Skip to content
ParamergeParamerge
Stephen Lieberman
Through Paramerge
Contact

PAN Lab example

SafeRent Tenant Screening Score

SafeRent Score: the number the leasing desk could not change

A screening company turns a rental applicant's file into a number between 200 and 800 and an accept, decline or conditional recommendation, measured against a cutoff the property company picked. Modelled on the SafeRent Score litigation: its shape, not the real system. Watch where the authority sits. The vendor alone weights the factors, and its marketing told landlords they cannot change the algorithm. Property management picks the cutoff without knowing how the score is computed. The leasing desk that signs the lease gets the bottom line and wrote to one applicant that it does not accept appeals and cannot override the outcome. That inversion is what a federal court reasoned from in 2023 when it let Fair Housing Act claims proceed against the vendor, holding that the company that builds and conceals the algorithm effectively controls the approval decision. The plaintiffs were housing-voucher holders, and their pleaded point is about which feed is missing: the housing authority pays the landlord directly, an average of 1,159 dollars a month against a tenant contribution of 423, and that payment is not one of the model's inputs. Credit history is. Disparate impact was never adjudicated here; the company settled for 2.275 million dollars without admitting liability and maintains its scores comply with all applicable laws. What the settlement built instead of a model fix is a gate on the output: for five years, no screening score and no accept or decline recommendation on a voucher application, with the score withheld on the other products unless the provider certifies the applicant holds no voucher, and a route back only if a named civil-rights organization finds the score valid for this population. Underlying records and third-party credit scores still flow. As of mid-2026 nothing in the public record shows that gate operating. The question is how you govern a decision whose owner is not the person making it. Before you pick a target level: this board cannot be won under Service and Safety Targets or All Governance Targets. The Lab offers this deployment every tool its own record supports, and the whole set costs 2.7 times the budget. Cost is not what blocks it. Containment is reached at the lowest effort, and two pathways still stay open: the vendor authoring the weights, and the provider setting the cutoff. No lever in this deployment's documented authority set closes either one, and the benefit reading stays well short of the margin it would need to clear. That is a measurement of the authority inversion this network is derived from, not a puzzle waiting to be cracked. Explore and Service Targets Only can be won.

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 Tenant-score-class voucher screening gate network: 13 components and 25 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: 5 assumed · 5 published baseline. In the Lab, the shaded evidence band behind each headline readout draws its width from the least-established class below.

  • assumed

    Binding framing. Disparate impact was never adjudicated on the merits: the screening company settled without admitting liability and publicly maintains that its scores comply with all applicable laws, which is a vendor claim and is labelled as one throughout. The disparity figures on this map (credit medians 612, 661 and 725; subprime shares 45.1, 31.5 and 18.3 percent; the 91,000 vouchers and 24 percent Black / 31 percent Hispanic demographics; the 73.26 percent payment share) are plaintiff-pleaded numbers citing Urban Institute, Department of Housing and Urban Development (HUD), Federal Reserve and Consumer Financial Protection Bureau (CFPB) data, accepted by the court as allegations at the motion-to-dismiss stage only. The court's holdings are drawn with their posture: the Fair Housing Act and chapter 151B claims survived, the chapter 93A counts were dismissed, and 'effectively controls' is the court's own language about the vendor.

  • baseline

    Three input-source components, and the third is the case. Credit bureau data and eviction/landlord-tenant court records are documented inputs and sit at full strength each: the model description names them, and the complaint pleads credit history as the score's principal input. The housing authority payment feed sits empty because the same source lists it under what is notably absent, and because the deployment's central pleaded defect is that the strongest protective factor in the real arrangement is not in the model's state. Drawing the feed empty rather than leaving it off the map is the honest form: the payment stream exists in this deployment and is measured in Department of Housing and Urban Development (HUD) data, and what the record documents is that it does not reach the model.

