[MaxGhenis]

This PR charges the Medicare Part B premium only to the people who pay it, through a new Part B enrollment input and a state buy-in for SSI recipients. The model charged every Medicare enrollee, which overstates premiums in Schedule A, Idaho's premium subtraction and every other reader.

Gaps and witnesses

Each witness takes a single 70-year-old in 2025, when the standard premium runs $185 a month ($2,220 a year), unless the row names 2026.

Gap	Witness on main	This PR
Part A-only enrollees. medicare_enrolled covers Part A and/or Part B, and the model gated the Part B premium on it, so no input could say a person has Part A only.	TX, $40,000 interest: medicare_part_b_premium 2,220	New medicare_part_b_enrolled input (defaults to medicare_enrolled). With it false: premium 0, medical_expense_health_insurance_premiums 0, gross_medicare_part_b_premium_if_enrolled 0
State buy-in for SSI recipients. States pay Part B for SSI recipients under their Part B buy-in agreements (42 U.S.C. 1395v; 42 CFR part 407 subpart C), and the model netted out only MSP coverage.	CA, reported receives_ssi, $30,000 interest (above the 135% FPL MSP limit): Medicaid via SSI, msp_part_b_premium_coverage 0, premium 2,220	New medicaid_part_b_buy_in_premium_coverage pays the standard premium, month by month, for an SSI recipient in a Section 1634 or SSI-criteria state and for a 209(b) SSI recipient who qualifies for Medicaid; premium 0
SSI recipients in 209(b) states. CMS puts in buy-in group 1 "Individuals who receive SSI or SSP or both and are covered under the State's Medicaid state plan as categorically needy" (42 CFR 407.42(b)(1)(i)), with no 209(b) exception.	CT, reported SSI, $30,000 interest, $40,000 other medical expenses: spends down to Medicaid (is_209b_ssi_recipient_for_medicaid true, medicaid_enrolled true), premium 2,220	Buy-in 2,220, premium 0. Medicaid eligibility counts the premium before the buy-in (medicare_part_b_premium_before_buy_in), since Congress starts buy-in coverage no earlier than "the first day of the first month in which he is both an eligible individual and a member of a coverage group" (42 U.S.C. 1395v(d)(2)(C)).
SSI receipt month by month. The year's receives_ssi holds December's value, so an annual SSI read would buy in December-only SSI for all 12 months and January-only SSI for none.	TX, $30,000 interest, 2026, SSI in one month (every month entered): premium 2,434.80 for December-only and January-only SSI alike	The buy-in counts each month with SSI (ssi > 0 or receives_ssi): 202.90 buy-in and 2,231.90 premium for both
Idaho's MSP case, which main already gets right.	ID, $12,000 Social Security: MSP pays all 12 months, premium 0, subtraction 0	A new test pins it

Readers of the modeled premium

All of these read medicare_part_b_premium (directly or through medical_expense_health_insurance_premiums or spm_unit_health_insurance_premiums) and inherit both fixes:

• Aggregates: medical_expense_health_insurance_premiums, spm_unit_health_insurance_premiums, then spm_unit_medical_out_of_pocket_expenses.
• Federal tax: itemized_medical_expenses and itemized_medical_expenses_excluded_premiums (Schedule A), self_employed_health_insurance_ald_excluded_premiums.
• State tax: ID id_qualified_health_insurance_premiums; MT mt_premium_deduction_joint, mt_medical_expense_deduction_joint, mt_eligible_medical_insurance_premiums; OH oh_medical_care_insurance_premiums; WA senior/disabled property tax exemption income deductions.
• Benefits: SNAP snap_allowable_medical_expenses (and OK LIHEAP through it), HUD hud_medical_expenses, PA CCW, UT HEAT, ME HEAP and NH FAP. The LIHEAP/EAP income definitions read it for AR, AZ, GA, IA, IN, KY, LA, MD, MN, MO, MT, ND, NJ, NV, NY, OH, PA, RI, SD, TN, VA, WA, WI, WY.
• Gross-premium readers: gross_medicare_part_b_premium_if_enrolled feeds medicare_cost and the CA disregard of the full Part B premium, and both now take the Part B enrollment gate.

Two Medicaid eligibility readers, medicaid_medically_needy_medical_expenses and medicaid_optional_senior_or_disabled_countable_income, read the premium before the buy-in (medicaid_eligibility_health_insurance_premiums, medicare_part_b_premium_before_buy_in). They take the Part B enrollment gate but not the buy-in, which starts only once the person qualifies for Medicaid.

Changes

• medicare_part_b_enrolled (new): the formula returns medicare_enrolled; enter false for Part A only.
• medicare_part_b_premium: defined_for Part B enrollment. It nets the larger of MSP coverage and the SSI buy-in, since both pay the same standard premium, and floors at 0. The enrollee keeps paying IRMAA above the standard premium.
• medicaid_part_b_buy_in_premium_coverage (new): counts each month the person receives SSI and either passes the Section 1634 or SSI-criteria test or counts as a 209(b) SSI recipient who qualifies for Medicaid. It tests eligibility, with no take-up step. It pays the standard premium only. SSA instructs that "States and territories pay the standard premium for individuals covered under state buy-in" (POMS HI 00815.001). CMS has beneficiaries pay the income-related amount "in addition to the Part B (SMI) standard monthly premium" (42 CFR 408.28).
• medicare_part_b_premium_before_buy_in (new): the gross premium net of MSP, before the buy-in.
• medicaid_eligibility_health_insurance_premiums (new): medical_expense_health_insurance_premiums with the premium before the buy-in. CMS lists "Medicare and other health insurance premiums" among the expenses a 209(b) state deducts (42 CFR 435.121(f)(1)(iii)).
• medicaid_medically_needy_medical_expenses reads medicaid_eligibility_health_insurance_premiums. medicaid_optional_senior_or_disabled_countable_income reads it and medicare_part_b_premium_before_buy_in for paid_premiums and paid_part_b.
• gross_medicare_part_b_premium_if_enrolled: defined_for Part B enrollment.
• medical_expense_health_insurance_premiums and the CA optional senior/disabled disregard gate Part B on medicare_part_b_enrolled.
• The documentation of spm_unit_medical_out_of_pocket_expenses and the comments in OK LIHEAP and the WA exemption sources now name the netting.

Not in this PR

• Other full-benefit Medicaid groups. CMS offers states a buy-in group of "all individuals eligible for Medicaid" (42 CFR 407.42(b)(3)), and coverage for those groups starts later (42 CFR 407.47(d)), so the result depends on each state's agreement. NY, $30,000 Social Security, $40,000 other medical expenses: medically needy, medicaid_enrolled true, premium 2,220.
• SSP-only and deemed SSI recipients (42 CFR 407.42(b)(1)(i) and (ii)). The buy-in reads ssi and receives_ssi only.
• Illinois SMIB. il_smib_categorical_eligible reads is_ssi_eligible, which applies no income test, so an IL 70-year-old with $40,000 of interest gets il_smib 2,220. Netting it would zero Part B for nearly every Illinois enrollee.
• State MSP income limits. The model applies the federal 100%, 120% and 135% of FPL limits in every state.
• The limit on Part B premium increases for Social Security recipients (42 U.S.C. 1395r(f)).
• Coverage that alternates between programs within a year. The premium nets the annual larger of MSP coverage and the buy-in. CA, $30,000 interest, msp_income_eligible true January through June, receives_ssi true July through December: MSP 1,110, buy-in 1,110, premium 1,110, where the state pays all 12 months. No annual input reaches this case, and the fix touches msp_part_b_premium_coverage, which Add MSP take-up gate and any-month Kansas CSFP categorical eligibility #9495 and Apply immigration tests to Medicare and Medicare Savings Programs #10632 edit.
• The enrollment gate on msp_part_b_premium_coverage. A Part A-only enrollee who qualifies for MSP shows 2,220 of coverage there, and no reader uses it. The gate can follow Add MSP take-up gate and any-month Kansas CSFP categorical eligibility #9495 and Apply immigration tests to Medicare and Medicare Savings Programs #10632.
• medicare_cost for a Part A-only enrollee. ID, $12,000 Social Security: 17,751 with Part A only against 15,531 with Part B, since the per-capita series includes Part B spending. Only the new input reaches this case.

Tests

• YAML: medicaid_part_b_buy_in_premium_coverage.yaml 10, medicare_part_b_premium.yaml 19, medicare_part_b_enrolled.yaml 4, medicaid_medically_needy_medical_expenses.yaml 3. The buy-in file covers CA full-year SSI, modeled SSI, no SSI, Part A only, three CT spend-downs, CT December-only SSI, and TX December-only and January-only SSI. One CT spend-down turns on the premium before the buy-in, and one lands on the limit.
• Existing Part B premium cases used people with no current income, whom the model now finds to be SSI recipients with a buy-in, so they enter ssi: 0 to keep testing the uncovered premium.
• Two cases entered medicare_part_b_premium and expected Medicaid to read it, and they now enter medicare_part_b_premium_before_buy_in with the same expected values.
• Property (test_part_b_msp_offset.py, 15 cases): 200 random draws of gross premium, MSP coverage, buy-in coverage and Part B enrollment in one simulation. The paid premium stays at or above 0 and at or below gross, and reads 0 without Part B.
• Mutant: dropping the modeled-SSI term from the buy-in fails the modeled-SSI case.
• Mutant: requiring strictly less than the limit in the 209(b) test fails the case that lands on the limit.
• CI: the Pull request run at this head (42054053269) succeeded, 36 of 36 jobs.

No partner test changed.

Invariants

These hold for every input, and test_part_b_msp_offset.py checks them with random draws:

1. 0 ≤ medicare_part_b_premium ≤ the gross premium.
2. Without Part B enrollment, the paid premium reads 0.

The YAML overlap case pins the netting: a person covered by both MSP and the buy-in counts once.

Overlaps with open PRs

• Limit SSI spousal-deeming child allocations to the couple's co-resident children #10146 moves medicaid_optional_senior_or_disabled_countable_income's formula into a new monthly file, which reads medical_expense_health_insurance_premiums and medicare_part_b_premium and gates both Part B terms on medicare_enrolled. Limit SSI spousal-deeming child allocations to the couple's co-resident children #10146's branch already conflicts with main in three files: the annual file, ca_state_supplement.py and ca_wdp_ssi_ssp_income_eligible.py. This PR edits only the annual file. Whoever lands second carries three changes from this PR into the monthly file: paid_premiums reads medicaid_eligibility_health_insurance_premiums, paid_part_b reads medicare_part_b_premium_before_buy_in, and both Part B terms take the medicare_part_b_enrolled gate. Without the two premium reads, the buy-in and the 209(b) test read each other, and the simulation stops with a cycle error. Without the gate, the model disregards the full Part B premium for a California enrollee with Part A only. Whoever lands second also deletes the monthly file's note that medicare_enrolled stands in for Part B enrollment.
• Exclude employer-plan-eligible premiums from the self-employed health insurance deduction #10122 and issue Self-employed health insurance deduction leaves out the Medicare premiums the model computes (26 U.S.C. 162(l)) #10530 concern the self-employed health insurance deduction. self_employed_health_insurance_ald_excluded_premiums reads medical_expense_health_insurance_premiums, which this PR edits by one gate line.
• Count only filer-paid premiums in Schedule A medical expenses #10270 edits itemized_medical_expenses, which reads the same aggregate, with no shared lines.
• Add MSP take-up gate and any-month Kansas CSFP categorical eligibility #9495 and Apply immigration tests to Medicare and Medicare Savings Programs #10632 edit msp_part_b_premium_coverage, and this PR leaves that file alone.

