SNAP payment error simulator

Beginning fiscal year 2028, a state's SNAP payment error rate sets its share of benefit costs — 0% below a 6% rate, 5% from 6–8, 10% from 8–10, 15% at or above 10 — keyed to the state's choice of its FY 2025 rate (published June 2026) or its FY 2026 rate, still being measured through September 30. One more twist sits at the top: a rate whose 1.5× reaches 20% (above 13.33%) delays the state's first billed year — FY 2025 crossing pushes the start to FY 2029, FY 2026 crossing to FY 2030 — so the highest-error states owe $0 in FY 2028. This simulator draws from the state's own USDA Quality Control sample to show the distribution of a measured rate, the odds its FY 2026 election beats its locked FY 2025 rate, the delay odds, and what targeted review of high-risk cases, audit volume, and a model-based standard medical deduction scenario do to the bill.

Beyond the state's FY 2024 sample size

Targeted review scenarios

Cut counted error dollars by a chosen amount inside a targeted group of cases, and see what the measured rate, the tier odds, and the expected bill do. An accounting construction: it prices the arithmetic of a smaller error pool, not any claim about how a state achieves it.

Oracle ranks by each case's actual recorded error — the ceiling any targeting method can reach. Random is the floor. The model row uses the fitted error model's out-of-sample FY2024 scores.
Of weighted caseload; one boundary case can be fractional.
If the cut persists, FY 2028–30
Sustained-intervention construction: the same rate delta applied to each horizon year of the persistence-anchored path.
Policy scenario — case-level model
Re-predicts every case from the fitted error model with the state's SMD documentation feature reversed. Loads the model on first use.
Rules-engine scenario — cause-coded
Scales each case's error dollars by its share of findings the reviewers coded as computer programming (17), computer-generated mass change (19), or arithmetic computation (20) — the class a correct rules engine removes. Data entry (18) and computer user error (21) stay. Steady-state FY 2024 carries little of it: 2.2% of counted error dollars nationally under this attribution. System transitions are where the class spikes — see the migrations paper in the events section.
Benefit recomputation
Shows how the open, certified Axiom RuleSpec computation compares against the QC file's FNA Minimodel benefit chain, stage by stage, with the formula-benefit divergence catalog. Precomputed artifacts; the simulation is unchanged.
Uncertainty
Adds process movement on top of sampling noise, calibrated to the single realized FY 2024→2025 transition (robust estimate). Off = sampling noise only, a disclosed lower bound.

Sampling engine: resamples the state's observed FY 2024 QC error dollars, centered on the official rate.

–
Official FY 2025 payment error rate (locked election year)
–
Expected FY 2028 bill, scenario (elected rate, delay-aware)
–
Most likely tier, scenario
–
Std. dev. of the FY 2028 bill

The FY 2027 measurement year

The FY 2025 vs FY 2026 election

The FY 2028 cost share keys to the state's FY 2025 or FY 2026 rate, at its election (7 U.S.C. 2013(a)(2)(B)(ii)). FY 2025 is published and locked; FY 2026 is being measured through September 30, 2026. The simulated FY 2026 below is a QC-sized sample centered at the FY 2025 level — sampling noise only, a lower bound on true FY 2026 uncertainty.

Adopting a verified rules engine

System replacements in the record

The adoption panel prices the computing-apparatus cause class by accounting convention. Three state eligibility-system replacements test how that class moves in practice — each estimated against a synthetic control built from never-treated states, with the decision rule frozen before results.

Rhode Island (UHIP, 2016): strict computing-apparatus error dollars rise $2.90 per weighted case-month against the synthetic control (permutation p = 0.023; the client-caused placebo stays null). The rise concentrates in FY 2017–19, the window for which FNS billed the state $37.3M — a descriptive consistency check the frozen protocol designates as verdict-inert.

Kentucky (Benefind, 2016): no protocol-defined signal (−$0.58, p = 0.30), despite a documented troubled launch.

Oregon (ONE expansion, 2021): refused by the placebo rule — the client-caused placebo fires inside the pandemic window, so attribution is impossible.

The pooled two-state statistic reaches p = 0.093. These estimate bundled system replacements as implemented — staffing, process, and software together — never “the effect of a rules engine.” Frozen protocols and full artifacts: the design memo, Rhode Island and Kentucky results, Oregon results, the migrations paper (the full design, both decomposition estimators, and the limits), and the companion measurement paper.

Measured-rate distribution

Scenario draws (filled) vs. baseline (outline); bands mark the cost-share tiers; the dashed line is the official FY 2025 rate.

