US: humanitarian immigration statuses absent from the imputation — H.R.1 §71109/§71302 eligibility channels compute ≈$0

[MaxGhenis]

Evidence

On the certified US release (populace-us-2024 Build P, read through policyengine.py 5.0.1), immigration_status carries variation but no records in the categories H.R.1 removes from program eligibility:

status	records
CITIZEN	153,491
UNDOCUMENTED	6,682
LEGAL_PERMANENT_RESIDENT	5,702
CUBAN_HAITIAN_ENTRANT	226
DACA	220
REFUGEE / ASYLEE / DEPORTATION_WITHHELD / CONDITIONAL_ENTRANT / PAROLED_ONE_YEAR / TPS	0

years_since_us_entry is populated (100% nonzero).

Why it matters

Three enacted H.R.1 eligibility channels key off exactly these categories and therefore compute ≈$0 on this file:

1. Medicaid — §71109 narrows qualified immigrants to LPR + Cuban/Haitian entrants + COFA, effective 2026-10-01. Encoded in policyengine-us (gov/hhs/medicaid/eligibility/eligible_immigration_statuses, 2026-10-01 value).
2. CHIP — same §71109, same date (KFF summary of the law). Not yet encoded in policyengine-us (filed separately), but the data gap binds regardless.
3. ACA PTC — §71302: below-FPL lawfully-present exception repealed from 2026; ineligible_immigration_statuses adds TPS/REFUGEE/ASYLEE/DEPORTATION_WITHHELD/PAROLED_ONE_YEAR from 2027-01-01. Encoded in policyengine-us.

CBO scores §71109 alone at −$6B federal Medicaid over 10 years and +100k uninsured by 2034. Any distributional analysis of the law on this file silently zeroes all three channels. Context: the IARIW OBBBA paper (PolicyEngine/obbba-paper) discloses this as a data boundary in its methods; the multi-year extension needs it fixed.

Ask

Carry humanitarian/temporary-protection statuses in the imputation so eligibility rules can see them. Design options, not mutually exclusive:

1. Impute the missing categories onto the CPS base — reassign a calibrated share of current LPR/citizen records (or model directly from ACS/SIPP citizenship + entry-year + origin) to REFUGEE / ASYLEE / DEPORTATION_WITHHELD / PAROLED_ONE_YEAR / TPS / CONDITIONAL_ENTRANT, targeted to external stocks (DHS/ORR admissions, KFF/ASPE coverage counts of refugees-asylees on Medicaid/CHIP).
2. A stored eligibility-check surface: alongside takes_up_*_if_eligible, a data-level flag family (e.g., passes_immigration_eligibility_check_{program}) that rules engines respect — so eligibility shocks flow to Medicaid, CHIP, and PTC coherently rather than being program-dollar overrides. This generalizes beyond immigration (verification/redetermination churn could use the same surface).
3. Interim (consumer-side, documented in the OBBBA repo): a seeded reform-level reassignment of statuses in the right proportions — acceptable for sensitivity work, not for certified estimates.

Acceptance criteria

• immigration_status distribution on the release includes the humanitarian categories with weighted totals validated against a stated external source.
• The three H.R.1 channels above produce nonzero, sign-correct effects when toggled on a certified release.
• Certification notes state the imputation source and target for each category.

[MaxGhenis]

Design proposal (implementing now; PR to follow)

Mechanism claims verified against policyengine-us 1.819.0 (the workspace lock) and the microcosm build code; pool numbers computed on the raw 2024 ASEC person file; every target below was read directly from the cited source today.

Why the current stage refuses these labels, and why that reason no longer holds

The immigration_status source stage deliberately does not emit REFUGEE/TPS/etc. — its docstring reasons that blanket labels would mislabel millions "while LPR treatment gives near-identical means-tested eligibility for those populations." H.R.1 repealed exactly that equivalence: after §71109, LEGAL_PERMANENT_RESIDENT stays Medicaid/CHIP-qualified while REFUGEE / ASYLEE / DEPORTATION_WITHHELD / PAROLED_ONE_YEAR / CONDITIONAL_ENTRANT fall out of gov/hhs/medicaid/eligibility/eligible_immigration_statuses (2026-10-01); §71301/71302 remove TPS / REFUGEE / ASYLEE / DEPORTATION_WITHHELD / PAROLED_ONE_YEAR from ACA lawful presence (2027-01-01) and repeal the below-FPL PTC exception (gov.aca.below_fpl_immigration_exception_in_effect flips false 2026-01-01). And SNAP §10108 already did the same thing (gov/usda/snap/eligibility/eligible_immigration_statuses narrows 2025-07-01 in the model) — the data gap silently zeroes a fourth enacted channel beyond the three in the issue body.