  • baseline

    Two staff-to-model pathways, neither from the front-line decision-maker. This map uses that pathway for documented operator input into the model's decision surface: the vendor's exclusive authorship of the weights at full strength (sole owner of the model; landlords told they cannot change the screening algorithm) and the housing provider's choice of cutoff and conditional band at a substantial level (a 500 minimum with a 450 to 499 band, chosen in consultation with the vendor and without knowledge of how scores are computed). The leasing desk that signs the lease has no pathway into the model at all, and its in-house second read is drawn empty on a written statement that appeals are not accepted and the outcome cannot be overridden. Those three facts together are the authority inversion the court's reasoning turned on, drawn as structure rather than asserted in prose.

  • baseline

    The output control is drawn as a guardrail plus a model-to-operator check at a low level, and that low value is derived, not chosen. It is above empty because the settlement agreement documents the control as existing and in force: practice changes fully effective no later than twelve months from the March 28, 2024 execution, obligations running five years from certification, under the court's continuing exclusive jurisdiction. It is not higher because the record bounds it three ways. Its scope is one population, voucher-holder applications. Its content is partial by construction: only the composite score and the accept or decline recommendation are withheld, while the underlying records and third-party credit scores continue to flow with disclosure of their source, so suppression shifts rather than removes the decision input. And its operation has never been observed: as of the research date of July 20, 2026 no compliance report, no enforcement motion and no validating-organization examination had been located in the public record.

  • baseline

    baselineDemand 3 comes from the documented flow: the settlement records show more than 18,000 Massachusetts applications scored below the housing provider's accept threshold between May 25, 2020 and September 27, 2023 (the records-pull window), roughly 5,400 a year, against a pleaded Massachusetts base of over 91,000 vouchers in use, about 7 percent of holders moving in a year, and over 1,500 Black and over 1,900 Hispanic voucher holders screened for an apartment per year. manualCapacity 1 comes from the documented counterfactual: the leasing office wrote that it does not accept appeals and cannot override the screening outcome, the leasing manager does not receive the detailed credit information at screening time, and the single reversal in the record took roughly six weeks and an outside organization drafting a second appeal. A human channel that produced one documented reversal against thousands of declines is a low counterfactual floor, not a competent one.

  • assumed

    The 18,000 figure is an upper-bound flow figure and is used as one. It counts Massachusetts applications scored below the property's accept threshold during the records window, not voucher holders denied: the screening company could not identify which of those applicants held vouchers or their race. The company's own October 2023 estimate of putative class members, after deduplication and estimating voucher-holding Black and Hispanic applicants, was between 3,300 and 4,200, and that band is the primary-source lower bound this map reads alongside the 18,000.

  • baseline

    Two checks are drawn empty and both are documented absences rather than blank space. The in-house second read of a decline: management staff told a named plaintiff in writing that they do not accept appeals and cannot override the outcome of the tenant screening. The tenancy-record reconciliation: records persist in the input stores and continue to depress scores, one plaintiff's denial rested partly on an eviction filed because the landlord wanted the unit back for a relative, and the record describes accuracy disputes reaching the record entry rather than the weighting that reads it. The one check the record shows working, the advocate appeal, drawn faint, carries its documented cost in the copy: roughly six weeks of organized advocacy for one reversal, worked from outside the company. Generalizing a reversal rate from that single case is unsupported.

  • assumed

    Served people are not in these dynamics and no served-person outcome is computed anywhere on this map. Rental applicants, voucher holders and their households are boundary-only: a score, a report or an appeal here is an institutional signal, never a person. The record's documented household harms live in the case file and are measured outside any diagram like this one, including a fallback apartment costing 200 dollars a month more with fewer amenities in a higher-crime area, and the hard 120-day initial voucher term that a denial consumes and that requires a discretionary extension or the loss of the voucher. The dossier also documents a denial-to-record loop in which a decline pushes a household into costlier housing whose financial strain credit-based inputs later read as risk; that loop runs through a person's life and is deliberately not drawn as a pathway, because doing so would compute a served-person outcome inside the dynamics.