Cost-share tier probabilities

All states at a glance

Baseline simulation for every jurisdiction (observed resample) — anchored at the official FY 2025 rate, FY 2024 sample size and composition, no levers. Click a column to sort, a row to load that state above.

Method and caveats
  • Error process from the FY 2024 SNAP QC public-use file, official error-rate universe: adjudicated errors with recorded amounts above the official $56 threshold. Scenarios re-draw QC-style samples and recompute the weighted rate, centered on the state's official FY 2025 published rate (the FY 2024 file is the newest public case file; case composition is held at FY 2024 while the level anchors at FY 2025).
  • The official rate exceeds the file-computable rate by a layer the file never records case-by-case — federal re-review integration and ineligible-case error. Anchoring carries that layer as a fixed offset: across jurisdictions it is a median 31% of the FY 2024 official rate (from −3% to 81%), it contributes no sampling variance to the draws, and no scenario reaches it. The state panel above reports each state's own file-covered share, and a committed registry decomposes the wedge by the published overpayment and underpayment components — nationally 95% of it sits on the overpayment side, consistent with the excluded ineligible-case error (overpayment by construction) dominating the layer.
  • The simulated distribution carries sampling noise only, and the FY 2024 to FY 2025 transition shows that is a lower bound: 18 of 53 jurisdictions changed cost-share tiers between the two published years while only 10 moved beyond the two-year sampling-noise 95% band (median absolute z 1.38 against the simulated SD; full table in the repository). Election probabilities here are therefore conservative about year-over-year process movement.
  • FY 2025 rates are from the Food and Nutrition Administration's June 24, 2026 release (FNS was renamed FNA effective June 2026); FY 2024 rates and the QC case file are from the predecessor FNS publications.
  • The policy scenario re-predicts every case from the fitted error model with the state's standard-medical-deduction documentation feature reversed — per-case deviation probabilities and all nine deviation quantiles, exported with 10,000-draw paired case-bootstrap 95% intervals (model frozen through FY 2022, evaluated on FY 2024). It is a model-implied association from burden features that add +0.006 ROC AUC, not a causal policy effect; states adopt options endogenously. Self-employment, heat-and-eat, and BBCE scenarios are deliberately absent: the fitted features cannot express a defensible policy flip for them (machine-readable reasons ship in the export).
  • The underlying error process is held fixed: no behavioral response, and no corrective-feedback channel from auditing more cases.
  • Case bootstrap approximates the stratified monthly QC design; within-state estimates carry wide uncertainty. Comparative statics are more robust than dollar levels.
  • The all-states table runs the same baseline simulation (4,000 draws per jurisdiction, official rate and sample size, no levers); "P(different tier)" is the chance a fresh QC-style sample lands in a different cost-share tier than the official point rate implies. Expected cost applies the FY 2028 tier shares to FY 2024 issuance — an illustration of stakes, not a forecast of FY 2028 budgets.
  • When the scenario is on, both baseline and scenario draw each case's payment deviation from distributions fitted on FY 2017–19 and 2022 QC data (pandemic years excluded), with levels anchored to the official rate; with it off, the simulator resamples observed QC error dollars. Model output is disabled for seven jurisdictions (AK, HI, ID, MN, SD, VI, WY) whose factor-adjusted FY 2024 model-to-observed dollar-rate ratio falls outside [0.7, 1.4]. Validation results — including quantile under-coverage at most levels and under-dispersion of simulated rates — and the adversarial review history are in the repository.
  • The delay clause is modeled: if a state's FY 2025 (or FY 2026) payment error rate times 1.5 reaches 20%, its start moves to FY 2029 (or FY 2030) — 7 U.S.C. 2013(a)(2)(B)(ii)–(iii). Expected-bill figures zero delayed years; the FY 2029 bill keys to the FY 2026 rate (the third preceding year). A start pushed to FY 2030 keys to the FY 2027 rate, which this simulator does not model — those draws contribute $0 to the shown FY 2028 and FY 2029 expectations.
  • The Axiom rules engine verification view displays precomputed artifacts only: stage-by-stage parity from the committed axiom-oracles comparison reports (byte-identical, SHA-256-pinned copies in paper/snapshot/oracle-suites) and the formula-benefit divergence catalog from analysis/engine_comparison.py. The browser verifies the artifact's SHA-256 against a committed pin before displaying it. Engine runs never execute in the browser, and this view does not alter the simulation.
  • Source code, tests, and data pipeline: PolicyEngine/snap-qc-sim. Data: SNAP QC public-use files and FNS's published payment error rates. Cost-share tiers: 7 U.S.C. 2013(a)(2).