Sparse-57k certified release drops SSI asset input columns; SSI resource-test reforms score $0 on bundle 4.18.8

[MaxGhenis]

Summary

The sparse-57k certified release (populace-us-2024-sparse-l0-refit-57k-71a0887-national-only-20260701, the bundle 4.18.8 default) drops the SSI countable-resource input columns from the person table entirely: bank_account_assets, stock_assets, and bond_assets are absent (222 person columns vs 249 in the dense f0af251 release certified for 4.18.7, which carries all three with real mass — 42.5% of people nonzero).

This is the dropped-column counterpart of #278's zeroed input bases: ssi_countable_resources (adds = gov.ssa.ssi.eligibility.resources.countable) computes 0 for every record, meets_ssi_resource_test is universally true, and any SSI resource-limit reform silently scores $0 for every 4.18.8 consumer.

Measured (TY2026, managed_microsimulation, policyengine-us 1.752.2)

run	baseline SSI	remove resource test	$10k/$20k limits
sparse 57k (4.18.8 certified)	8.27M recipients / $66.0B, no resource-test bite	+$0.00B / +0 recipients	+$0.00B / +0
dense f0af251 (4.18.7 certified) native asset columns	8.05M / $63.6B	+$16.05B / +3.23M	+$1.60B / +499k

SSA administrative totals for reference: ≈7.4M recipients / ≈$61B federal payments (2024). The dense baseline sits near them; the sparse baseline runs above (no resource-test exclusions) and its reform surface for this policy family is dead.

External comparators for the $10k/$20k scenario (both modeled for 2026): SSA OACT scored S.4102 at ≈$1.0B federal SSI cost in 2026; CBPP estimated ≈224k new beneficiaries. The dense-native run lands within ~1.6× / ~2.2× of those; prior ECPS-based runs (PolicyEngine/ssi-asset-limits) were similar or further.

Mechanism

The three asset inputs sit on no calibration target, so the sparse build path lost them altogether (here as dropped columns rather than zeroed mass). The #257-style gate (flag inputs absent, constant-at-default, or zero-mass vs the dense parent / an admin anchor) would have caught this row too — bank_account_assets et al. would diff as column missing against the dense parent.

Suggested fix

• Carry the three asset columns through the sparse export (they exist in the dense parent; the 57k household subset just needs to keep them), or re-impute at build time via the us-data SIPP QRF stage.
• Add them to the Gate releases on degenerate PE input columns (constant at the engine default) #257/Sparse-57k certified release zeroes untargeted input bases (IRA/HSA/SE-pension/childcare); CDCC and ALD reforms silently score ~$0 on bundle 4.18.8 #278 release-gate input inventory so SSI/TANF/Medicaid resource-test reform classes can't silently zero out again.
• Classify them explicitly in the Verify US wealth, property, and asset imputation backlog #49 wealth inventory (populated / upstream-populated / reviewed-excluded).

Found while re-running the PolicyEngine/ssi-asset-limits analysis (SSI resource-test reforms) against the certified populace default via policyengine.py. Relates to #278, #49, #257; the next release build (#324) is a natural vehicle.

[MaxGhenis]

Three follow-ups from re-running the SSI analysis against the certified small default on bundle 4.20.0 (same sparse-57k artifact and pe-us 1.752.2 as 4.18.8):

1. The regression is now pipeline-level, not just the sparse export. The Build-H dense parent published today (populace-us-2024-buildh-dense-warmstart-b449eb7-20260708T100442Z) also lacks bank_account_assets/stock_assets/bond_assets (and ~129 of the same 133 dropped columns — full inventory in #361). The Build-I successor (#359) inherits this unless the export carries the columns. "Carry through the sparse subset" alone no longer fixes it; the source stage has to run (or persist) in the Build-H+ lineage.

2. Consumer-side repair is not viable. Two attempts to restore the inputs downstream, both on the certified artifact:

• Carry-by-join from the dense parent: only 45.2% of sparse persons match dense f0af251 on PERIDNUM (different pools, cloning) — a partial carry leaves half the file asset-less.
• Post-hoc QRF overlays (the two models in policyengine-us-data storage, mean_quantile=0.5, assets set at the dataset year): both collapse the baseline and swing the reform score by an order of magnitude.

TY2026, pe-us 1.752.2	baseline SSI recipients	remove resource test	$10k/$20k limits
certified sparse-57k, as shipped	8.27M (no test bite)	$0	$0
sparse + 5-predictor liquid_assets.pkl overlay	4.02M	+$33.2B	+$26.2B
sparse + 10-predictor WIP-spec overlay	3.05M	+$39.6B	+$9.0B
dense f0af251 native columns (4.18.7 default)	8.05M	+$16.1B	+$1.60B
external anchors	SSA ≈7.4M	prior ECPS +$15.75B	OACT ≈$1.0B / CBPP 224k new

The overlays put far more $2k+ asset mass on SSI-income-eligible people than the in-build imputation did (dense: 57.5% of people at exactly zero; overlays: 32–48%), knocking out half the baseline caseload and making the $10k/$20k delta range from $1.6B to $26.2B across vintages. Only the dense-native columns sit near the external anchors.

3. Practical implication for the fix. Scoring this reform class on the small default needs the asset inputs restored in-build (source-stage imputation persisted through selection/export, presence-gated per #361) — not consumer-side patches. Until then, 4.18.7 dense-native is the only populace artifact that can score SSI resource-test reforms.

Repro scripts: ssi-asset-limits repo, .local-results/populace-rerun-scripts/ (variants: certified+overlay via SSI_ASSETS_CSV, dense-native via SSI_DATASET_URI + SSI_NATIVE_ASSETS=1).

[MaxGhenis]

Build I status (populace-us-2024-buildi-sparse-rmloss100-6e8e929, certified as the US default today via policyengine.py#463): unchanged for this issue. The artifact restores 7 input columns (IRA/SE-pension contributions, hours, citizenship categoricals — detail on #361) but none of the asset columns; measured on the new default with pe-us 1.764.6, ssi_countable_resources = 0 for every record and the $10k/$20k reform scores $0.000B.

Reference check: the dense f0af251 native-asset runs re-scored on pe-us 1.764.6 are identical to the 1.752.2 results (remove test +$16.05B / +3.23M; $10k/$20k +$1.60B / +499k) — the model bump doesn't move SSI mechanics, so those remain the current-model reference for this reform class.

Calibration interaction worth noting for the fix: Build I's baseline SSI lands at 7.97M recipients / $61.1B — on top of SSA admin totals with no resource test binding. The weight solve is absorbing the missing test, so restoring the asset columns without refitting the SSI targets will push baseline SSI below target; the restore and the SSI-target refit should ship together.

[MaxGhenis]

Resolved on the certified default: Build M (populace-us-2024-buildm-sparse-rmloss100-98bf731-20260717T075101Z, certified via policyengine.py#475) ships the SSI asset columns (scf_wealth stage on) and the $10k/$20k resource-limit smoke probe scored +$5,350,231,200 on the certified artifact — the exact reform class this issue reported as $0 on bundle 4.18.8. Residual asset-magnitude refinement (SIPP blend + SSI-aware refit) continues as #374 in the Build N bundle (#449).