PolicyBench · expansion

Expand PolicyBench to your region, population, or program

Families already ask AI about the questions that decide their month: Do I qualify for SNAP? How much is my credit? Will this job cost me Medicaid? The public board tests 39 frontier models on 100 real households. The best model, GPT-5.6 Sol, computes 89.2% of requested outputs exactly, household-impact weighted (amounts within $1, eligibility flags exact). On Medicaid eligibility, the median model misclassifies about 1 in 16 people; the weakest, about 1 in 3 people. A family told “$0” does not apply. Those are national numbers; nobody measures this for your region.

A PolicyBench slice measures it. We draw households from survey microdata weighted to your population, cover your programs, benchmark the models behind the tools your people use, and compute every reference from the law with PolicyEngine. Rows selected for audit receive a diagnosed failure mode. You get a public slice leaderboard, a written analysis of where models fail your population, and a briefing. The answer key checks itself in public: references come from open-source code, cross-checked against other calculators where they exist, and challenged values get adjudicated against the statute — the methodology and adjudication record are published.

Program deep-dive

from $7,500

One program family: SNAP, Medicaid, child care, or tax credits, across all 39 board models. Per-model accuracy, diagnosed failure modes, written analysis, and a briefing. Fast: the board already holds the raw material.

State or city slice

from $20,000

New households weighted to your area and program mix. A published slice leaderboard beside the national board, the full audit, the analysis, and a briefing for your team or grantees.

National or portfolio

let’s talk

A 50-state observatory, or coverage across a whole grantee portfolio — including your grantees’ own tools, run through the same households as the board. We scope these directly.

We size samples to the confidence width your question needs, from your population — not a fixed household count. Any slice can add standing coverage: quarterly refresh and re-analysis, new models as they ship, and technical assistance to grantees building AI tools — priced with the slice.

What funding buys — and what it never buys

Funders buy coverage: households, programs, regions, refresh cadence. Funding never buys scores, rankings, or placement. No model vendor pays for evaluation — vendor funding for separate projects never touches the benchmark. Every slice stays public — prompts, references, predictions, and diagnoses.

Contact uscontact@policybench.org

PolicyBench is a project of PolicyEngine, a 501(c)(3) nonprofit — engagements work as grants or contracts. We also brief funder networks — one slice presented to a convened room goes further than a dozen pitches, and we are glad to present at yours.