  • assumed

    Documented actors this map does not draw, and why. Federal fair-housing agencies filed a statement of interest on January 9, 2023 arguing that the Fair Housing Act reaches algorithm-based tenant screening and screening companies, and the court's ruling followed that position; their documented action is a one-time legal-position filing rather than a recurring pathway into this deployment, so it is narrated in the supervising court's role rather than drawn as a reviewer that would need an invented inbound flow. Public housing authorities are documented as sitting entirely outside the screening loop, so they appear as the payment-records feed, empty, rather than as a group of staff with no pathways. No external boundary is drawn: the record documents no pathway by which applicant data leaves the governed system, and drawing a data-leaving pathway that no source describes would be decoration.

  • assumed

    Geographic scope is ambiguous in the record and is not resolved here. The practice-change text is not expressly limited to Massachusetts, but the classes, the notice population and the plaintiffs' framing are Massachusetts-centred, and some coverage characterizes the changes as company-wide. This map is drawn at the level of one screening product and its housing-provider customers and asserts no nationwide reach. The end of court supervision is likewise approximate: the five-year clock runs from a certification date that is not in the public record reviewed, so supervision runs to roughly 2029 or 2030.

What this example does not show

  • Served people are not modeled here; the Lab models institutional propagation only, and the outcomes that land on rental applicants and their households are documented in the case file and measured outside any diagram like this one. A score, a report or an appeal on this map is an institutional signal, never a person. The documented household harms sit outside these dynamics by construction: a fallback apartment costing 200 dollars a month more with fewer amenities in a higher-crime area, and a hard 120-day initial voucher term that each denial consumes and that requires a discretionary extension or the loss of the voucher. The record also describes a loop in which a denial pushes a household into costlier housing whose financial strain credit inputs later read as risk; that loop runs through a person's life and is deliberately not drawn as a pathway, because drawing it would compute a served-person outcome inside the dynamics.
  • Discrimination here is alleged and settled, never found. The disparate-impact merits were never adjudicated: the screening company settled for 2.275 million dollars without admitting liability and publicly maintains that its scores comply with all applicable laws, which is a vendor claim and is labelled as one. The statistical disparities used throughout — credit medians of 612, 661 and 725; subprime shares of 45.1, 31.5 and 18.3 percent; over 91,000 Massachusetts vouchers with about 24 percent Black and 31 percent Hispanic holders; the 73.26 percent payment share — are plaintiff-pleaded figures citing Urban Institute, HUD, Federal Reserve and CFPB data, accepted by the court as allegations at the motion-to-dismiss stage only. The rulings are stated with their posture: the Fair Housing Act and chapter 151B race and source-of-income claims survived on 26 July 2023, the chapter 93A consumer counts were dismissed, and 'effectively controls' is the court's own language.
  • The output control is described mechanically, and it has never been observed operating. What the settlement suppresses is the composite screening score and the accept or decline recommendation on a voucher application; what continues to flow is the underlying tenant-screening information and any third-party credit score, with disclosure of its source. Suppression shifts the decision input rather than removing it. A screening score may return to this channel if the National Fair Housing Alliance, or an organization class counsel and the company agree on, finds it valid for use with voucher holders. As of the research date of 20 July 2026 no compliance report, no enforcement motion and no such examination had been located in the public record, so nothing here should be read as evidence that the gate works, only that it exists on court-approved terms.
  • The 18,000 figure is an upper bound on flow, not a count of harm. It counts Massachusetts applications scored below the housing provider's accept threshold between 25 May 2020 and 27 September 2023, the records-pull window, and the company could not identify which of those applicants held vouchers or their race. Its own October 2023 estimate of putative class members, after deduplication and estimating voucher-holding Black and Hispanic applicants, was between 3,300 and 4,200 — the primary-source lower band this map reads alongside it. The single documented reversal supports the shape of an override cost and supports no reversal rate at all.
  • Scope and duration are both ambiguous in the record and are left ambiguous here. The practice-change text is not expressly limited to Massachusetts, but the classes, the notice population and the plaintiffs' framing are Massachusetts-centred and some coverage characterizes the changes as company-wide; nothing on this map asserts a nationwide reach. The five-year obligation period runs from a certification date that is not in the public record reviewed, so court supervision ends at roughly 2029 or 2030 rather than on a known day. Federal fair-housing agencies' January 2023 statement of interest is narrated in the supervising court's role rather than drawn as its own reviewer, because their documented action is a one-time legal-position filing and drawing it would require inventing a recurring pathway.

