Diagnose the child-vs-total SPM composition anomaly (poverty stays a permanent holdout)

From scorecard diagnosis batch 1, item #8 (classification: pe_gap, confidence high; full memo: PolicyEngine/policyengine-scorecard diagnosis/DIAGNOSES.md).

Divergence

PE exceeds Urban by 3.1689 percentage points (22.96% relative). More decisively, PE exceeds the same-year Census 2024 child SPM benchmark by 3.5689 points (26.63%), while its total rate exceeds Census by only 0.7042 points. Urban's 2023 13.8% is only 0.4 point above the separate Census 2024 child benchmark despite the year mismatch. The age-specific pattern and same-year benchmark reject a general 2023-to-2024 shift as the main explanation. EHS contributes exactly zero to the scorecard SPM calculation.

Proposed fix

Build P's 2024 child SPM poverty rate is 16.97% versus Census P60-287's same-year 13.4%, even though total poverty is much closer (13.60% versus 12.9%). Promote the existing national total and child P60-287 rows from passive backtests to explicit release diagnostics with thresholds. Add child-versus-adult decompositions of market income, cash and in-kind benefits, taxes, medical expenses, work/child-care expenses, and SPM thresholds. Poverty rates are PERMANENT calibration holdouts (doctrine 2026-08-02): use these comparisons to locate the child-specific resource or threshold error; fitting poverty directly is prohibited, and a deviation from the official rate is not itself the defect — the child-vs-total asymmetry is what warrants investigation.

Evidence

• c80b6c4:data/comparison.json, child SPM poverty_rate: held_out with no Build P poverty targets; PE 16.9689% versus Urban 13.8%.
• POP@8828dee:packages/populace-build/src/populace/build/us/state_spm_poverty_levels.json:1-32 records Census P60-287 single-year 2024 benchmarks of 13.4% for children under 18 and 12.9% overall; these are out-of-sample.
• data/pe/pe_metrics.json: PE 2024 child poor count 13,012,226.578 / population 76,683,005.281 = 16.9689%; total is 46,264,651.701 / 340,077,321.941 = 13.6042%.
• PEUS@92e6052d3e spm_unit_is_in_spm_poverty.py:10-13 compares spm_unit_net_income with spm_unit_spm_threshold; in_poverty.py:11-12 and poverty_gap.py:12-15 implement the same strict comparison used by the Census backtest; spm_unit_net_income.py:11-18 adds market income and benefits and subtracts taxes and SPM expenses.
• docs/replication-assessment.md:19,130-133 documents PE annual 2024 versus ATTIS average-month 2023 on pooled 2022+2023 ACS.
• PEUS@92e6052d3e:policyengine_us/variables/household/income/spm_unit/spm_unit_benefits.py:11-84 omits Head Start and Early Head Start, ruling out the $112.3B EHS valuation defect as a cause of this SPM rate.

Adjudication note

Adjudicator amendment (Max, 2026-08-02): (1) poverty metrics are PERMANENT calibration holdouts — never targets, not 'held out initially'; targets come from administrative systems only. (2) Deviations from official poverty statistics are never inherently problematic — populace corrects survey benefit underreporting (admin-calibrated benefit dollars exceed survey-reported amounts), so sitting below survey-based rates is the expected direction, all else equal. What flags THIS item is the anomaly: the child rate sits ABOVE Census while the total sits near it. Investigate composition; the deviation itself is not the defect and the official number is a comparator, not truth.

🤖 Generated with Claude Code

Start with data/comparison.json and the referenced SPM variables, especially spm_unit_net_income.py, spm_unit_is_in_spm_poverty.py, in_poverty.py, and poverty_gap.py. Trace the child-versus-total resource and threshold components, using the Census benchmarks in state_spm_poverty_levels.json as comparison data. Done means explicit release diagnostics and decompositions identify the child-specific asymmetry without fitting poverty metrics as targets.