Early Head Start ships at engine-default 100% take-up: $112.3B (≈30× ACF spending) inside household_net_income on the certified 2024 artifact; CHIP at 100% is program-aggregate-only

Observed on the certified artifact

Certified release populace-us-2024-buildp-sparse-rmloss100-cae8640-20260728T011454Z, loaded via policyengine.py 5.0.1 managed_microsimulation (engine policyengine-us 1.764.6, period 2024), weighted sums (2026-08-01 take-up mechanics audit):

The artifact stores nine takes_up_* columns (aca, eitc, head_start, housing_assistance, medicaid, medicare, snap, ssi, tanf). takes_up_early_head_start_if_eligible and takes_up_chip_if_eligible are not among them, so the engine's default_value = True applies to every record — the "mechanical 100% participation" the take-up contract's own doctrine text predicts for unseeded data_seeded leaves.

Administrative anchors

• The engine's own gov.hhs.head_start.early_head_start parameters (2023-09 vintage, summed across states): $3.796B spending / 180,873 enrollment (≈$21.0k per enrollee). The artifact distributes ≈29.6× the engine's own spending parameter.
• ACF/OHS PIR service snapshot 2023–24 (the same PDF pinned in test_us_early_head_start_exclusion.py): 187,966 funded slots; 232,842 cumulative enrollment (219,916 children + 12,926 pregnant women). The artifact's 5.16M weighted participants are ≈22–27× either count.
• CHIP: the engine's separate-CHIP calibration parameters (2024) sum to 3,341,711 enrollment / $8.394B spending; the artifact computes 4,322,070 / $12.977B (≈1.3× / 1.5×). Much milder, and the fix path is already adjudicated (below).

Mechanism (each step read in code at policyengine-us 1.764.6, commit 92e6052d3e)

• takes_up_early_head_start_if_eligible and takes_up_chip_if_eligible are bare inputs, default_value = True, no formula.
• early_head_start pays state spending / enrollment (≈$21.8k average across the artifact's takers) to every taker; is_early_head_start_eligible = (age < 3 or pregnant) and (Head Start income-eligible or categorically eligible) → 5.16M weighted persons.
• chip pays per_capita_chip (separate-CHIP state spending / enrollment) gated by chip_enrolled = is_chip_eligible & takes_up_chip_if_eligible.

This is not a build-process failure: the take-up contract deliberately records both flags as rate_unsourced ("left unseeded", #313 parity-gate exemption), and the EHS evidence contract (test_us_early_head_start_exclusion.py, ecps_parity_known_gaps.json) proves the source unavailability. The problem is the composition: adjudicated non-seeding + engine default True + downstream aggregation = a ~30× phantom on the certified surface.

Downstream exposure, verified at 1.764.6

• household_net_income is contaminated. early_head_start and head_start are in the gov.household.household_benefits parameter list (2022 and 2024 brackets) → household_benefits → household_net_income.adds. ≈$112B of phantom benefits, concentrated at ≈$21.8k/person on households with an under-3 child or pregnant member inside the Head Start eligibility domain. Anything computed from household_benefits or household_net_income (benefit aggregates, net-income levels and decile outputs) inherits it.
• SPM poverty is clean. Neither program (nor chip/medicaid) appears in spm_unit_benefits, and in_poverty ← poverty_gap ← spm_unit_net_income ← spm_unit_benefits. Poverty baselines are unaffected.
• CHIP is program-aggregate-only by default. chip sits in household_health_benefits, which returns 0 unless gov.simulation.include_health_benefits_in_net_income (default false). Exposure is the chip aggregate itself plus the health-inclusive net-income variants.
• Weights are unaffected. The fiscal-targets registry (fiscal_targets.py) and target-parity manifest contain no head_start reference and no household_benefits expression, so EHS never enters calibration; the distortion is purely output-side. (CHIP enrollment targets do exist and carry the concept mismatch tracked in #321.)

Why the obvious data-side fixes are already foreclosed

• A sourced admin rate: the contract's provenance rule requires an administrative participation rate; ACF publishes none. PIR cumulative enrollment (232,842) includes turnover; funded enrollment (187,966) is capacity. Both are pinned as "semantic non-substitutes" in test_us_early_head_start_exclusion.py, and the retired us-data 9% NIEER scalar is the fabricated-rate class ledger#77 removed. PIR enrollment counts therefore do not clear the bar for a rate-seeded stage.
• A count_calibrated stage: that treatment (medicaid pattern, #331) requires a survey-reported anchor column, and the evidence contract proves none exists for EHS in any locked input (SIPP EEDHEADST has zero under-3 positives, the child-care items begin at age 2, there is no pregnancy branch; ASEC/PUF carry no enrollment signal).
• CHIP: already adjudicated — blocked on the #321 M-CHIP/separate-CHIP concept split, then follows the #331 pattern. Its numbers above are for the record, not a new ask.

Options

1. Engine-side gate (recommended). policyengine-us moves head_start/early_head_start out of the always-on household_benefits list or behind an include-switch — the exact precedent used for medicaid/msp/chip/aca (household_health_benefits behind gov.simulation.include_health_benefits_in_net_income, default false). Fixes any dataset that does not store the flag, and preserves per-program calculator display. Companion issue: PolicyEngine/policyengine-us#9192. Note this only cleans the headline surface — early_head_start itself remains a 100%-take-up aggregate until a source exists, so aggregate consumers still must not sum it.
2. Engine default flip for these flags: conflicts with the household-calculator contract (default True means "show what an enrollee would get"); not recommended.
3. New populace treatment class ("unrepresentable program": store all-False with the same evidence): converts the +$108B error into −$3.8B, but is a doctrine change — rate_unsourced today deliberately means "leave the engine default", and all-False is as unsourced as any other rate. Needs explicit adjudication if chosen.
4. Status quo + disclosure: keep the parity-gap exemption and document the certified-surface contamination in release notes so consumers don't sum early_head_start/chip or read household_net_income levels as calibrated.

Adjacent, tracked elsewhere: WIC at 1.764.6 runs its own machinery (52.8% of eligibles on this artifact; no takes-up flag consumed) and the 1.779.4 takes_up_wic_if_eligible contract drift belongs to the #324 pin-bump lane. takes_up_basic_health_program_if_eligible and takes_up_dc_ptc are the same rate_unsourced class (no numbers recorded here). #591 tracks the separate default-value-hole class (is_tanf_enrolled etc.).

Repro: weighted sums of early_head_start, chip, head_start and their eligibility variables on the certified artifact via managed_microsimulation (period 2024); engine wiring verified with git show 92e6052d3e:<path> against policyengine-us (household_benefits.yaml, spm_unit_benefits.py, household_net_income.py, household_health_benefits.{py,yaml}, include_health_benefits_in_net_income.yaml, the early_head_start/chip variable files).

Reproduce the weighted sums with managed_microsimulation for the certified 2024 artifact, then read the named policyengine-us files, especially household_benefits.yaml, household_health_benefits.py, and household_net_income.py. Compare the proposed engine-side gate with the documented alternatives and companion issue #9192. Done requires an adjudicated treatment that prevents the contaminated household_net_income surface while preserving the stated contract and documenting remaining aggregate limitations.