic-violence-comparable database); it has no live administrative-data feed, so there is deliberately no record-to-model pathway - the county-data linkage was used only to derive the weights, not to score a person live. The identifiable client self-report (housing history, health, justice involvement, wellness) is entered into a shared record on which the prioritization list is built. data-minimization and write-gate act on that intake write; there is no automated denial and no individual-case override of the computed score - the threshold gates consideration mechanically, and the human-judgment floor (case conferencing as an alternate route, retained matching discretion) sits outside the score.

  • assumed

    Served people experiencing homelessness are not in the dynamics; this Lab reads institutional propagation only. No allocation or housing outcome to any person is computed here, and the harm surface the model reads is an institutional one - a biased channel capturing decisions during a transition, a persisted legacy score shaping the list, an unwatched eligibility gradient - never an individual determination. The demographic harm is documented outside any diagram like this one: every equity and accuracy figure is a pre-deployment estimate on 2015 to 2018 held-out historical data, the LA HAT bias-reduction numbers are the builders' predicted values rather than observed post-launch outcomes (no post-deployment outcome evaluation has been published, though Arc4Justice is commissioned to publish a first-phase implementation evaluation in 2026), and LAHSA's disparate-eligibility claim cites initial quantitative data without releasing the underlying numbers. A score, a rank, or a threshold on this map is an institutional signal, never a person.

What this example does not show

  • Served people experiencing homelessness are not modeled here; the Lab reads institutional propagation only. No allocation or housing outcome to any person is computed, and the harm surface is institutional - a biased channel capturing decisions during a transition, a persisted legacy score on the list, an unwatched eligibility gradient - never an individual determination. The demographic harm the case is about is documented outside any diagram like this one.
  • Every equity and accuracy figure is a pre-deployment estimate on 2015 to 2018 held-out historical data, not an observed post-launch outcome: the Vulnerability Index-Service Prioritization Decision Assistance Tool (VI-SPDAT)'s near-chance area under the curve (AUC of 0.54) and false-negative gaps (up to 8.5 percentage points), and the corrected LA HAT's AUC (0.60, deliberately traded down from 0.64 for equity) and its predicted gap reductions (Black 5.9 to 0.7, Latinx 3.2 to 0.2), all come from the CESTTRR research report. No post-deployment outcome evaluation has been published, though a first-phase implementation evaluation is commissioned for 2026. A safe-looking baseline is a property of this model, not a safety promise for any real deployment.
  • LAHSA's central claim - that the mismatched dual-tool thresholds made clients likelier to qualify under the Vulnerability Index-Service Prioritization Decision Assistance Tool (VI-SPDAT), so providers administered it and this perpetuated the tool's racial bias in the System - is a first-party statement that cites initial quantitative data and provider feedback without releasing the underlying numbers; the direction is documented but the magnitude is not. The implementation timeline is best stated as a phased 2024 to 2026 rollout rather than pinned to a single go-live date.
  • This is the successor-and-transition event, distinct from the Atlas's separate Vulnerability Index-Service Prioritization Decision Assistance Tool (VI-SPDAT) case (the legacy instrument itself, cross-referenced rather than re-litigated here) and from the separate University of Southern California (USC)-built housing-matching optimization algorithm, which Los Angeles deliberately did not deploy - it swapped only the scoring instrument and kept human matching discretion. The deployed tool is a self-report questionnaire with no live administrative feed; its atlas relevance is the two-channel transition dynamic and the instrument-arbitrage it created, not a claim about model internals.

Sources and evidence

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

  • 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.

    empirical
    • Academic Rice, Milburn, Vayanos, Rountree, Hill, Petering, Blackwell, Santillano and colleagues, CESTTRR Coordinated Entry System Triage Tool Research and Refinement Final Report (USC Center for Artificial Intelligence in Society, 2023) https://cais.usc.edu/wp-content/uploads/2023/11/CESTTRR-Final-Report-2023.pdf
    • Government Los Angeles Homeless Services Authority, Los Angeles Housing Assessment Tool (LA HAT) (2025) https://www.lahsa.org/news?article=1033-los-angeles-housing-assessment-tool-la-hat-

  • During the dual-tool transition the two instruments' PSH-consideration thresholds were 8-plus on the VI-SPDAT and 17-plus on the LA HAT, and 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 it, a trend LAHSA states 'perpetuated the racial bias of the VI-SPDAT in the System'; on April 22, 2026 the CES Policy Council lowered the LA HAT threshold to 12-plus, ruled the most recent LA HAT score supersedes a coexisting VI-SPDAT score, and forced deactivation of new VI-SPDAT completions (for LA HAT-access programs on May 1, 2026 and system-wide on June 30, 2026), though LAHSA has not released the underlying eligibility-rate numbers.

    empirical
    • Government Los Angeles Homeless Services Authority, Los Angeles Housing Assessment Tool (LA HAT) (2025) https://www.lahsa.org/news?article=1033-los-angeles-housing-assessment-tool-la-hat-
    • Government Los Angeles Homeless Services Authority, Los Angeles Housing Assessment Tool (LA HAT) Spring 2026 Implementation Updates (2026) https://www.lahsa.org/documents?id=9877-los-angeles-housing-assessment-tool-la-hat-implementation-improvements-spring-2026-
    • Government Los Angeles Homeless Services Authority and the LA CES Policy Council, CES Permanent Supportive Housing Prioritization and Matching Guidance (2026) https://www.lahsa.org/documents?id=7658-ces-psh-prioritization-and-matching-guidance-effective-07-01-2026-.pdf

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