Skip to content
ParamergeParamerge
Stephen Lieberman
Through Paramerge
Contact

PAN Lab example

Benefits Data Trust wind-down

The node that could not be kept: winding down a benefits-navigation nonprofit

For twenty years this nonprofit sat between low-income people and the agencies that run public benefits: it used agency data-sharing to find likely-eligible people, ran a call center where navigators screened callers across many programs at once, and stayed on the line through submission. In 2023 it reported helping more than 120,000 people. Modeled on the Benefits Data Trust wind-down. Nothing here decides anything and nothing here is a scoring model — that is the point. In June 2024 its board voted to wind the whole node down in sixty days, and the danger is not a wrong answer. It is what happens to the links a deleted intermediary held: the referral partners whose access ran through it, and the agency data-sharing store whose fate was left unresolved. Watch the hand-off that was never built, and watch the store no one wound down.

Stylized model of a documented deploymentBenefits navigation & public-facing chat

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 Intermediary-node-deletion of a benefits-navigation nonprofit network: 6 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: 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 intermediary-node-deletion pattern documented in the Benefits Data Trust wind-down case file - not a reconstruction of the actual organization, its call center, or its Community Prism screening tool. The atlas-relevant object is the node itself and its links, not a model inside it: Benefits Data Trust was a data-driven navigation intermediary, not an algorithmic decision or risk-scoring system, and the record carries no error, override, or accuracy metrics.

  • baseline

    The binding governance constraint was the funding-authority actor: a nonprofit board with unilateral, unreviewable authority over the intermediary's continuity, sitting atop philanthropy-subsidized unit economics (funded roughly two-thirds by philanthropy). The board is drawn as oversight holding continuity authority, and the case turns on the fact that no agency partner, funder, or regulator held an edge to observe or veto its wind-down vote. The staff-sourced per-application cost figures behind the unit-economics diagnosis (a few hundred dollars per successful application versus about two dollars for online self-service) come from anonymous staff sourcing in the Technical.ly post-mortem and are treated as staff-attributed, not audited.

  • baseline

    The modeled harm channel is dependency loss for referral partners. The intermediary's partner network reached benefits through this single node, so the empty peer hand-off pathway is the defining absence: the self-imposed 60-day clock foreclosed an orderly acquisition or transition, and the work rerouted to higher-friction channels (redistribution across partner agencies, a state subcontractor, months-long referral waits). peer-governance and oversight-cadence open an orderly, dated hand-off. The claim that the 60-day timeline was self-imposed and avoidable is contested ex-employee framing carried honestly, not adjudicated fact; the board's only official explanation was a perfect storm of circumstances.

  • baseline

    The privacy-side absence is the orphaned data-sharing store, drawn as the empty data-leaving pathway: the agency data-sharing links and the custom screening tools and chatbots had reported-unresolved fate at closure. It is drawn empty at baseline because that is the pathway an abrupt wind-down opens and a governed one closes. data-minimization (bounding and ageing down what is kept and how identifiable it is) and connection-auth (vetting or closing the links out) act on it, and oversight-cadence sets the dated teardown schedule; a content-blind record-purge does not close it. The store's inflow is a consent-adjacent one - identifiable benefits data used for targeted outreach under data-sharing agreements - so the record-to-model and staff-to-model pathways are marked privacy-sensitive.

  • assumed

    This is a node deletion rather than a node failure: the intermediary was reported to be exceeding its contract performance requirements to its final months, then was removed from the network in 60 days by its own governing board. The danger on this map is therefore not a runaway model error - the screening engine's low error is a modest modeling choice for a human-mediated assistive workflow, not a measured rate - but what happens to the accumulated links and the dependent partner network when a running intermediary is decommissioned, and whether the exit is executed on a dated schedule with an orderly hand-off or left abrupt.

  • assumed

    Served households, and the benefits they do or do not ultimately receive, are not in the dynamics; this Lab models institutional propagation only. The backdrop figure of more than 450 million dollars per year in benefits going unclaimed by eligible Philadelphians is a Philadelphia-only estimate cited via a funder-affiliated analysis, not a measured post-closure enrollment effect, and no counterfactual enrollment-decline measurement has been published. A call, an application, or a reroute on this map is an institutional signal, never a person.