So the fix is not "loosen the old rationale" but "replace blanket labels with externally-targeted draws": assign each category only as much weighted mass as the published stock supports, from candidates whose survey signature (origin country × arrival window × legal-status indicators) is consistent with the status.

Chosen design: option 1, as an extension of the existing source stage

Raw-only compliant: the logic lives in the immigration_status stage kernel (us_runtime/immigration.py), parameterized by the source-stage manifest with one citation per number, running over raw ASEC columns the stage already requires (PENATVTY, PEINUSYR, PRCITSHP, the ASEC-UA indicators). No new raw columns, no new output columns, no post-hoc artifact edits; certified releases untouched — this lands in the next build lane.

Phase order inside the kernel (all draws reuse the existing seeded blake2b threshold-selection machinery, keyed on stable source identity so support-channel clones stay consistent):

1. Citizenship + ASEC-UA legal-status indicators (unchanged).
2. New: humanitarian draws, sequential and mutually exclusive, in decreasing order of target precision — PAROLED_ONE_YEAR (per-origin) → REFUGEE → ASYLEE → TPS (per-country) — each to a manifest-cited stock. Candidates exclude Cuba/Haiti-born (the CUBAN_HAITIAN_ENTRANT class is the better statutory label and keeps H.R.1 eligibility) and the DACA statutory cohort. Draws may take residual-pool people only for parole and TPS (mixed benefit signatures); REFUGEE/ASYLEE draw from the indicator-documented pool (those statuses carry immediate benefits access — 99%/83% of the 2020+ Afghan/Ukrainian documented pools report Medicaid).
3. EAD worker/student spill to the Pew controls (unchanged semantics — the controls bind the remaining residual pool, so the final undocumented worker/student counts still land on their published anchors).
4. Status tags (unchanged CHE/DACA/LPR/UNDOCUMENTED logic; humanitarian marks are disjoint from CHE/DACA by candidate construction). Residual-pool draws flip ssn_card_type NONE → NON_CITIZEN_VALID_EAD (parole and TPS both confer employment authorization), preserving the gate invariant NONE ⇔ UNDOCUMENTED — and fixing a pre-existing mislabel: TPS holders are ACA-lawfully-present but were being emitted as UNDOCUMENTED (ACA-ineligible) or LPR (Medicaid-qualified), each wrong in a different direction post-H.R.1.

Targets and sources (data year 2024)

category	target	source (read today)
PAROLED_ONE_YEAR — Afghanistan	73,566	DHS OAW report to Congress: "The number of Afghan evacuees who have been paroled into the United States: 73,566" (Jul 2021–Mar 2022)
PAROLED_ONE_YEAR — Ukraine	158,000	DHS "Uniting for Ukraine: Process Overview and Assessment": "over 158,000 individuals arrived … via the U4U process" as of 2023-09-30
PAROLED_ONE_YEAR — Nicaragua	93,070	CBP CHNV release (through end of Dec 2024): "93,070 Nicaraguans … arrived and were granted parole"
PAROLED_ONE_YEAR — Venezuela	117,330	same CBP release: "117,330 Venezuelans arrived and were granted parole"
REFUGEE	160,000	OHSS Refugees Annual Flow Report FY2023 (60,050 admitted, Table 3) + OHSS FY-24 Refugees Flow Report (100,060 admitted). Stock model: refugees must apply for LPR after 1 year, so the not-yet-adjusted stock ≈ trailing two years of admissions
ASYLEE	160,000	OHSS Asylees FY2023 flow report (35,720 granted in 2022 as revised, 54,350 in 2023) + OHSS FY-24 Asylees report ("nearly 70,000" in 2024). OHSS: 88% of adults granted affirmative asylum 2015–22 had adjusted to LPR by end-2024 — a trailing three-year window matches the adjustment lag
DEPORTATION_WITHHELD	0 (explicit, cited)	EOIR FY2024 decision outcomes: 1,386 INA + 770 CAT withholding grants in I-862 cases (+162 in I-863); CRS IN12501: <1% of 666,177 decisions. No published stock exists; the accumulated population (low tens of thousands) is below the CPS's resolution for a defensible draw. Emitting zero is explicit and gate-enforced until a citable stock appears
TPS — Venezuela	605,015	CRS RS20844 Table 1 (as of 2025-03-31): 2021 designation 252,825 + 2023 designation 352,190; required arrival ≤ 2023-07-31
TPS — El Salvador	170,125	same table; required arrival ≤ 2001-02-13 (PEINUSYR ≤ code 17)
TPS — Honduras	51,225	same table; arrival ≤ 1998-12-30 (code ≤ 16)
TPS — Nicaragua	2,910	same table; arrival ≤ 1998-12-30
TPS — Nepal	7,160	same table; arrival ≤ 2015-06-24 (code ≤ 24)
TPS — other designated	21,215	same table: Burma 3,670 + Cameroon 4,920 + Ethiopia 4,540 + Lebanon 140 + Somalia 705 + South Sudan 210 + Sudan 1,790 + Syria 3,860 + Yemen 1,380

