Benchmark build: carry the disability inputs the references depend on (SSI criterion, SSDI months, permanent-and-total disability)

Summary

PolicyBench draws its households from the certified US build and scores frontier models against PolicyEngine references computed from exactly the inputs the build carries. On 2026-09-05 the 39-model board had to exclude eleven outputs from scoring because their references depend on inputs the June 2026 build (populace-us-2024-5da5a95-20260611) never carried, so the prompt could not list them (PolicyEngine/policybench#164, reference_exclusions.json). The v2 household draw needs a build that carries them.

What the June build lacks

Builds from 2026-07-17 carry the SIPP-imputed SSI criterion (7,275 to 7,515 persons). #361 documents the sparse export dropping 94 no-formula inputs, including is_disabled, is_blind, employer_sponsored_insurance_premiums and the asset columns; the benchmark build must carry every no-formula input the engine reads.

Asks

1. Certify a v2 candidate build that carries every no-formula PolicyEngine-US input (the #361 fix), the SIPP SSI criterion, and is_incapable_of_self_care.
2. Populate is_permanently_and_totally_disabled (impute from SSI/SSDI receipt or the SIPP donor) or document that the tax code's disability pathways are out of scope.
3. Carry SSDI entitlement duration or start date (months_receiving_social_security_disability) with a defined reference date, if the SIPP donor can supply it; otherwise record the assumption explicitly so the benchmark can state it.
4. Record provenance per disability fact (observed, imputed, unknown). An unknown must not export as an observed negative.
5. Keep person_marital_unit_id in the export; the benchmark now builds marital units itself (head and spouse) after finding policyengine-core lumps a household into one unit when none is given.

Related: #424, #453 (SSI candidate undercount, child SSI absent), PolicyEngine/policyengine-us#9330 (disability input wiring audit), PolicyEngine/policybench#139 (v2 board).

Start by reviewing #361 and the June and July 2026 build differences, especially the sparse export and SIPP-imputed disability inputs. Trace how the benchmark build carries PolicyEngine-US inputs and provenance, then verify that a v2 candidate includes the requested disability fields and person_marital_unit_id. Done means the candidate is certified or its explicit assumptions and out-of-scope pathways are documented.