What this example does not show

  • Served households, and the benefits they do or do not ultimately receive, are not modeled here; the Lab models institutional propagation only. The backdrop figure of more than $450 million a year in benefits going unclaimed by eligible Philadelphians is a Philadelphia-only estimate cited via a funder-affiliated analysis, not a measured post-closure effect, and no counterfactual enrollment-decline measurement has been published.
  • Benefits Data Trust was a navigation intermediary, not an algorithmic decision or risk-scoring system, so there are no published error, override, or accuracy metrics; the per-application cost figures come from anonymous staff sourcing in a post-mortem and are treated as staff-attributed, not audited. The board's only official explanation was "a perfect storm of circumstances," and the ex-employee claim that the 60-day timeline was self-imposed and avoidable is contested framing carried honestly, not adjudicated fact.

Sources and evidence

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

  • In June 2024 the board of Benefits Data Trust, a Philadelphia benefits-navigation nonprofit that reported helping more than 120,000 people access about $182 million in benefits in 2023, voted unanimously to wind the organization down within a self-imposed 60-day window, citing only 'a perfect storm of circumstances'; the organization closed on August 24, 2024, laying off 273 employees, despite roughly $12 million in unrestricted reserves at the end of 2023 and about $32 million in projected 2024 revenue.

    empirical
    • Investigative Brubaker, Benefits Data Trust is shutting down in 60 days (The Philadelphia Inquirer, 2024) https://www.inquirer.com/health/benefits-data-trust-bdt-shutting-down-20240625.html
    • Investigative Brubaker, Benefits Data Trust is leaving employees and supporters in the dark over its abrupt closure (The Philadelphia Inquirer, 2024) https://www.inquirer.com/health/benefits-data-trust-bdt-surprise-closure-philadelphia-20240627.html
    • Investigative Wink, Why Benefits Data Trust fell apart despite millions from philanthropy and government contracts (Technical.ly, 2024) https://technical.ly/civic-news/benefits-data-trust-shutdown-trooper-sanders/
    • Investigative Mosbrucker-Garza, Philly's Benefits Data Trust shutters after 20 years. Laid-off workers say they still want answers (WHYY News, 2024) https://whyy.org/articles/philadelphia-benefits-data-trust-closure-employees-laid-off/

  • The closure left active government partnerships without a designated successor, including a Pennsylvania Department of Aging workload of nearly 48,000 applications from 27,018 households in the final year and a Philadelphia BenePhilly call-center contract the organization was reported to be exceeding through mid-2024; the navigation function fragmented to higher-friction channels, with the work redistributed across partner agencies and a subcontractor and referral waits reported as several months, which a Pew analyst described as a 'cascading effect.'

    empirical
    • Investigative Brubaker, What the loss of Benefits Data Trust means for two government agencies in Harrisburg and Philly (The Philadelphia Inquirer, 2024) https://www.inquirer.com/health/benefits-data-trust-closing-august-23-20240823.html
    • Trade press Burnley, After the abrupt closure of Benefits Data Trust, Philly nonprofits are stepping up to fill in the gaps (Technical.ly and The Philadelphia Citizen, 2024) https://technical.ly/civic-news/philadelphia-senior-care-benefits-navigation/
    • Investigative Mosbrucker-Garza, Philly's Benefits Data Trust shutters after 20 years. Laid-off workers say they still want answers (WHYY News, 2024) https://whyy.org/articles/philadelphia-benefits-data-trust-closure-employees-laid-off/

Where this connects

Institutional pressures in this domain

  • Workload surge — Demand outruns staffing; per-case attention shrinks and review becomes triage.
  • 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).
  • Reviewer bottleneck — One fixed-capacity checking stage sits between AI output and consequence; everything queues behind it.
  • Compliance over substance — Paper controls (sign-offs, checklists) satisfy audits while the behavior they describe erodes.

All of them in context on the Benefits navigation & public-facing chat domain page.

Levers available here and the patterns behind them

Documented case histories