Impact

We estimate that this branch changes none of the outputs we measured in 2026 on the default dataset. We compared main at the merge base 1104c5edb5 with this branch at 6c5d72e29f.

The branch lowers a Part B premium only where the new buy-in would pay more of it than the Medicare Savings Program already pays. No record on the default dataset meets that test in 2026, so no premium, tax or net income moves. The dataset sets no medicare_part_b_enrolled, so that variable follows medicare_enrolled for every record, and the branch's Part B gates match main's. Main lacks medicaid_eligibility_health_insurance_premiums, which this branch adds, so we couldn't compare it.

Output	Main total	Change	Records changed (weighted)
medicare_part_b_premium	$144.4bn	$0	0 (0)
medical_expense_health_insurance_premiums	$649.4bn	$0	0 (0)
state_income_tax	$578.0bn	$0	0 (0)
income_tax	$2,595.1bn	$0	0 (0)
household_benefits	$2,060.0bn	$0	0 (0)
household_net_income	$17,161.7bn	$0	0 (0)

impact: no change in 2026 on the default dataset to Part B premiums, health insurance premiums, taxes, benefits or household net income.
axiom: TheAxiomFoundation/rulespec-us#2032 queued (corpus release re-pin and encoder approval); TheAxiomFoundation/rulespec-us#2006, TheAxiomFoundation/rulespec-us#1931, TheAxiomFoundation/rulespec-us#1933 and TheAxiomFoundation/rulespec-us#1625 queued, extended with comments (209(b) premiums, medically needy expenses, California Part B enrollment and the Idaho payer-paid regression).

🤖 Generated with Claude Code

[MaxGhenis]

Ready for hub review

Head: be36a24573367c783b5e1e3fac2c3efa878012fa (draft; branch claude/medicare-part-b-audit-xwdnw6. The cloud session was pinned to this branch name, not medicare-part-b-payers. Both names returned 404 before I pushed, and the search for open PRs found none making the same change.)

What the item asked, and how I resolved it

1. List every reader of the modeled premium. The full table is in the PR body under "Readers of the modeled premium". Every reader goes through medicare_part_b_premium, directly or via medical_expense_health_insurance_premiums / spm_unit_health_insurance_premiums. The gross-premium readers are gross_medicare_part_b_premium_if_enrolled → medicare_cost, and the CA full-premium disregard.
2. A witness for each gap at origin/main 4e71481. Each is a single 70-year-old in 2025, standard premium $2,220:
   • Part A only: no input exists. TX, $40,000 interest → premium 2,220.
   • Reported SSI, CA, $30,000 interest → Medicaid via SSI, MSP 0, premium 2,220.
   • Medically needy, NY, $30,000 Social Security + $40,000 other medical → medicaid_enrolled true, premium 2,220.
   • Idaho, $12,000 Social Security → MSP pays, premium 0, subtraction 0. Already correct.
3. Fix the coverage cases.
   • Part A-only enrollees: new medicare_part_b_enrolled (medicare_part_b_enrolled.py:24, defaults to medicare_enrolled). The premium is gated on it at medicare_part_b_premium.py:10, and so are gross_medicare_part_b_premium_if_enrolled.py:10, medical_expense_health_insurance_premiums.py:25-26 and the CA disregard at medicaid_optional_senior_or_disabled_countable_income.py:83-86.
   • State buy-in for SSI recipients: new medicaid_part_b_buy_in_premium_coverage (medicaid_part_b_buy_in_premium_coverage.py:43), covering Section 1634 and SSI-criteria states. It is netted with MSP as max_ at medicare_part_b_premium.py:23 and floored at 0 at :27.
4. Idaho regression where MSP pays and the subtraction is 0. id_health_insurance_premiums_subtraction.yaml:499.
5. Property: paid is between 0 and the premium. test_part_b_msp_offset.py:422. It runs 200 seeded random draws in one simulation and checks: paid ≥ 0, paid ≤ gross, paid = 0 without Part B, and paid = max(gross − max(MSP, buy-in), 0).

Tests run (locally, one file per process)

• Changed YAML files:
  • medicare_part_b_premium.yaml: 19/19
  • medicare_part_b_enrolled.yaml: 4/4
  • medicaid_part_b_buy_in_premium_coverage.yaml: 4/4
  • Idaho subtraction: 24/24
  • CA optional senior/disabled countable income: 42/42
• test_part_b_msp_offset.py: 15/15.
• Sweep: all 125 YAML files under tests/ that mention medicare_part_b, medicare_enrolled, medical_expense_health_insurance_premiums, msp_part_b, spm_unit_health_insurance_premiums or medical_out_of_pocket, excluding partners. All passed: 1,921 cases, 0 failures. No partner file references these variables, and none was edited.
• No microsimulation was run; the hub measures that.
• Cost of the property test, measured locally while the sweep shared the CPU (not a Linux CI figure): 17.1 s in pytest, 59.5 s end to end including model build, peak RSS 1.45 GB. It reuses the file's module-level SYSTEM, so it builds no new tax-benefit system. It belongs to whichever CI group already runs test_part_b_msp_offset.py; I didn't change any routing.

Hand computations

• Standard premium: 2025 $185 × 12 = $2,220; 2026 $202.90 × 12 = $2,434.80.
• CA Part A-only: (1,000 − 20) × 12 = 11,760, with no Part B disregard. The Part B case gives 9,325.20.
• MSP + buy-in overlap: gross 4,440 − max(2,220, 2,220) = 2,220. The income-related amount stays with the enrollee.

Mutants (each run in a separate worktree and process, code restored after each)

Mutant	Result
M1 premium gated on medicare_enrolled	killed (2)
M2 buy-in dropped	killed (1)
M3 MSP + buy-in summed instead of max	killed (2)
M4 no floor at 0	killed (YAML 1, property 1)
M5 gross-if-enrolled gated on medicare_enrolled	killed (1)
M6 CA disregard gated on medicare_enrolled	killed (1)
M7 buy-in includes 209(b) states	killed (1)
M8 buy-in ignores SSI receipt	killed (1)
M9 medicare_part_b_enrolled always true	killed (2)
M10 aggregate gated on medicare_enrolled	survived. It is equivalent for computed premiums, because the premium is already zero without Part B; it differs only when someone enters medicare_part_b_premium directly for a person who isn't in Part B.

CI

Draft, so the Pull request workflow at be36a245 shows completed / skipped. CI has not run.

Judgment calls

• Couldn't read the law. www.ecfr.gov and www.law.cornell.edu are blocked in this session (web.archive.org and govinfo too), so I could not read 42 CFR 406-407, 42 U.S.C. 1395v or 1396a(a)(10)(E). The code cites them (1395o; 1395v; 42 CFR part 407 subparts B and C) and quotes nothing. Please check those references before landing.
• Default for medicare_part_b_enrolled. I made it a formula returning medicare_enrolled, not an input with default_value=True. An input-only variable ignores defined_for, and with that version people outside Medicare came out as true (caught by tests). Data outputs don't change unless a dataset supplies the new input.
• SSI buy-in scope. It covers only the Section 1634 and SSI-criteria branch of is_ssi_recipient_for_medicaid. The 209(b) branch reads medicaid_optional_senior_or_disabled_countable_income, which deducts this premium, so including it gave a CycleError. A test pins the 209(b) exclusion, using CT.
• No take-up or immigration gate on the buy-in. It follows the existing SSI-recipient Medicaid test, with no take-up or immigration gate added.
• Existing tests changed. The 11 existing medicare_part_b_premium.yaml cases and test_medicare_part_b_premium_does_not_depend_on_calculation_order used people with no current-year income. The model now finds them SSI recipients with a buy-in, so they enter ssi: 0 to keep testing the uncovered premium. No expected values changed.
• Files left alone.
  • msp_part_b_premium_coverage is left alone because Add MSP take-up gate and any-month Kansas CSFP categorical eligibility #9495 and Apply immigration tests to Medicare and Medicare Savings Programs #10632 both edit it. Its only reader is the gated premium, but it can still show coverage for a Part A-only person.
  • msp_benefit_value is left alone.
• Medically needy and other full-benefit Medicaid not fixed. Whether a state's buy-in pays for them depends on its agreement, which I couldn't source. Their Medicaid tests also deduct this premium, so reading them here would be circular.
• Illinois SMIB not netted. il_smib pays $2,220 to an IL senior with $40,000 of interest, because il_smib_categorical_eligible reads is_ssi_eligible, which isn't income-tested. Netting it would wipe out Part B for nearly every IL enrollee.
• Shared lines.
  • Self-employed health insurance deduction leaves out the Medicare premiums the model computes (26 U.S.C. 162(l)) #10530 and Exclude employer-plan-eligible premiums from the self-employed health insurance deduction #10122 are untouched. self_employed_health_insurance_ald_excluded_premiums reads medical_expense_health_insurance_premiums, whose Part B gate (lines 25-26) this PR changes.
  • Count only filer-paid premiums in Schedule A medical expenses #10270 edits itemized_medical_expenses, which reads the same aggregate; there are no overlapping lines.

Proposed follow-up issues

1. Net state buy-ins for full-benefit Medicaid groups beyond SSI from the Part B premium. Add a by-state parameter for buy-in coverage groups from the State Buy-In Manual. Break the medically needy spend-down cycle, for example by testing spend-down with the gross premium.
2. Illinois SMIB treats every aged/blind/disabled person as an SSI recipient. il_smib_categorical_eligible uses is_ssi_eligible, which isn't income-tested; it should use SSI receipt. The fix should then net SMIB from medicare_part_b_premium.
3. State MSP income limits above 135% FPL (42 U.S.C. 1396a(r)(2)). msp_income_eligible uses only the federal QI limit, so states with higher MSP limits (for example CT, DC) leave the premium on enrollees whose premium the state pays. I didn't verify this against state sources.