Reconciliation against the table total (1,297,635 as of 2025-03-31; 1,095,115 as of 2024-09-30 per the Dec 2024 CRS update): the emitted TPS sum is 857,650 because Ukraine TPS (101,150) and Afghanistan TPS (8,105) are carried by the parole draw (those registrants are overwhelmingly the same U4U/OAW parolees, whose benefits flow through parole under the Ukraine supplemental and Afghan Allies acts), and Haiti TPS (330,735) is carried by the existing CUBAN_HAITIAN_ENTRANT label, which is the statutorily favorable class H.R.1 preserves — conservative for §71109 loss estimates. Origin PENATVTY codes were verified against Census CPS technical documentation Appendix I (Cuba 327 / Haiti 332 match the existing module constants); arrival bins against the 2024 ASEC data dictionary (PEINUSYR 27 = 2020–21, 28 = 2022–24; bin-boundary slop at the Venezuela 2023 cutoff is documented).

The parole targets are gross program admissions: a cohort member who later won asylum (large for Afghans: Afghanistan was 38% of FY2024 asylum grants) stays labeled PAROLED_ONE_YEAR. This is deliberately not corrected — both classes are qualified aliens pre-§71109 and lose Medicaid/SNAP/PTC on identical dates, so the split is channel-irrelevant in the current engine; noted in the stage docs.

Empirical feasibility (raw 2024 ASEC, weighted, verified by running the current stage): candidate pools carry 832k (parole origins 2020+, documented+residual: AFG 94k, UKR 118k, NIC 144k, VEN 476k), 567k (refugee origins 2020+, documented), 2.36M (asylee origins 2016+, documented), 3.9M (TPS origins). The Ukraine parole target (158k) exceeds its ASEC pool (~118k) — the 2022–24 inflow is undercovered by the survey — so the draw will saturate and the gate records achieved-vs-target per category (band-checked, coarse). Reconciling levels in weights is the calibration lane's job, same division of labor as the undocumented total today. Pew's 2025 report (2023 estimates) independently validates the extraction design: its 14.0M "unauthorized" total explicitly includes ~6.0M with temporary protections (parole ~700k, TPS ~650k, DACA ~600k, pending asylum), so pulling TPS/parolees out of the residual pool is consistent with the residual-method concept the stage already uses; the emergent undocumented total stays well inside the existing gate band around the Pew anchor.

Option 2 (the flag surface): what I'm doing and why

The issue proposes stored passes_immigration_eligibility_check_{program} flags. Verdict after reading the engine: for immigration eligibility, the coherent flow surface already exists — immigration_status_str + the engine's parameter lists + the takes_up_*_if_eligible propensities. The engine recomputes is_medicaid_immigration_status_eligible / is_aca_ptc_immigration_status_eligible / is_snap_immigration_status_eligible from stored inputs under whatever law vintage or reform is simulated, and the medicaid take-up stage already keeps off-domain persons at a draw-based propensity precisely so an eligibility shock enrolls newly-eligible people "at a plausible propensity instead of a silent zero" (its own docstring). A stored pass/fail verdict would freeze one law vintage into the artifact and stop reforms from recomputing eligibility — the opposite of "flows to other policies."

Where a stored flag family is the right mechanism is verification/documentation churn (whether a person can document status at redetermination — a person attribute the law can't derive), which needs engine-side consumer variables first, exactly like takes_up_*. I'll file that as a linked follow-up (microcosm + policyengine-us) rather than shipping a dead column now. If you want the passes_* family anyway, it layers on top of this design without rework — say so here.

