Skip to content
ParamergeParamerge
Stephen Lieberman
Through Paramerge
Contact

PAN Lab example

GetCalFresh

Most of a state's online intake with no authority at all: an assisted-application node and its handoff

For six years, most of a state's online food-benefit applications went through a web form built by a nonprofit - not a government system, not a scoring model, not a chatbot that decides anything. Modeled on GetCalFresh. It made no determinations at all: every application it helped complete was handed to a county caseworker who interviewed the applicant and decided, and anyone could skip the node entirely at a county office. That is exactly why it was safe to run - and why the real question is not whether it is accurate, but what happens when over 70% of a state's online intake depends on one voluntary node with no designated backstop. Watch the succession that was never built into the design, and the one that finally was.

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 GetCalFresh-class assisted-application navigation node network: 5 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 assisted-application navigation-node pattern documented in the GetCalFresh case file - not a reconstruction of the actual service. The atlas-relevant object is the topology, not a model inside it: GetCalFresh was a deterministic, structured-workflow digital assister (a guided intake form, a document uploader, a reminder scheduler, live chat, and a CBO portal), not a machine-learning, generative, or scoring system, and the record carries no error, override, or accuracy metric because it made no automated decisions. Any reading that implies algorithmic decision-making misreads it.

  • baseline

    The defining structural property is a single point of dependency: between 2019 and 2025 more than 70% (about 73% per the ten-year retrospective) of California's online Supplemental Nutrition Assistance Program (SNAP) applications were submitted through one nonprofit-built node sustained by philanthropy and a state partnership rather than statute. That concentration is carried by the single dominant intake pathway, drawn at full strength, a topology property, not by demand pressure. The headline scale figures (6.2 million people, over 12.8 billion dollars, over 70% of online applications) are organization-published by Code for America and not independently audited.

  • baseline

    The load-bearing control is a statutory determination floor entirely outside the node: the county caseworker conducts an independent intake interview and verifies the submitted application against county administrative records before making every determination, and an applicant can bypass the node entirely at a county office. That is why the county-to-assister check is drawn strong, at full strength - a step above the US sibling copilots' individual-habit checks and structurally different from Caddy's inside-the-org supervisor gate. This is the case's distinctive safety shape: a critical node that holds no authority over any outcome.

  • baseline

    The node's measured influence runs entirely through completion, timing, and churn on the applicant side, never through a decision: a peer-reviewed randomized controlled trial of roughly 65,000 Los Angeles applicants found access to applicant-initiated flexible interviews raised approvals by about 6 percentage points, and an in-house reminder experiment reported lifting renewal submissions among prior non-responders from about 1.5% to about 12%. These are served-people outcomes recorded in the case file and measured outside any diagram like this one; they are not computed here, and the reminder figures are organization-published without sample sizes or confidence intervals.

  • assumed

    The defining latent absence is a dormant succession/backstop pathway (the assister-to-successor peer check, empty at baseline): for six years no statutory backstop was designated for the navigation function - county-office and paper channels persisted throughout, roughly 27% of online applications flowed through other online channels the whole time, and BenefitsCal itself rolled out across 2021-2023 - but no plan covered where the dominant online channel's volume would go. It is drawn empty at baseline because that is the pathway a governed exit opens - CDSS (the state social-services department that had adopted the node as the statewide assister in 2019) directed a dated, five-phase transfer of the node's functions into the state-owned BenefitsCal portal in 2024 and 2025. peer-governance and oversight-cadence build and schedule the orderly hand-off, and connection-auth authorizes the successor path one grant at a time. This is the exact pathway the intermediary-node-deletion sibling leaves closed (an abrupt board-vote deletion); here it is the positive counterexample, opened on a dated schedule into government ownership.

  • assumed

    Served people, and the benefits they do or do not ultimately receive, are not in the dynamics; this Lab reads institutional propagation only. The participation-rate rise (66% of eligible people in 2014 to 81% in 2022) is correlational context Code for America presents alongside its growth, not a proven causal effect, and no post-sunset participation measurement has been published. The Benefits Data Trust collapse is an analogical counterfactual about a different organization with a different funding structure, not evidence about GetCalFresh, whose wind-down was planned and state-coordinated - the opposite failure mode. A completion, an application, or a reroute on this map is an institutional signal, never a person.