---

Generated by Claude Code

[MaxGhenis]

Hub review in progress

Independent review of be36a24573367c783b5e1e3fac2c3efa878012fa, continued by a third session. The head hasn't moved since the hand-off. The final review is posted: #10634 (comment). It supersedes this comment.

Rechecked at this head (third session)

• CI. Pull request run 41579751139 at this head: 34 of 34 jobs success.
• The defect reproduces at current main (89a15b72): TX, $40,000 interest, Part B 2,220 with no Part A-only input; CA, reported SSI, $30,000 interest, Medicaid via SSI, Part B 2,220. At the head: Part A-only gives 0, CA SSI gives buy-in 2,220 and premium 0.
• Changed tests, one file per process at the head: Part B premium 19/19, medicare_part_b_enrolled 4/4, buy-in 4/4, Idaho subtraction 24/24. The 11 existing premium cases only gain ssi: 0; no expected value changed.
• Core claim holds. policyengine-core 3.34.2 simulation.py masks a defined_for result with np.where(mask, array, variable.default_value), so an input-only variable with default_value=True would read true outside Medicare. The formula version avoids that.
• Sources. From this VM, ecfr.gov, govinfo.gov, uscode.house.gov, law.cornell.edu and ssa.gov all fail at the proxy. I quote only the texts the hub supplied.

Findings so far

• P2: the 209(b) gap can close cheaply. The PR's CT test person isn't on Medicaid (is_ssi_recipient_for_medicaid false at main and head), so it can't show the gap. A real witness adds $40,000 of other medical expenses: 209(b) Medicaid true, buy-in 0, premium 2,220 at the head. On a scratch tree, letting Medicaid's eligibility chain (optional senior/disabled countable income and medicaid_medically_needy_medical_expenses) read premiums net of MSP but before the buy-in, and having the buy-in read is_ssi_recipient_for_medicaid, removes the CycleError and gives buy-in 2,220 and premium 0 for that person. On the scratch tree, the 104 YAML files that touch this chain pass 1,797 cases and fail 2, both of which enter the premium directly.
• Corrected: the January-only SSI witness here was wrong. A single monthly receives_ssi input carries over to every month. With every month entered, the annual read follows December: December-only SSI buys in 12 months, and January-only buys in none. See the final review, P2-1.
• P3: the property test re-derives the formula from inputs. It enters gross premium, MSP and buy-in as inputs, so it never runs the MSP or buy-in formulas. Its seed (10634) is fixed, and its draws do reach each netting branch: 62 MSP only, 20 buy-in only, 53 both, 39 where the floor binds, 43 outside Part B.
• Mutants: M3, M7, M8 and M13 are killed; M10 and M14 survive as equivalent. Details are in the final review.

---

Generated by Claude Code

[MaxGhenis]

Ready for hub review
VERDICT: REQUEST_CHANGES
be36a24
METHODOLOGY_CHOICE: yes

Independent review of #10634 at be36a245. The head hasn't moved since the hand-off. The PR picks readings where the law leaves room: Part B enrollment defaults to every Medicare enrollee, the buy-in pays only the standard premium, and SSI receipt is read per year. Two P2 findings ask for small changes; both fixes below were run on scratch trees. Nothing here is a P1.

Sources: from this VM, ecfr.gov, govinfo.gov, uscode.house.gov, law.cornell.edu and ssa.gov all failed at the proxy. I quote only the texts the hub supplied: 42 CFR 407.10, 407.40 and 407.42 (eCFR), and 42 U.S.C. 1395v, 1395o, 1383c and 1396a(a)(10)(E).

Findings

P2-1. December SSI buys in all 12 months; SSI that ends before December buys in none

medicaid_part_b_buy_in_premium_coverage.py:40-42 reads add(person, period, ["receives_ssi"]) > 0 for the year. receives_ssi is a monthly boolean, and the annual read returns December's value alone. (Correction to my interim comment: its witness entered SSI for January only, and core carried that input into every month, so that person had SSI all year.)

Witness: TX, age 70, $30,000 interest, 2026, receives_ssi entered for every month:

SSI months	receives_ssi (year)	buy-in at head	premium at head	premium at main (89a15b72)
December only	True	2,434.80	0	2,434.80
January only	False	0	2,434.80	2,434.80

Under 42 U.S.C. 1395v(d)(2)(C), coverage starts with "the first month in which he is both an eligible individual and a member of a coverage group". Under (d)(3)(A), it ends with the month he is found ineligible for money payments. So the State pays about one month either way: 202.90 of buy-in, 2,231.90 paid by the enrollee. For the December case the head is off by 2,231.90, where main was off by 202.90. This PR makes that reader worse.

Fix (the second line keeps ssi > 0 as well as reported receipt):

        monthly_premium = person("base_part_b_premium", period) / MONTHS_IN_YEAR
        coverage = 0
        for month in period.get_subperiods(MONTH):
            receives_ssi = (person("ssi", month) > 0) | person("receives_ssi", month)
            covered = receives_ssi & is_covered & covered_classification
            coverage += where(covered, monthly_premium, 0)
        return coverage

On a scratch tree this gives buy-in 202.90 and premium 2,231.90 for both rows. It still gives 2,220 for CA full-year SSI. At this patch, the premium (19), buy-in (4) and Idaho (24) YAML files and test_part_b_msp_offset.py (15) all pass. Please add a December-only case to medicaid_part_b_buy_in_premium_coverage.yaml. Enter every month, because one monthly input carries over to the rest.

Follow-up, same on main: is_ssi_recipient_for_medicaid and is_209b_ssi_recipient_for_medicaid read receives_ssi the same way, so the SSI route to Medicaid also follows December alone.

P2-2. The 209(b) gap closes cheaply, and the test that pins it uses someone who isn't on Medicaid

medicaid_part_b_buy_in_premium_coverage.yaml:31-41 (CT, reported SSI, $30,000 interest) and the body's 209(b) witness use a person with is_ssi_recipient_for_medicaid false and medicaid_enrolled false, at main and at the head. The case passes with or without 209(b) netting, so it pins nothing. 42 CFR 407.42(b)(1)(i) puts in group 1, which every buy-in group includes, "Individuals who receive SSI or SSP or both and are covered under the State's Medicaid state plan as categorically needy".

A real witness adds other_medical_expenses: 40_000. At the head: is_209b_ssi_recipient_for_medicaid true, medicaid_enrolled true, buy-in 0, premium 2,220.

The cycle has two arms: the 209(b) test reads medicaid_optional_senior_or_disabled_countable_income and also medicaid_medically_needy_medical_expenses, and both read the net premium. It breaks if Medicaid eligibility reads premiums net of MSP but before the buy-in. On a scratch tree I did that:

• a new medicare_part_b_premium_before_buy_in: defined_for = "medicare_part_b_enrolled", returning max_(gross_medicare_part_b_premium - msp_part_b_premium_coverage, 0);
• a new medicaid_eligibility_health_insurance_premiums: the same formula as medical_expense_health_insurance_premiums, but reading the premium before the buy-in;
• medicaid_medically_needy_medical_expenses adds medicaid_eligibility_health_insurance_premiums instead of medical_expense_health_insurance_premiums;
• in medicaid_optional_senior_or_disabled_countable_income.py:84-85 (lines 84 and 85 at this head), paid_premiums reads medicaid_eligibility_health_insurance_premiums, and paid_part_b reads medicare_part_b_premium_before_buy_in;
• the buy-in formula becomes return person("is_ssi_recipient_for_medicaid", period) * person("base_part_b_premium", period). That also removes the copy of that variable's first branch.

Result: no CycleError, and the CT witness gives buy-in 2,220 and premium 0. Across all 104 YAML files that mention the premium, its aggregate, 209(b), the SSI-recipient Medicaid test, the countable income or medically needy expenses (partners excluded), 1,797 cases pass and 2 fail. Both failures enter medicare_part_b_premium directly and expect it to reach Medicaid: medicaid_medically_needy_medical_expenses.yaml:15 and the CA countable-income case at line 398, "California does not count a self-paid Part B premium twice…". They would enter the new before-buy-in variable instead.

The reading this takes, that Medicaid eligibility counts premiums the person owes before the State's buy-in starts, is a methodology choice. If the author keeps the exclusion instead, please (a) replace the CT case with the spend-down witness above, expecting buy-in 0 and premium 2,220, and (b) say in the body that 209(b) SSI recipients who get Medicaid still pay Part B in the model. The hub would then file a follow-up.

P3-1. The property test restates the formula from inputs

test_part_b_msp_offset.py:422-466 enters gross_medicare_part_b_premium, msp_part_b_premium_coverage and medicaid_part_b_buy_in_premium_coverage as inputs, so the MSP and buy-in formulas never run. Line 465 recomputes max(gross - max(msp, buy_in), 0), the formula under test. The bounds (≥ 0, ≤ gross, 0 outside Part B) are real properties. The seed (10634) is fixed, so the test is deterministic, and the draws reach each netting branch:

• 62 MSP only;
• 20 buy-in only;
• 53 both;
• 39 where the floor binds;
• 43 outside Part B;
• 37 where sum and max differ (M3 fails on exactly these 37).

Suggestion: keep the bounds, and drop line 465 or move it into the YAML overlap case that already exists.

P3-2. Citations should point at the sections the code relies on

• medicaid_part_b_buy_in_premium_coverage.py:12-13. 1395v as a whole supports a State agreement. The rule the code applies is 42 CFR 407.40(c)(5) ("In a 1634 State, CMS enrolls SSI beneficiaries in Medicare Part B, on behalf of the State") and 407.42(b)(1)(i). Replacement:

      reference = (
          "https://www.law.cornell.edu/uscode/text/42/1395v#d",
          "https://www.ecfr.gov/current/title-42/section-407.40#p-407.40(c)(5)",
          "https://www.ecfr.gov/current/title-42/section-407.42#p-407.42(b)(1)(i)",
      )

  (1395v(d)(1) is the premium the State pays, "without any increase under subsection (b)".)
• medicare_part_b_enrolled.py:17-18. Eligibility to enroll is 1395o(a) and 42 CFR 407.10. Use .../uscode/text/42/1395o#a and https://www.ecfr.gov/current/title-42/section-407.10. I couldn't fetch Cornell's subpart pages, so I can't say whether the current subpart-B and subpart-C links land on these sections.

P3-3. The body has stale lines

• "CI has not run: this is a draft." The PR is no longer a draft. Pull request run 41579751139 at this head finished success, 34/34 jobs.
• The 209(b) row's witness: see P2-2.
• The axiom: line reads axiom: (hub files pe-parity before landing), which is the brief's wording. The hub should replace it with a final form before landing.

Checked and sound

• Defect reproduces at main (89a15b72):

  • TX, 70, $40,000 interest: premium 2,220, and no input can mark Part A only.
  • CA, reported SSI, $30,000 interest: is_ssi_recipient_for_medicaid true, MSP 0, premium 2,220.