Sources and evidence

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

  • SafeRent Solutions (formerly CoreLogic Rental Property Solutions) sold landlords a Registry ScorePLUS tenant-screening model returning a single 200-800 'lease performance risk' score plus an accept/decline/conditional recommendation measured against a landlord-chosen cutoff (500 in the complaint's worked example, with an optional 450-499 conditional band), built from credit bureau reports and scores including non-tenancy debt, bankruptcy records, past-due accounts, payment performance, and eviction and landlord-tenant court records, weighted 'according to their statistical significance in predicting lease performance'; factor weights were undisclosed to landlords, applicants, and the public, and the company's marketing told buyers a landlord cannot change the screening algorithm. The plaintiffs' pleaded theory concerns a missing input: as of 2021, HUD data cited in the complaint shows Black and Hispanic voucher holders in Massachusetts paying an average of $423/month toward rent and utilities while public housing authorities paid landlords an average of $1,159/month directly (at least 73.26% of the expected payment), on tenancies averaging over 21 years in the same unit, and that subsidy is not a model input. These are plaintiff-pleaded figures accepted by the court as allegations at the motion-to-dismiss stage only.

    empirical
    • Government Louis v. SafeRent Solutions LLC, Class Action Complaint, No. 1:22-cv-10800-AK (D. Mass. 2022) https://clearinghouse-umich-production.s3.amazonaws.com/media/doc/160025.pdf
    • Government Louis v. SafeRent Solutions LLC, Memorandum and Order on Defendants' Motions to Dismiss, No. 1:22-cv-10800-AK, Doc. 64, 685 F. Supp. 3d 19 (D. Mass. 2023) https://www.justice.gov/crt/media/1310736/dl
    • Reference FindLaw Caselaw, Louis v. SafeRent Solutions LLC, full-text opinion (D. Mass. 2023) (2023) https://caselaw.findlaw.com/court/us-dis-crt-d-mas/114706064.html

  • The authority over a SafeRent-screened tenancy decision was documented as inverted: the vendor authored the weights and returned the answer without any housing relationship to the applicant, property management picked the cutoff 'in consultation with SafeRent' without knowing how scores are computed, and the leasing desk that signed the lease received only the bottom-line score ('The Leasing Manager does not receive the detailed credit information at the time of running the applicant screening'; 'CoreLogic sends us a number, and if it is above the predetermined approved number, we move forward... We do not know why they were denied') and disclaimed override authority to a named plaintiff in writing ('we do not accept appeals and cannot override the outcome of the Tenant Screening'). On July 26, 2023 Judge Angel Kelley denied the motions to dismiss the FHA 3604(a),(b) and Massachusetts c.151B race and source-of-income claims on the reasoning that SafeRent 'effectively controls' approval decisions because it alone builds and conceals the algorithm, while dismissing the c.93A consumer counts. The single documented pre-settlement reversal ran outside the screening system entirely: a July 22, 2021 denial, a personal appeal rejected August 26, and a second appeal drafted by the tenant-advocacy organization City Life/Vida Urbana delivered September 1-8, 2021 — roughly six weeks of organized advocacy for one reversal, from which no reversal rate can be generalized.

