Corrected October 26, 2026. Earlier versions of this page overstated the rankers: their inputs let them partly see errors (the issued benefit beside values recorded by QC reviewers). Every figure now comes from the leak-free rebuild, and the models no longer use the issued benefit. Later the same day the models were retrained on the official error definition: they rank by the dollars of errors the QC reviewer found with an amount over the federal tolerance, the ones the payment error rate counts, and the cost scenarios use the lift on those dollars.
—
Jurisdictions with a ranker
—
Median top-20% lift over random review
—
Jurisdictions at 2.0× or better
—
FY2028 cost share avoided, all states (central scenario, fairness floor on)
—
Where the ranker helps most
Click a state for its numbers. Lift and dollars come from each state's held-out QC cases; error rates are USDA's official FY2025 figures.
All jurisdictions
Sort by any column; search by name or code. Click a row to show it on the map. ⚑ = fewer than 1,500 QC cases of its own.
How the rankers were built
One model per jurisdiction from FY2020–2024 federal QC public-use data, using case facts (income, deductions, rent and utilities, household size) and the household's maximum allotment, but not the benefit issued. In the federal QC data the income and deduction fields are what the QC reviewer recorded, so pairing them with the issued benefit would let a model see the error.
It estimates each case's chance of an error over the federal tolerance (the kind the official payment error rate counts) times the likely error dollars, then ranks cases by that.
"Lift" is error dollars (errors over the tolerance) found in the top 20% (or 10%) of the ranked list, divided by what reviewing the same number of random cases finds, on cases held out from training.
"2020–22 → 2023–24" trains on the earlier years and tests on the later ones. "Forecast ÷ actual" compares the model's expected dollars with the dollars found in its top 20%.
Review time is assumed to be 30 minutes a case.
Each jurisdiction's model is trained on its own cases pooled with every other state's, plus the case's maximum allotment, which tells it the state's benefit levels. Pooling beat state-only models when it was adopted (that test used the earlier inputs).
Ranking purely by dollars under-reviews elderly households for their risk in every jurisdiction. The "fairness floor" splits the review budget across household types by their share of errors; it brings review into proportion at a cost of about a tenth of the lift in the median state.
Cost-share scenarios
From FY2028, states pay part of SNAP benefits if their error rate is 6% or more: 5% of benefits at 6–8%, 10% at 8–10%, 15% at 10% or more. The scenarios estimate how far ranker-guided reviews of active cases could lower the rate. They're the size of the opportunity, not a forecast. They use the current models' lifts with the fairness floor on (the recommended default).
Scenario
Cases reviewed a month
Lift used
Errors fixed
Months an error lasts
Read with care
Small samples make small jurisdictions noisy. Those with fewer than 1,500 QC cases are flagged.
The scenarios don't model whether corrections land before the federal sample month, procedural errors outside the public file, or the option to delay the cost share.
The ranker orders a state's own reviews. It doesn't touch the federal QC sample or decide anything about a household.