  At the head, Part A only gives 0 for the premium, medical_expense_health_insurance_premiums and gross_medicare_part_b_premium_if_enrolled, and the CA case gives buy-in 2,220 and premium 0.

• Buy-in coverage against 407.42 group 1 (check 1). It tests SSI receipt (modeled ssi or reported receives_ssi), the state's ssi_recipient.is_covered, and a 1634 or SSI-criteria classification. That is exactly the first branch of is_ssi_recipient_for_medicaid, the model's categorically needy SSI route. The parameters classify 35 jurisdictions as 1634, 8 as SSI-criteria and 8 as 209(b), and all eight 209(b) states have is_covered false.

• Groups 2 and 3 (MSP and other full-benefit Medicaid). The PR changes nothing for anyone without SSI receipt or a Part A-only input: the premium only drops by max(MSP, buy-in), never rises. So no case is worse than main except the December SSI case in P2-1. Medically needy NY (hand-off witness) still pays 2,220, the same as main.

• medicare_part_b_enrolled as a formula (check 3). In policyengine-core 3.34.2, simulation.py masks a defined_for result with np.where(mask, array, variable.default_value). An input-only variable with default_value=True would therefore read true outside Medicare, so the hand-off's reason holds. No file in this repo besides the PR's own reads or sets it. A code search across the repositories this session can reach found it in no PolicyEngine dataset code.

• Netting and hand values (check 4):

  • Model standard premium: 2025 base_part_b_premium 2,220 = 185 × 12; 2026 2,434.80 = 202.90 × 12. I couldn't reach CMS to confirm the parameter.
  • CA Part B: (1,000 − 20 − 202.90) × 12 = 9,325.20.
  • CA Part A-only: (1,000 − 20) × 12 = 11,760. Both CA cases pass.
  • Overlap: 4,440 − max(2,220, 2,220) = 2,220, so the State pays once and the income-related amount stays with the enrollee. 1395v(d)(1) excludes only the 1395r(b) increase; see "What I couldn't check" for IRMAA.

• Changed tests (check 5). The 11 existing premium cases and the calculation-order test only gain ssi: 0; the diff changes no expected value. At the head, one file per process: premium 19/19, enrolled 4/4, buy-in 4/4, Idaho 24/24, CA countable income 42/42, test_part_b_msp_offset.py 15/15. One IRMAA case without ssi: 0 could test the buy-in netting through real formulas, but the YAML overlap case covers the arithmetic.

• Mutants (check 6). Each ran in its own worktree and process against the six changed files, and the trees were removed afterwards:

  Mutant	Result
  M3 sum instead of max	killed: premium YAML 2, property test (37/200 draws)
  M7 209(b) included	killed: CycleError in premium 16, buy-in 3, CA 32, Idaho 2, Python
  M8 SSI receipt ignored	killed: premium 12, buy-in 1, CA 2, Idaho 1, Python 1
  M10 aggregate gated on medicare_enrolled	survives, equivalent as stated. Part A-only with medicare_part_b_premium: 1000 entered: aggregate 0 at head, 1,000 under M10; computed premiums are equal.
  M13 (mine) buy-in ignores reported receives_ssi	killed: premium 1, buy-in 1
  M14 (mine) buy-in drops the classification test	survives, equivalent today: every 209(b) state also has is_covered false

• Readers (check 5 and 8). medicare_part_b_premium has 26 reading variables and 3 parameter lists: SPM premiums, the premium aggregate, CA countable income, NH FAP, LIHEAP/EAP in 22 states, and the RI and OH LIHEAP and WA exemption lists. medical_expense_health_insurance_premiums has 14 variables and 2 parameter lists: HUD, Schedule A and its excluded premiums, the self-employed ALD exclusion, medically needy, SNAP, PA CCW, UT HEAT, OH, ID, three MT readers, ME HEAP and WA. gross_medicare_part_b_premium_if_enrolled feeds only medicare_cost. msp_part_b_premium_coverage has no new reader. msp_part_a_premium_coverage names it only in a comment.

• Neighbours (check 7). git merge-tree of this head merges cleanly with Add MSP take-up gate and any-month Kansas CSFP categorical eligibility #9495, Apply immigration tests to Medicare and Medicare Savings Programs #10632, Exclude employer-plan-eligible premiums from the self-employed health insurance deduction #10122, Count only filer-paid premiums in Schedule A medical expenses #10270, Add California LIHEAP regular heating assistance #10607, Add Connecticut CEAP regular heating assistance #10611, Add Delaware LIHEAP regular heating assistance #10612, Add Idaho LIHEAP regular heating assistance #10613, Apply 416.1163(e)(2) to deemed SSI and identify SSI couples by marriage, not tax roles #10141, Exclude people otherwise eligible for Medicaid from Illinois HBWD and use Illinois's income rules for its disability pathway #10631 and Count dependents' other medical expenses in Idaho's claimant medical worksheet #10278. Limit SSI spousal-deeming child allocations to the couple's co-resident children #10146's conflicts lie in files this PR doesn't touch.

  • Count only filer-paid premiums in Schedule A medical expenses #10270 edits the shared Idaho YAML only up to about line 430; this PR appends after line 497.
  • Add MSP take-up gate and any-month Kansas CSFP categorical eligibility #9495 and Apply immigration tests to Medicare and Medicare Savings Programs #10632 change MSP and Medicare eligibility. Their added lines don't read these variables, and either landing order keeps both.
  • Add Connecticut CEAP regular heating assistance #10611, Add Delaware LIHEAP regular heating assistance #10612 and Add Idaho LIHEAP regular heating assistance #10613 add LIHEAP readers of medicare_part_b_premium, and Add California LIHEAP regular heating assistance #10607 one of the aggregate. They inherit the net premium, which is the intended reading.
  • Self-employed health insurance deduction leaves out the Medicare premiums the model computes (26 U.S.C. 162(l)) #10530 (issue) and Exclude employer-plan-eligible premiums from the self-employed health insurance deduction #10122 read the aggregate whose gate changes on one line. There's no conflict in meaning.

• Scope (check 7). Nothing changes under .github/ or the partner tests. The changelog fragment is a top-level changelog.d/medicare-part-b-payers.fixed.md.

• CI: run 41579751139 at this head, 34/34 jobs success.

Baseline movement for the hub's impact pair (2026)

medicare_part_b_enrolled defaults to medicare_enrolled, so outputs move at baseline only through the buy-in. They move for Part B enrollees with SSI (modeled or reported) in 1634 and SSI-criteria states, where MSP pays less than the full standard premium. Typical cases are reported SSI with income or assets over the MSP limits, and, until P2-1 is fixed, people whose December SSI flag covers the whole year. Variables to report:

• medicaid_part_b_buy_in_premium_coverage (sum and count > 0);
• medicare_part_b_premium, medical_expense_health_insurance_premiums;
• spm_unit_health_insurance_premiums, spm_unit_medical_out_of_pocket_expenses, spm_unit_net_income, SPM poverty;
• itemized_medical_expenses, medical_expense_deduction, income_tax;
• id_health_insurance_premiums_subtraction, MT and OH premium deductions, state_income_tax;
• snap, HUD hud_medical_expenses / housing assistance;
• the state LIHEAP/EAP benefits that read the premium;
• household_net_income.

If P2-2's change lands, add medicaid_enrolled and is_medically_needy_for_medicaid, since the Medicaid tests would then read the premium before the buy-in.

What I couldn't check

• Any source beyond the hub's inputs. In particular, I couldn't read 42 U.S.C. 1395r(i), so I can't confirm that the income-related amount stays with a buy-in enrollee. 1395v(d)(1) excludes only the 1395r(b) increase.
• Whether states still use 1395v(b)'s exclusion of Social Security beneficiaries absent a (g) modification. The current 42 CFR 407.42(b)(1)(i) states no such exclusion.
• The 2025 and 2026 standard premiums against CMS.
• The size of the baseline movement on population data. The earlier session's synthetic grid (no buy-in changes among 286 SSI enrollees with Part B) rests on that session's run; I didn't rerun it.

---

Generated by Claude Code

[MaxGhenis]

Picking this up for the US hub. I'll post a hand-off comment here when it's ready.

---

Generated by Claude Code

[MaxGhenis]

Ready for hub review
790a911

Fix round 1 on #10634, answering the round 1 review. The head was be36a245 when I picked this up; no one else pushed in between. Commits: cc9d3654 (fixes), 682cebcf (changelog), 790a9116 (merge of current main).