    empirical
    • Government Louis v. SafeRent Solutions LLC, Class Action Complaint, No. 1:22-cv-10800-AK (D. Mass. 2022) https://clearinghouse-umich-production.s3.amazonaws.com/media/doc/160025.pdf
    • Government Louis v. SafeRent Solutions LLC, Memorandum and Order on Defendants' Motions to Dismiss, No. 1:22-cv-10800-AK, Doc. 64, 685 F. Supp. 3d 19 (D. Mass. 2023) https://www.justice.gov/crt/media/1310736/dl
    • Government United States Department of Justice and Department of Housing and Urban Development, Statement of Interest of the United States, Louis et al. v. SafeRent et al. (2023) https://www.justice.gov/d9/2023-01/u.s._statement_of_interest_-_louis_et_al_v._saferent_et_al.pdf
    • Reference Civil Rights Litigation Clearinghouse, Louis v. SafeRent Solutions LLC, No. 1:22-cv-10800 (D. Mass.) case page (2025) https://clearinghouse.net/case/45888/

  • The Louis v. SafeRent settlement (executed March 28, 2024; finally approved November 20, 2024; $2.275M total, $1.175M fund with no reversion) remedied the case on the output channel rather than in the model: for five years SafeRent may return no SafeRent Score, no other tenant screening score, and no accept/decline recommendation on a report for a voucher-holder application, providing a report of underlying information instead (3.5.2); for the 'market' and 'no-credit' products the score is suppressed by default unless the landlord affirmatively certifies the applicant is not a voucher recipient (3.5.3); any tenant screening score may re-enter the voucher channel only if 'found to be valid when used for voucher-holders by the National Fair Housing Alliance' or another organization mutually agreed by class counsel and SafeRent (3.5.5(ii)(1)); third-party credit scores (FICO, VantageScore) may still be passed through with source disclosure (3.5.5(iii)); customer training is required (3.5.4); and the court retains continuing exclusive enforcement jurisdiction for five years from SafeRent's certification (6.9), approximately through 2029-2030. The geographic reach of the practice-change terms is ambiguous: 3.5.2 and 3.5.3 are not expressly limited to Massachusetts, while the classes, the notice population, and the plaintiffs' framing are Massachusetts-centered, and some coverage characterizes the changes as company-wide. Suppression therefore shifts rather than eliminates the decision input. SafeRent settled without admitting liability and states it 'continues to believe the SRS Scores comply with all applicable laws' (vendor claim); as of 2026-07-20 no public post-settlement compliance report, enforcement motion, or validating-organization event had been located, so the gate exists on court-approved terms and has never been observed operating. The >18,000 figure is Massachusetts applications scored below the housing provider's accept threshold between May 25, 2020 and September 27, 2023 (the records-pull window), an upper-bound flow figure that does not identify voucher status or race, paired with SafeRent's own October 2023 estimate of 3,300-4,200 putative class members.

    empirical
    • Government Louis v. SafeRent Solutions LLC, Memorandum in Support of Unopposed Motion for Preliminary Approval of Class Action Settlement, with the executed Settlement Agreement as Exhibit 1, Doc. 114 and 114-1 (D. Mass. 2024) https://www.cohenmilstein.com/wp-content/uploads/2022/05/SafeRent-Memo-in-Support-of-Unopposed-Motion-to-Settle-and-Certify-the-Classes-and-Exhibits-March-28-2024.pdf
    • Government Epiq Class Action Services, Louis et al. v. SafeRent official settlement website and frequently asked questions (2024) https://matenantscreeningsettlement.com/Home/FAQ
    • Investigative Associated Press via Fortune, Renter Scoring Firm Agrees to Pay 2.2 Million Dollars to Settle Case Accusing Its Algorithm of Discriminating on Race and Income (2024) https://fortune.com/2024/11/21/renter-scoring-saferent-million-settle-case-algorithm-discriminating-race-income/
    • Reference Greater Boston Legal Services, Cohen Milstein Sellers and Toll, and National Consumer Law Center, Rental Applicants Using Housing Vouchers Settle Ground-Breaking Discrimination Class Action Against SafeRent Solutions (2024) https://www.gbls.org/sites/default/files/2024-04/SafeRent-press-release-settlement-reached-4-26-2024.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