Validation: gates and release probes

• Composition gate extends: per-category emitted-vs-target relative band (coarse, tolerant of ASEC undercoverage; explicit-zero categories must emit zero), plus the existing enum-domain / column-consistency / Pew-anchor checks. Gate details record every achieved-vs-target ratio and the humanitarian extraction so certification notes can quote them (acceptance criterion 3).
• Reform-coverage smoke probes (the release-blocking "$0 bound reform fails publication" machinery): (1) restore the pre-§71109 Medicaid status list at 2027 → medicaid must rise; (2) restore the below-FPL PTC exception in 2026 → aca_ptc must rise; (3) restore the pre-§71301 ACA lawful-presence list at 2027 → aca_ptc must rise; (4) restore the pre-§10108 SNAP status list at 2026 → snap must rise. Each binds only through immigration_status_str, so a release where the imputation regresses fails before publication — acceptance criterion 2, enforced permanently rather than checked once.
• CHIP: on engine 1.819.0 CHIP tests only ~UNDOCUMENTED, so no reform toggle can bind — that's CHIP immigration eligibility does not encode H.R.1 §71109 narrowing (Medicaid side is encoded; CHIP check is only ~undocumented) policyengine-us#9346. The data-side precondition (humanitarian-status children present in CHIP-age range; USCIS reports 8.9% of TPS beneficiaries under 18) is gated now; the CHIP probe gets added with the engine bump that lands #9346.
• years_since_us_entry stays at its engine default (5) in this PR. It is not stored on the release (the ABI lock declares it an absent engine input — the issue body's "fully populated" was the engine default, not data). Storing true entry years from PEINUSYR is the natural companion but flips recent-arrival LPRs out of baseline federal Medicaid — statutorily right for the five-year bar, wrong for reported-receipt children in the ~half of states using the CHIPRA-214/ICHIA lawfully-residing option, which the engine does not yet model. Filing as its own issue with that prerequisite. Consequence documented: PAROLED_ONE_YEAR is treated as past the bar at baseline, which matches the statutory position of the two biggest Medicaid-using parole cohorts (OAW Afghans and U4U Ukrainians hold refugee-equivalent benefits access by statute; the CHNV FR notice itself records that two-year paroles made CHNV parolees "generally … considered qualified aliens").
• CONDITIONAL_ENTRANT is not emitted: the INA §203(a)(7) class closed in 1980; survivors are indistinguishable from LPRs at this granularity. Documented in the stage notes.

Targets land in source_stages.json → regenerated us/spec/sources.yaml (schema extended; spec digests re-pinned as the attestation machinery requires); new seed-protocol sites per draw; tests cover assignment logic, manifest validation, determinism/clone-consistency, gate behavior, and the probe definitions.

[MaxGhenis]

Addendum: spill controls re-founded on Pew's 2023 universe (Aug 29 change, reviewed Sep 4)

One part of the design above is superseded in #779. The proposal said the EAD worker/student spill would still bind the remaining residual pool to Pew's 8.3M workers after the humanitarian draws. On the 2024 ASEC that left the residual worker count below the control with no spill (8.14M), because Pew's own count already includes the TPS/parole people the draws extract. The Aug 29 commit (405f9c2d) fixes the concept rather than the number:

• The controls now bind Pew's modeled universe — residual undocumented rows plus DACA, parolees, TPS holders and similar temporary protections (refugees and granted asylees excluded, as Pew does) — using Pew's superseding 2025 report and its coherent 2023 pair: 14.0M unauthorized / 9.7M in the labor force.
• The worker margin uses measured current labor-force status (A_LFSR, age 16+), restored from the pinned Census ASEC archives by exact PERIDNUM join (checkpoint schema v4), instead of a prior-year-earnings proxy.
• The composition gate anchors the same universe against 14.0M, identifying residual-derived Cuban/Haitian entrants by the (CUBAN_HAITIAN_ENTRANT, NON_CITIZEN_VALID_EAD) pair — exactly the set the spill counts.
• Side effect worth noting: it restores the DACA population the first draft had eroded (draws no longer starve the EAD pool DACA tags depend on).

Fresh-eyes check on the raw 2024 ASEC after merging current main: Pew labor force lands at 9.698M against the 9.7M control (binds), the broad universe at 14.71M (1.05× anchor), humanitarian categories PAROLED 372k / REFUGEE 129k / ASYLEE 156k / TPS 865k (all in band; parole and refugee saturate on survey undercoverage of 2022–24 arrivals plus the tighter U4U/CHNV arrival windows), NONE ⇔ UNDOCUMENTED intact. The four H.R.1 probe channels still move sign-correctly. Everything else in the proposal stands; #776/#777/#778 remain the filed follow-ups.