Items

1. SSI receipt month by month (P2-1). Fixed. medicaid_part_b_buy_in_premium_coverage.py:53 loops over period.get_subperiods(MONTH) and counts a month when ssi > 0 or receives_ssi holds that month and the state test holds. New cases at medicaid_part_b_buy_in_premium_coverage.yaml:76 (December only) and :102 (January only): TX, age 70, $30,000 interest, 2026, receives_ssi entered for all 12 months.
   • By hand, before running: 2026 base 202.90 × 12 = 2,434.80. $30,000 of interest is over the MSP limits, so MSP pays 0, and there is no IRMAA. One month of SSI gives a buy-in of 202.90 and a premium of 2,434.80 − 202.90 = 2,231.90. Both pass.
   • The CA full-year case (:1) still gives 2,220.
   • is_ssi_recipient_for_medicaid and is_209b_ssi_recipient_for_medicaid are untouched.
2. 209(b) buy-in (P2-2). Fixed, using the review's methodology.
   • New medicare_part_b_premium_before_buy_in: gross premium minus MSP, floored at 0, defined_for Part B enrollment. Cites 42 CFR 407.42 and 42 U.S.C. 1395v(d).
   • New medicaid_eligibility_health_insurance_premiums: the same formula as medical_expense_health_insurance_premiums, but reading the premium before the buy-in. Cites 435.121(f) and 435.831.
   • Each new variable has a label and a documentation line saying why it exists.
   • medicaid_medically_needy_medical_expenses.py:25 adds the new premiums variable.
   • medicaid_optional_senior_or_disabled_countable_income.py:86 (paid_premiums) and :88 (paid_part_b) read the two new variables.
   • Buy-in, medicaid_part_b_buy_in_premium_coverage.py:46-47: the review's annual is_ssi_recipient_for_medicaid × form isn't used, as the brief directs. The loop counts a month when the person receives SSI that month and either (is_covered & covered_classification) holds or is_209b_ssi_recipient_for_medicaid (read once, for the year) is true.
   • The formula's 209(b) exclusion comment is gone. The documentation now says 209(b) SSI recipients on Medicaid are covered under 407.42(b)(1)(i), and that other full-benefit groups such as the medically needy stay out.
   • CT spend-down witness (:31, replaces the old CT case). By hand, before running: 2025 base 185 × 12 = 2,220. $30,000 of interest is over the MSP limits, so MSP pays 0 and the premium before the buy-in is 2,220. Countable income is at most 30,000 − 240 = 29,760 before CT's own disregard. Medical expenses are 40,000 + 2,220 = 42,220, so income after spend-down is negative, under any nonnegative limit, and the income test in is_209b_ssi_recipient_income_eligible_for_medicaid passes. There are no assets (limit 1,600). With SSI reported, CT's 209(b) classification and age 70 (no nonblind-child exclusion), is_209b_ssi_recipient_for_medicaid is true. So is_ssi_recipient_for_medicaid is true, the Medicaid category isn't NONE, take-up defaults to true, and medicaid_enrolled is true. Buy-in 12 × 185 = 2,220; premium 0. Passes.
   • CT December-only witness (:49, every month entered, plus the spend-down). By hand: is_209b_ssi_recipient_for_medicaid reads the year's receives_ssi, which is December's value, so it returns true. The loop counts December only: buy-in 185, premium 2,220 − 185 = 2,035. Passes. A CT January-only recipient would get 0 buy-in, because the 209(b) variable reads December. That is the follow-up below.
   • Moved inputs. Each now enters medicare_part_b_premium_before_buy_in, with the same expected value:
     • medicaid_medically_needy_medical_expenses.yaml:15, "SSI-based spenddown retains decomposed non-pretax Medicare premiums."
     • medicaid_optional_senior_or_disabled_countable_income.yaml:591, "California deducted a Part B premium that the person paid before 12/2020."
   • Correction to the review: the CA case that fails is this 2019 case at :591, not the one at line 398 ("…does not count a self-paid Part B premium twice…"). The line-398 case enters only health_insurance_premiums and passes unchanged.
   • No cycle. The CT cases, the whole buy-in file, the CA countable-income file and the sweep below ran with no CycleError.
3. Buy-in at the standard premium. No code change. Added https://www.ecfr.gov/current/title-42/section-408.28 to the buy-in references (:15). The body says 1395v(d)(1) ("without any increase under subsection (b)") and 408.28 set the reading, so the income-related amount stays with the enrollee.
4. Property test (P3-1). test_part_b_msp_offset.py keeps the three bounds. The formula restatement is gone; a comment at :466 points to the YAML overlap case ("MSP and buy-in coverage of the same premium are not added twice."), which M3 fails.
5. Citations (P3-2).
   • The buy-in references are the review's tuple plus 408.28 (medicaid_part_b_buy_in_premium_coverage.py:11-16).
   • medicare_part_b_enrolled.py:17-18 cites …/uscode/text/42/1395o#a and https://www.ecfr.gov/current/title-42/section-407.10.
6. Body (P3-3).
   • "CI has not run" is gone.
   • The 209(b) row now uses the spend-down witness, and there is a new row for the monthly SSI read.
   • The Changes section lists the 209(b) buy-in and both new variables.
   • The impact: line now carries a variable list, ending with medicaid_enrolled and is_medically_needy_for_medicaid.
   • axiom: is left for the hub.
   • Changelog fragment updated (changelog.d/medicare-part-b-payers.fixed.md).

Tests (local, after the merge with main, one file per process, each with its own TMPDIR)

• Changed files:

  • medicaid_part_b_buy_in_premium_coverage.yaml 7/7;
  • medicare_part_b_premium.yaml 19/19;
  • medicare_part_b_enrolled.yaml 4/4;
  • id_health_insurance_premiums_subtraction.yaml 24/24;
  • medicaid_optional_senior_or_disabled_countable_income.yaml 42/42;
  • medicaid_medically_needy_medical_expenses.yaml 3/3;
  • test_part_b_msp_offset.py 15/15.

• Sweep: every YAML file outside partners/ whose text contains any of these names:

  • medicare_part_b_premium (as a substring, so it also matches gross_… and …_before_buy_in);
  • medical_expense_health_insurance_premiums, medicaid_eligibility_health_insurance_premiums;
  • is_209b_ssi_recipient_for_medicaid, is_ssi_recipient_for_medicaid;
  • medicaid_optional_senior_or_disabled_countable_income, medicaid_medically_needy_medical_expenses;
  • medicaid_part_b_buy_in_premium_coverage, medicare_part_b_enrolled.

  That is 105 files: 1,806 passed, 0 failed, at most 4 at once. The review's 104 files gave 1,797 passed and 2 failed; the 2 were the moved inputs, and the extra cases are this round's new ones plus the merge with main.

• The memory cgroup OOM-killed one sweep process and one mutant process after each had printed its pytest summary. All 105 sweep logs end in a passed summary, so no result is missing.

• No microsimulation was run, and no private data was used.

Mutants (each on a scratch git worktree in its own process, run against the 5 key YAML files and test_part_b_msp_offset.py; worktrees removed, git status clean afterwards)

Mutant	Result
M3 sum instead of max (medicare_part_b_premium.py)	killed by medicare_part_b_premium.yaml (2: "MSP and buy-in coverage of the same premium are not added twice.", "MSP coverage preserves IRMAA above the standard Part B premium") and test_medicare_part_b_premium_preserves_only_irmaa_above_msp_support. The property test no longer kills it, as expected once the restatement is gone.
M8 SSI receipt ignored (covered = ssi_medicaid_state)	killed: premium 12, buy-in 4, CA countable income 2, Idaho 1, Python 1 (…does_not_depend_on_calculation_order)
M13 reported receives_ssi ignored	killed: buy-in 5 (including both CT cases and TX December/January), premium 1
Read the year's receives_ssi again inside the loop	killed by buy-in 3: TX December-only, TX January-only, CT December-only
209(b) countable income reads medicare_part_b_premium (after the buy-in)	killed with CycleError: premium 16 failed, buy-in 6, CA countable income 33, Idaho 2, Python 4. Medically needy expenses still passed 3/3, since that file doesn't reach the buy-in.

CI

Pull request run 41737420675 at 790a9116: still in progress at 60 minutes after the push (00:14Z). 35 of 36 jobs had finished, all success; still running: Contrib tests / Full Suite - Contrib (other-shard-1). No job had failed. The hub should read the run's final conclusion before landing.

Judgment calls

• The medicare_part_b_premium_before_buy_in references add 42 U.S.C. 1395v(d) next to 42 CFR 407.42, since (d)(2)(C) is the rule that the buy-in starts only once the person is eligible.
• medicaid_eligibility_health_insurance_premiums lives next to medical_expense_health_insurance_premiums in variables/household/expense/health/.
• medical_expense_health_insurance_premiums still reads the premium after the buy-in, so tax, SNAP, HUD, LIHEAP and SPM readers see what the person pays.
• The merge conflicts were only in files this PR doesn't change: CHANGELOG.md, pyproject.toml, uv.lock, the trustees 2026 reform and its test, and an MO itemized-deductions YAML. I took main's version of each, and git diff origin/main --stat shows only the PR's 19 files.
• I couldn't fetch eCFR or the U.S. Code from this VM to re-check the quotes; the ones in the body and code come from the hub's brief.

Proposed follow-up

• Read SSI receipt month by month in the SSI-recipient Medicaid tests. is_ssi_recipient_for_medicaid and is_209b_ssi_recipient_for_medicaid read the year's receives_ssi, which is December's value. So the SSI route to Medicaid follows December alone, and a January-only 209(b) recipient gets no buy-in month.

---

Generated by Claude Code

[MaxGhenis]

Hub review in progress

Done: the full review is at #10634 (comment).

---

Generated by Claude Code

[MaxGhenis]

Ready for hub review
VERDICT: APPROVE
790a911
METHODOLOGY_CHOICE: yes

Round 2 review of #10634. The head hasn't moved from 790a9116. This round reviews cc9d3654 (fixes), 682cebcf (changelog) and the main merge 790a9116 against the round 1 review at be36a245.

The code does what the hub's rulings ask:

• The buy-in pays the standard premium only, counted month by month while the person receives SSI.
• It now covers 209(b) SSI recipients on Medicaid.
• Medicaid eligibility counts the premium net of MSP but before the buy-in, with no cycle.

All hand values reproduce, every mutant is killed by a named test, and CI is green. Nothing here is a P1.

• P2-1 is a landing-order hazard with open Limit SSI spousal-deeming child allocations to the couple's co-resident children #10146. It isn't a defect at this head.
• P2-2 asks for two corrections to the PR body.
• P3s ask for a stronger 209(b) test and note follow-ups.

The PR picks readings where the law leaves room:

• Part B enrollment defaults to every Medicare enrollee.
• The buy-in pays only the standard premium.
• Medicaid eligibility counts the premium before the buy-in.

The hub ruled on the last two.

Sources. From this VM, ecfr.gov, uscode.house.gov, govinfo.gov and law.cornell.edu all failed: refused at the proxy, or not resolved. I could not re-read 42 CFR 407.40, 407.42, 408.28, 435.121 or 435.831, or 42 U.S.C. 1395v. Everything below rests on the hub's quoted texts in the brief, and I quote no law beyond them. Nothing in the code contradicts those texts.

Findings

P2-1. Landing #10146 after this PR, taking its side of the conflict, brings back the cycle

#10146 ("Limit SSI spousal-deeming child allocations…", open, head 0b6e566f) moves the body of medicaid_optional_senior_or_disabled_countable_income into a new medicaid_optional_senior_or_disabled_countable_income_monthly.py. The new file's lines 98 and 102 read medical_expense_health_insurance_premiums and medicare_part_b_premium, both after the buy-in.

Git flags a conflict only in the annual file. The monthly file is new to #10146, so git merges it silently. A resolver who takes #10146's side gets this chain:

209(b) test → countable income → monthly → medicare_part_b_premium → buy-in → 209(b) test

Witness: merge pr-10146 into this head with git merge -X theirs pr-10146, then run the buy-in YAML:

$ policyengine-core test .../medicaid_part_b_buy_in_premium_coverage.yaml -c policyengine_us
=================== 6 failed, 1 passed ===================
CycleError: Circular definition detected on formula medicaid_part_b_buy_in_premium_coverage@2025

CI on the second PR would catch this, so it's a hazard, not a defect here. Whichever lands second should carry this PR's reads into the monthly file:

        part_b_enrolled = person("medicare_part_b_enrolled", period.this_year)
        paid_premiums = (
            person("medicaid_eligibility_health_insurance_premiums", period.this_year)
            / MONTHS_IN_YEAR
        )
        paid_part_b = (
            person("medicare_part_b_premium_before_buy_in", period.this_year)
            / MONTHS_IN_YEAR
            * part_b_enrolled
        )
        full_part_b = (
            person("gross_medicare_part_b_premium", period.this_year)
            / MONTHS_IN_YEAR
            * part_b_enrolled
        )

The same check against #10141, #10631, #10111, #10532 and #10103 merged into this head passes 7/7.