What this example does not show

  • Served people - the applicants, and the benefits they do or do not ultimately receive - are not modeled here; the Lab reads institutional propagation only. The measured completion effects (the flexible-interviews trial's roughly 6-percentage-point approval increase, the reminder experiment's rise from about 1.5% to about 12% of lapsed renewals) are served-people outcomes reported in the case file and measured outside any diagram like this one, never computed here.
  • The headline scale figures - 6.2 million people helped, more than $12.8 billion in benefits, over 70% (about 73%) of the state's online Supplemental Nutrition Assistance Program (SNAP) applications - are organization-published by Code for America and not independently audited; the participation-rate rise (66% to 81%) is correlational context, not a proven causal effect; and the reminder-experiment figures come from an in-house experiment reported without sample sizes or confidence intervals.
  • This is explicitly not an AI, machine-learning, or generative system: no risk scoring, no automated eligibility decision, no language model. Its atlas relevance is the navigation-node / single-point-of-dependency topology and the measured completion-dynamics levers, and any reading that implies algorithmic decision-making misreads it. A safe starting baseline is a property of this model, not a safety promise for any real deployment.
  • The Benefits Data Trust closure is an analogical counterfactual about a different organization with a different funding structure, not evidence about GetCalFresh; GetCalFresh's wind-down was planned and state-coordinated - the opposite failure mode - and no post-sunset participation measurement has been published.

Sources and evidence

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

  • Between 2019 and 2025 more than 70% (about 73% per its ten-year retrospective) of California's online SNAP applications were submitted through GetCalFresh, a deterministic, structured-workflow application assister built and operated by the nonprofit Code for America, which reports helping 6.2 million people obtain more than $12.8 billion in food benefits from 2017 to 2025 (organization-published figures that are not independently audited); the node made no eligibility determinations, and in 2024 and 2025 the California Department of Social Services coordinated a dated, phased transfer of its functions into the state-owned BenefitsCal portal.

    empirical
    • Vendor Code for America, Reflecting on 10 Years of Food Assistance in California (2024) https://codeforamerica.org/news/reflecting-on-10-years-of-getcalfresh/
    • Vendor Code for America, Food benefits (program page, 2025) https://codeforamerica.org/programs/social-safety-net/food-benefits/
    • Government California Department of Social Services, GetCalFresh Transition to BenefitsCal (2025) https://www.cdss.ca.gov/inforesources/cdss-programs/calfresh-outreach/getcalfresh-transition

  • A randomized controlled trial of roughly 65,000 Los Angeles GetCalFresh applicants (Giannella, Homonoff, Rino, and Somerville, American Economic Journal: Economic Policy 16(4), 2024) found that access to applicant-initiated flexible interviews increased SNAP approvals by about 6 percentage points, doubled early approvals, and raised long-term participation by over 2 percentage points, identifying the intake interview as a key procedural-denial barrier; Code for America separately reported an in-house experiment lifting renewal-form submissions among prior non-responders from about 1.5% to roughly 12% (organization-published, without sample sizes or confidence intervals).

    empirical
    • Academic Giannella, Homonoff, Rino, Somerville, Administrative Burden and Procedural Denials: Experimental Evidence from SNAP (American Economic Journal: Economic Policy 16(4), 2024; NBER Working Paper 31239, 2023) https://www.nber.org/papers/w31239
    • Vendor Code for America, Think Big, Start Small: How Implementing Flexible Interviews Improves Benefit Delivery (2021) https://codeforamerica.org/news/think-big-start-small-how-implementing-flexible-interviews-improves-benefit-delivery/
    • Vendor Code for America, How Experimentation Helps Us Meet Our Clients' Needs (2024) https://codeforamerica.org/news/how-experimentation-helps-us-meet-our-clients-needs/

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