P2-2. Two body statements are false

1. The monthly SSI row's witness isn't main. The table is headed "Gaps and witnesses (at main 4e71481)". The monthly row says that at main December-only SSI gets "buy-in 2,434.80, premium 0; January only: buy-in 0, premium 2,434.80". Main has no buy-in. Those are round 1's numbers at be36a245. Run at current main (25c42534) through the Python API, TX, age 70, $30,000 interest, 2026, receives_ssi entered for all 12 months:

   TX Dec  medicare_part_b_premium 2434.8  (medicaid_part_b_buy_in_premium_coverage: no such variable)
   TX Jan  medicare_part_b_premium 2434.8
   

   Replace that row's middle cell with:

   TX, $30,000 interest, 2026, SSI in one month (every month entered): premium 2,434.80 for December-only and January-only alike. Main has no buy-in. Round 1 of this PR bought in December-only SSI for all 12 months.

   Replace the row's gap text, "so December-only SSI bought in all 12 months and January-only SSI none", with "so an annual SSI read would buy in December-only SSI for all 12 months and January-only SSI for none."

2. The readers list says Medicaid readers inherit both fixes. "Readers of the modeled premium" opens "All of these … inherit both fixes", and its Benefits bullet lists "Medicaid medically needy medical expenses, Medicaid optional senior/disabled countable income (CA Part B disregard)". Both now read the premium before the buy-in (medicaid_medically_needy_medical_expenses.py:25, medicaid_optional_senior_or_disabled_countable_income.py:86-88). They inherit the Part B enrollment gate, not the buy-in. Delete those two items from the Benefits bullet and add:

   • Medicaid eligibility: medicaid_medically_needy_medical_expenses and medicaid_optional_senior_or_disabled_countable_income read the premium before the buy-in (medicaid_eligibility_health_insurance_premiums, medicare_part_b_premium_before_buy_in). They take the Part B enrollment gate but not the buy-in, which starts only once Medicaid finds the person eligible.

P3-1. The CT spend-down witness doesn't depend on the premium

medicaid_part_b_buy_in_premium_coverage.yaml:31-47 says expenses "of 40,000 plus the 2,220 premium before the buy-in exceed countable income". But 40,000 alone does. The model's numbers for that person:

other_medical_expenses 0:      countable income 23,580, limit 10,016, medically needy expenses 2,220,  is_209b False
other_medical_expenses 40,000: countable income 23,580, limit 10,016, medically needy expenses 42,220, is_209b True

So the case passes whether or not Medicaid eligibility counts the Part B premium. I tested that with a mutant setting the Part B term in medicaid_eligibility_health_insurance_premiums to 0:

• all 7 buy-in cases pass;
• the medically needy expenses file (1 failure) and the CA countable-income file (2) kill it.

So the premium is pinned at the unit level, not at the 209(b) decision. The premium decides eligibility when other expenses fall between 11,344 and 13,564. Suggested added case, which passes at this head and fails under that mutant:

- name: The Part B premium before the buy-in decides a 209(b) spend-down.
  # Countable income 23,580 and limit 10,016. 23,580 - 12,000 = 11,580 is over
  # the limit; 23,580 - 12,000 - 2,220 = 9,360 is under it.
  period: 2025
  input:
    age: 70
    taxable_interest_income: 30_000
    other_medical_expenses: 12_000
    receives_ssi: true
    state_code: CT
  output:
    medicaid_optional_senior_or_disabled_countable_income: 23_580
    medicaid_medically_needy_medical_expenses: 14_220
    is_209b_ssi_recipient_for_medicaid: true
    medicaid_part_b_buy_in_premium_coverage: 2_220
    medicare_part_b_premium: 0

head:    1 passed
mutant:  1 failed

P3-2. The annual SSI read in the Medicaid tests (follow-up, same on main)

The 209(b) branch of the buy-in reads is_209b_ssi_recipient_for_medicaid for the year, and that variable reads add(person, period, ["receives_ssi"]) > 0. Through the API, the year's receives_ssi is December's value: true for December-only, false for January-only. Witnesses at this head (2025, CT, $30,000 interest, $40,000 other medical expenses; and TX in 2026):

Case	is_209b…	medicaid_enrolled	buy-in	premium	premium at main
CT, SSI January only	False	True	0	2,220	2,220
CT, SSI December only	True	True	185	2,035	2,220
TX, SSI January only	n/a	False	202.90	2,231.90	2,434.80

So the CT January-only recipient gets no buy-in month. That's the same result as main, and the brief told the author to leave both variables alone, so it's the hand-off's follow-up, not this PR's defect. The TX January-only row shows the follow-up's other face: the buy-in counts January, but is_ssi_recipient_for_medicaid and medicaid_enrolled are false for the year. The hand-off's proposed follow-up issue covers both.

P3-3. The axiom: line isn't in a final form

The body has axiom: (hub files pe-parity before landing). CONTRIBUTING's forms are encoded-correct | encoded | queued | n/a. The brief left this to the hub, so it's a hub action before landing.

Checked and sound

• CI. Pull request run 41737420675 at 790a9116: completed, success, 36/36 jobs success (none cancelled or skipped), including Household API Partners and all baseline, contrib and Python shards.

• The defect at main (25c42534, Python API):

  • CA full-year reported SSI, $30,000 interest, 2025: premium 2,220.
  • CT spend-down: is_209b… true, medicaid_enrolled true, premium 2,220.
  • TX one-month SSI: premium 2,434.80.

  Main has no medicare_part_b_enrolled. At the head these cases give 0, 0 and 2,231.90.

• Hand values, computed before reading the code's output. base_part_b_premium is 2,220 in 2025 and 2,434.80 in 2026.

  • TX one month: 2,434.80 / 12 = 202.90 of buy-in; 2,434.80 − 202.90 = 2,231.90.
  • CT December only: 2,220 / 12 = 185; 2,220 − 185 = 2,035.
  • CT full year: 2,220 and 0. CA full year: 2,220.

  The API gives exactly these: 202.90001 and 2,231.9001 in float32, and 185, 2,035, 2,220 and 0. MSP is 0 in each case.

• The rules in the code (medicaid_part_b_buy_in_premium_coverage.py:34-57):

  • Each month counts when ssi > 0 or receives_ssi holds that month, and either is_covered & (SECTION_1634 | SSI_CRITERIA) or the 209(b) test is true.
  • It pays base_part_b_premium / 12, the standard premium without IRMAA, which matches 1395v(d)(1) as the hub quotes it.
  • The premium nets max(MSP, buy-in) and floors at 0 (medicare_part_b_premium.py:21-27).
  • Medicaid eligibility reads the new medicare_part_b_premium_before_buy_in (gross − MSP, floored, defined_for Part B) and medicaid_eligibility_health_insurance_premiums. The latter is line for line medical_expense_health_insurance_premiums except for that one read.
  • Both new variables have a label, a documentation line that says why they exist, and the references the brief asked for.
  • The buy-in references are the round 1 tuple plus 408.28.
  • Entity and period: all are Person/YEAR. ssi and receives_ssi are MONTH, read per month in the loop.
  • Filing status doesn't enter; the 209(b) income test sums over the marital unit, unchanged.

• Baseline movement in Medicaid eligibility. With medicare_part_b_enrolled defaulting to medicare_enrolled, the premium before the buy-in equals main's medicare_part_b_premium (gross − MSP, floored). So the 209(b), optional senior/disabled and medically needy tests compute the same values as main at baseline. For all eight hand cases, is_209b…, is_ssi_recipient_for_medicaid and medicaid_enrolled match between main and the head. One indirect path can still move medicaid_enrolled:

  is_medicaid_eligible → medicaid_work_requirement_eligible → medicaid_community_engagement_pass_through_eligible → snap → snap_allowable_medical_expenses → medical_expense_health_insurance_premiums → medicare_part_b_premium

  SNAP's medical deduction falls by the buy-in, so a work-requirement adult on SNAP pass-through with a buy-in could move. Such a person would need SSI and be under 65, which should be rare. Keep medicaid_enrolled in the impact list, as the body does.

• No cycle.

  • The hand-off's mutant (paid_part_b reads medicare_part_b_premium) gives CycleError … medicaid_part_b_buy_in_premium_coverage@2025: premium 16 failed, buy-in 6, CA countable income 33, Idaho 2, Python 4. Medically needy expenses passed 3/3. That matches the hand-off.
  • My mutant on the other arm (medicaid_medically_needy_medical_expenses adds medical_expense_health_insurance_premiums again) also gives CycleError: buy-in 6, premium 16, medically needy 1, CA 4, Idaho 2, Python 4. So both arms need the before-buy-in premium.
  • A static scan of every formula's quoted variable names, plus adds/subtracts lists including parameter lists, finds 1,563 variables that depend on medicare_part_b_premium, directly or through others, and 328 that the buy-in reads, with no overlap.

• Other Medicaid paths to medicare_part_b_premium. Yes, two, both on main:

  • is_medicaid_eligible/medicaid_enrolled → the work-requirement SNAP pass-through (path above);
  • medicaid_category → ms_hmw_eligible → is_working_disabled_buy_in_for_medicaid → il_hbwd_… → childcare_expenses → pa_ccw → pa_ccw_medical_expenses → medical_expense_health_insurance_premiums.

  Neither is among the buy-in's inputs, so neither closes a cycle today. They would if the SSI or 209(b) tests ever read medicaid_category or is_medicaid_eligible.

• The moved inputs.

  • medicaid_medically_needy_medical_expenses.yaml:11-21 still tests "retains decomposed non-pretax Medicare premiums": 600 + 1,200 + 300 = 2,100, with the 2,400 pretax excluded.
  • The CA 2019 case at medicaid_optional_senior_or_disabled_countable_income.yaml:584-601 still tests a person-paid Part B before 12/2020: (1,450 − 20 − 230 − 100 − 135.50) × 12 = 11,574.

  Putting both inputs back on medicare_part_b_premium fails exactly these two cases and no others in the six key files.

• The correction. At be36a245, line 398 is "California does not count a self-paid Part B premium twice…", which enters only health_insurance_premiums and asserts medicare_part_b_premium as an output. With the input put back, only the 2019 case fails in that file. The hand-off is right: the round 1 review named the wrong case.

• Changed files at the head (one per process, own TMPDIR):

  • buy-in 7/7, premium 19/19, enrolled 4/4, Idaho 24/24;
  • CA countable income 42/42, medically needy expenses 3/3;
  • test_part_b_msp_offset.py 15/15.

  Case counts against main: premium 13 → 19, Idaho 23 → 24, CA 41 → 42. These match the body.

• The sweep. The hand-off's grep, re-run here, gives 105 files. Five spot-checked:

  • is_209b_ssi_recipient_for_medicaid.yaml 18/18;
  • is_ssi_recipient_for_medicaid.yaml 9/9;
  • is_medically_needy_for_medicaid.yaml 37/37;
  • hud_medical_expenses.yaml 3/3;
  • itemized_medical_expenses.yaml 9/9.

• Mutants (each on a scratch worktree, run against the buy-in, premium, CA, medically needy and Idaho YAML and test_part_b_msp_offset.py):

  Mutant	Killed by
  M3 sum instead of max	premium YAML 2, Python 1 (test_medicare_part_b_premium_preserves_only_irmaa_above_msp_support); the property test no longer kills it, as the hand-off says
  M8 SSI receipt ignored	buy-in 4, premium 12, CA 2, Idaho 1, Python 1
  M13 reported receives_ssi ignored	buy-in 5 (CA full year, both CT, TX December and January), premium 1
  Year's receives_ssi read in the loop	buy-in 3: CT December-only, TX December-only, TX January-only
  209(b) countable income reads the after-buy-in premium	CycleError (counts above)
  (mine) 209(b) branch dropped	buy-in 2: both CT cases
  (mine) medically needy arm reads the after-buy-in premium	CycleError (counts above)
  (mine) Medicaid premiums drop Part B	medically needy 1, CA 2; survives the buy-in file (P3-1)

  All worktrees were removed, and git status is clean.

• The property test (test_part_b_msp_offset.py:422-466) keeps the three bounds and no longer restates the formula. It runs one 200-person simulation.

• Readers.

  • medicare_part_b_premium: 25 variables, including SPM premiums, the premium aggregate, NH FAP and 22 LIHEAP/EAP income definitions.
  • medical_expense_health_insurance_premiums: 12, including HUD, Schedule A and its excluded premiums, the self-employed ALD exclusion, SNAP, PA CCW, UT HEAT, OH, ID and three MT readers.
  • medicare_part_b_premium_before_buy_in and medicaid_eligibility_health_insurance_premiums: only the two Medicaid variables.
  • medicaid_part_b_buy_in_premium_coverage: only the premium.
  • gross_medicare_part_b_premium_if_enrolled: only medicare_cost.

  Every after-buy-in reader sees a premium that is equal or lower, and lower only for buy-in months, so none is worse than main except as P3-2's follow-up describes.

• Docs. Each added documentation line and comment matches the code read here: the buy-in, premium, before-buy-in and Medicaid-premiums docs, the countable-income comment, and the WA and OK comments. The changelog fragment is a top-level changelog.d/medicare-part-b-payers.fixed.md.

• Scope. Nothing changes under .github/ or the partner tests.

  • Open PRs naming the changed variables: Add California LIHEAP regular heating assistance #10607, Project Medicare premiums and IRMAA past the last announced year #10103, Add Idaho LIHEAP regular heating assistance #10613, Add Connecticut CEAP regular heating assistance #10611, Add Delaware LIHEAP regular heating assistance #10612, Count only filer-paid premiums in Schedule A medical expenses #10270, Apply 416.1163(e)(2) to deemed SSI and identify SSI couples by marriage, not tax roles #10141, Count eligible long-term care premiums in Schedule A medical expenses #10265, Exclude employer-plan-eligible premiums from the self-employed health insurance deduction #10122, Exclude people otherwise eligible for Medicaid from Illinois HBWD and use Illinois's income rules for its disability pathway #10631, Limit SSI spousal-deeming child allocations to the couple's co-resident children #10146, Include farm and partnership earnings in SSI and correct loss netting #10532, Subtract the expected Part D premium from medicare_cost #10111.
  • git merge-tree against this head is clean for all but three: Project Medicare premiums and IRMAA past the last announced year #10103 (conflict in medicare_part_b_premium.yaml), Count only filer-paid premiums in Schedule A medical expenses #10270 (conflict in test_medical_expense_premium_conservation.py) and Limit SSI spousal-deeming child allocations to the couple's co-resident children #10146 (P2-1).

• Impact. The impact: line exists and lists the buy-in, premium, aggregates, SPM, tax, SNAP, HUD, LIHEAP, medicaid_enrolled and is_medically_needy_for_medicaid. The diff moves outputs at baseline through the buy-in alone. Suggested population checks:

  • buy-in sum and count, split by 1634/SSI-criteria versus 209(b) states;
  • the count of people with buy-in in fewer than 12 months;
  • that is_medically_needy_for_medicaid doesn't change, and that any medicaid_enrolled change traces to the SNAP work-requirement path.

What I couldn't check

• Any statute or regulation text: eCFR, the U.S. Code sites, govinfo and Cornell's LII were all unreachable from this VM. The rulings rest on the hub's quotes.
• The 2025 and 2026 standard premiums against CMS.
• Baseline movement on population data, including how many SSI recipients gain a buy-in and how many 209(b) recipients are newly covered.
• The full 105-file sweep: I ran 5 of its files plus the 7 changed files, and leave the rest to CI, which passed.

---

Generated by Claude Code

[MaxGhenis]

Picking this up for the US hub. I'll post a hand-off comment here when it's ready.

---

Generated by Claude Code

[MaxGhenis]

Ready for hub review

Head: aa46e20cd18adc1334fe83ee4e221e9c254b1fd6

Built on 790a911670, the head the round 2 review approved. No one else pushed in between. I didn't merge main: git merge-tree against origin/main (25c4253483) is clean, and no check needed it.

Items

Item 14 (P3-1), the 209(b) spend-down case. medicaid_part_b_buy_in_premium_coverage.yaml:49-64 (commit aa46e20cd1):

• The review's case, as given. I put it right after the existing Connecticut spend-down case so the 209(b) cases sit together.
• At this head: 8 passed.
• The review's mutant sets the Part B term to 0 (medicaid_eligibility_health_insurance_premiums.py:29-31 becomes medicare_part_b = 0). It fails 1 and passes 7: only the new case fails, at its first output, medicaid_medically_needy_medical_expenses@2025: [12000.] differs from 14220.0.
• A probe of the same household under the mutant shows the 209(b) result itself flips, not just the expense line:

Output	This head	Mutant
medicaid_optional_senior_or_disabled_countable_income	23,580	23,580
medicaid_medically_needy_medical_expenses	14,220	12,000
is_209b_ssi_recipient_for_medicaid	true	false
medicaid_part_b_buy_in_premium_coverage	2,220	0
medicare_part_b_premium	0	2,220

Item 15 (P2-2), the two false statements.

• The month-by-month SSI row. I replaced the middle cell and the gap text with the brief's wording. Before keeping "since main has no buy-in", I ran the witness on main 25c4253483, in a worktree with the import path checked: TX, age 70, $30,000 interest, 2026, receives_ssi entered for all 12 months.
  • medicare_part_b_premium is 2,434.80 for December-only and January-only SSI alike.
  • medicaid_part_b_buy_in_premium_coverage raises "does not exist".
  • The table's heading names main 4e71481e. No commit since then touches variables/gov/hhs/medicare, parameters/gov/hhs/medicare or variables/gov/ssa/ssi.
• The readers list. I deleted the two Medicaid items from the Benefits bullet and added the brief's paragraph after the list. I checked it against the code:
  • medicaid_medically_needy_medical_expenses adds medicaid_eligibility_health_insurance_premiums.
  • medicaid_optional_senior_or_disabled_countable_income.py:86-88 reads that and medicare_part_b_premium_before_buy_in.
  • medicare_part_b_premium_before_buy_in is defined_for Part B enrollment, and medicaid_eligibility_health_insurance_premiums multiplies by medicare_part_b_enrolled.

Item 16 (P2-1), the overlap with #10146. I added "Overlaps with open PRs" with the brief's bullet. It sits just above the impact: line, which comes directly before axiom:, so the two lines stay together. I checked each claim:

• Limit SSI spousal-deeming child allocations to the couple's co-resident children #10146 is open at 0b6e566ff4. It adds medicaid_optional_senior_or_disabled_countable_income_monthly.py, which reads medical_expense_health_insurance_premiums (line 98) and medicare_part_b_premium (line 102), both after the buy-in. Its annual file now just calls the monthly one.
• git merge-tree of 790a911670 with Limit SSI spousal-deeming child allocations to the couple's co-resident children #10146's head flags a conflict only in the annual file.
• In a scratch worktree, git merge -X theirs of Limit SSI spousal-deeming child allocations to the couple's co-resident children #10146 into 790a911670, then the buy-in YAML: 6 failed, 1 passed. The error is CycleError: Circular definition detected on formula medicaid_part_b_buy_in_premium_coverage@2025, with the import path checked.

Test counts in the body.

• The Tests bullet now says the buy-in file has 8 cases and names the new one.
• The "Local results" sentence now says its counts are at 790a911670, and adds this head's four runs.
• I left the axiom: and impact: lines alone, and added this session's link after the one already there. The readback matches what I sent.

Item 17, runs at this head, one file per process:

File	Result
medicaid_part_b_buy_in_premium_coverage.yaml	8 passed
medicare_part_b_premium.yaml	19 passed
medicare_part_b_enrolled.yaml	4 passed
medicaid_medically_needy_medical_expenses.yaml	3 passed

Hand computations

The new case:

• Countable income. Connecticut's 2025 individual disregard is $535 a month (senior_or_disabled/income/disregard/individual.yaml), and Connecticut deducts no premiums from countable income. So countable income is 30,000 − 535 × 12 = 23,580.
• Medically needy expenses. 12,000 + 2,220 = 14,220.
• The spend-down. 23,580 − 12,000 = 11,580, over the 10,016 limit. 23,580 − 14,220 = 9,360, under it.
• The buy-in. 185 × 12 = 2,220, which leaves a premium of 0.

I took the 10,016 limit from the review. No case asserts it, and I didn't trace its parameter.

CI

Pull request run 41868367787 at aa46e20cd1 had no conclusion 60 minutes after the push (16:32Z). At 17:38Z the run still read in_progress.

• 32 of 36 jobs had concluded success, including Household API Partners, Quick Feedback, PR validation, and every Baseline and Rest Python shard.
• The other four still read in_progress at the job level, but every step of each had completed, with "Complete job" success:
  • Full Suite - Microsimulation, at 17:25:16Z;
  • Contrib (states-shard-1), at 17:26:42Z;
  • Contrib (other-shard-1), at 17:30:12Z;
  • Contrib (other-shard-3), at 17:35:37Z.
• In states-shard-1 and other-shard-3 the "Run Contrib Python tests" step was skipped, as it was in Reduce Colorado's federal deduction addback by the state income tax already added back #10094's runs.
• No job or step had failed or been cancelled.

Only the job and run records lag. #10094's run showed the same pattern: its run record concluded success while two of its job records still read in_progress. The hub should confirm the run's final conclusion.

Judgment calls

• Where the new section sits: above impact:, for the reason above.
• Where the new case sits: after the existing Connecticut spend-down case.
• The "Local results" sentence now says which head its counts come from.
• Ruff. This round changed no Python file. The locked ruff 0.15.5 passes on the tree.

Follow-ups

None new. The review's P3-2 follow-up on the annual SSI read in the Medicaid tests stays with the hub.

---

Generated by Claude Code

[MaxGhenis]

Hub review in progress

The final review is posted: #10634 (comment). It supersedes this interim note.

---

Generated by Claude Code

[MaxGhenis]

Ready for hub review
VERDICT: APPROVE
aa46e20
METHODOLOGY_CHOICE: yes

Round 3 delta review of #10634. The head is still aa46e20cd1, the head in the hand-off, so nothing moved after it.

What changed since 790a911670. The code diff is one YAML case, the round 2 review's P3-1 case, at medicaid_part_b_buy_in_premium_coverage.yaml:49-64. No variable, parameter or Python file changed. The body changed in three places: the month-by-month SSI row, the readers paragraph, and a new "Overlaps with open PRs" section. I reran all three round 2 items against the code.

Verdict. The new case asserts what it claims and kills the review's mutant. The two corrected statements match the code at main and at the head. CI is green. Nothing here is a P1.

• P2-1: the new Limit SSI spousal-deeming child allocations to the couple's co-resident children #10146 bullet is incomplete. Following it literally clears the cycle but breaks a test.
• P3-1: a new open PR, Add 2015-2023 Medicare IRMAA brackets and 2015-2020 Part B premiums #10112, conflicts with this PR in a comment.
• P3-2: no test pins the spend-down at exactly the limit.

METHODOLOGY_CHOICE: yes carries over from round 2:

• Part B enrollment defaults to Medicare enrollment.
• The buy-in pays only the standard premium.
• Medicaid eligibility counts the premium before the buy-in.

This round adds no new reading.

Sources. From this VM, ecfr.gov and aspe.hhs.gov failed at DNS. ecfr.gov, govinfo.gov, uscode.house.gov, law.cornell.edu, aspe.hhs.gov and portal.ct.gov all returned 403 at the proxy's CONNECT. I could read no statute, regulation, poverty guideline or Connecticut source, and I quote none. The values below are traced to the repository's parameters, not to the law.

Findings

P2-1. The #10146 bullet leaves out the Part B enrollment gate, so the merge it describes still fails a test

The body's "Overlaps with open PRs" bullet tells whoever lands second to carry two reads into #10146's new monthly file: medicaid_eligibility_health_insurance_premiums and medicare_part_b_premium_before_buy_in. That file (#10146 at 0b6e566ff4, medicaid_optional_senior_or_disabled_countable_income_monthly.py:96-108) also gates both Part B terms on person("medicare_enrolled", period.this_year). Its comment says "medicare_enrolled (Part A and/or Part B) stands in for Part B enrollment". This PR's annual file gates on medicare_part_b_enrolled instead.

gross_medicare_part_b_premium is defined_for = "is_medicare_eligible", not Part B enrollment. So if the resolver keeps medicare_enrolled, a Part A-only California enrollee still gets the full Part B premium disregarded.

Failure scenario. I merged #10146 into the head with git merge -X theirs pr-10146, which reproduces the cycle (below). Then I made exactly the two substitutions the bullet names in the monthly file and ran the files one per process:

medicaid_part_b_buy_in_premium_coverage.yaml:               8 passed
medicaid_optional_senior_or_disabled_countable_income.yaml: 1 failed, 41 passed
  Test 'California deducts no Part B premium for a Part A-only Medicare enrollee.':
  medicaid_optional_senior_or_disabled_countable_income@2026: [9325.2] differs from 11760.0 with an absolute margin [2434.7998]

Then I also replaced person("medicare_enrolled", period.this_year) with person("medicare_part_b_enrolled", period.this_year):

medicaid_optional_senior_or_disabled_countable_income.yaml: 42 passed

CI on the second PR would catch this, because this PR's Part A-only case would fail. But the bullet is the hand-off a resolver will follow. The round 2 review's code block had the gate. The body's bullet dropped it.

Replacement for the bullet in "Overlaps with open PRs":

• Limit SSI spousal-deeming child allocations to the couple's co-resident children #10146 moves medicaid_optional_senior_or_disabled_countable_income's formula into a new monthly file. That file reads medicare_part_b_premium after the buy-in and gates Part B on medicare_enrolled. Git flags a conflict only in the annual file. Whoever lands second must carry all three of this PR's reads into the monthly file: medicaid_eligibility_health_insurance_premiums, medicare_part_b_premium_before_buy_in, and the medicare_part_b_enrolled gate on paid_part_b and full_part_b. Without the two premium reads, the buy-in and the 209(b) test read each other, and the simulation stops with a cycle error. Without the gate, a Part A-only California enrollee gets the full Part B premium disregarded.

That resolution should also drop the monthly file's "medicare_enrolled … stands in for Part B enrollment" note, which would then be false.

P3-1. #10112 now conflicts with this PR, and the body doesn't list it

#10112 ("Add 2015-2023 Medicare IRMAA brackets and 2015-2020 Part B premiums", open at eaf0867dda) deletes the countable-income comment "NOTE: gov.hhs.medicare.part_b.base_premium starts in 2021…". This PR rewrote the lines next to it.

$ git merge-tree --write-tree --name-only pr-10634 pr-10112
CONFLICT (content): Merge conflict in policyengine_us/variables/gov/hhs/medicaid/income/medicaid_optional_senior_or_disabled_countable_income.py

CHANGELOG.md, pyproject.toml and uv.lock also conflict, because this branch's base is older than main. #10112 merges cleanly with origin/main.

The resolution is comment-only:

• keep this PR's "Only a Part B enrollee (medicare_part_b_enrolled)…" lines;
• drop the 2021 note, as Add 2015-2023 Medicare IRMAA brackets and 2015-2020 Part B premiums #10112 does;
• drop Add 2015-2023 Medicare IRMAA brackets and 2015-2020 Part B premiums #10112's "medicare_enrolled … stands in" note, as this PR does.

With that resolution, the merged tree passes:

• buy-in 8/8;
• premium 19/19;
• CA countable income 42/42;
• medically needy expenses 3/3.

Add to "Overlaps with open PRs":

• Add 2015-2023 Medicare IRMAA brackets and 2015-2020 Part B premiums #10112 deletes the countable-income note that the Part B base premium starts in 2021, next to lines this PR rewrote. The conflict is in comments only: keep this PR's Part B enrollment comment and drop both NOTE lines.

P3-2. No test pins the 209(b) spend-down at exactly the limit (same on main)

The spend-down test in is_209b_ssi_recipient_income_eligible_for_medicaid.py:40 is income_after_spenddown <= income_limit, and this PR doesn't change that file. My mutant replaces <= with <. All five files survive it:

• buy-in 8/8;
• medically needy 3/3;
• CA countable income 42/42;
• is_209b_ssi_recipient_for_medicaid.yaml 18/18;
• premium 19/19.

The new case sits at 9,360, well inside the window, so it can't tell the two apart. Optional case for the buy-in file, after line 64. It passes at the head and fails under the mutant (is_209b_ssi_recipient_for_medicaid@2025: [0.] differs from 1.0):

- name: A 209(b) spend-down that lands exactly on the limit qualifies.
  # Countable income 23,580 and limit 0.64 x 15,650 = 10,016.
  # 23,580 - 11,344 - 2,220 = 10,016, at the limit; 11,343 would leave 10,017.
  period: 2025
  input:
    age: 70
    taxable_interest_income: 30_000
    other_medical_expenses: 11_344
    receives_ssi: true
    state_code: CT
  output:
    medicaid_medically_needy_medical_expenses: 13_564
    is_209b_ssi_recipient_for_medicaid: true
    medicaid_part_b_buy_in_premium_coverage: 2_220
    medicare_part_b_premium: 0

I couldn't read 42 CFR 435.121 to confirm that "at the limit" qualifies, so the case pins the code's reading, not the regulation's.

Carried from round 2 (hub actions, unchanged)

• The axiom: line still reads axiom: (hub files pe-parity before landing). That isn't one of CONTRIBUTING's forms.
• The follow-up on the annual SSI read in the Medicaid tests (round 2 P3-2) still stands. It's the same on main.

Checked and sound

• CI. Pull request run 41868367787 at aa46e20cd1 is completed / success. All 36 of 36 jobs are completed / success, none cancelled or skipped at the job level. That includes the four jobs the hand-off saw still in_progress: Microsimulation, Contrib states-shard-1, other-shard-1 and other-shard-3. The job records have caught up.

• The defect at main (25c42534, Python API, import path checked). For the new case's household (CT, 70, $30,000 interest, $12,000 other medical expenses, reported SSI, 2025), main gives:

  • countable income 23,580, limit 10,016, medically needy expenses 14,220;
  • is_209b_ssi_recipient_for_medicaid True, medicaid_enrolled True;
  • medicare_part_b_premium 2,220, and medicaid_part_b_buy_in_premium_coverage doesn't exist.

  At the head: buy-in 2,220 and premium 0.

• The 10,016 limit, traced to its parameters. For CT, medicaid_optional_senior_or_disabled_income_limit takes the default branch, limit_pct * unit_fpg:

  • uses_ssi_federal_benefit_rate CT false;
  • covers_non_ssi_recipients CT true;
  • uses_state_standard CT false;
  • not MO.

  The rate is senior_or_disabled/income/limit/individual.yaml:37, CT 2025-01-01: 0.64. The guideline is gov/hhs/fpg.yaml:16, first person, contiguous US, 2025-01-01: 15_650. 0.64 × 15,650 = 10,016, and the model gives 10,016.0. I could not check the 0.64 rate or the 15,650 guideline against CT's guide or HHS (unreachable).

• The other hand values, computed before reading the model's outputs:

  • Countable income. CT uses _apply_ct_husky_c_exclusions: (30,000/12 − 535) × 12 = 1,965 × 12 = 23,580. The disregard is CT 2025-01-01: 535 in disregard/individual.yaml:59, and CT deducts_health_insurance_premiums is false.
  • Medically needy expenses. 12,000 + 2,220 = 14,220.
  • Spend-down. 23,580 − 12,000 = 11,580 > 10,016. 23,580 − 14,220 = 9,360 ≤ 10,016.
  • Buy-in. 185 × 12 = 2,220, so the premium is 0.

  The model gives exactly these.

• The window where the premium decides (head, Python API). With other medical expenses x:

  x	is_209b…	buy-in	premium
  11,343	False	0	2,220
  11,344	True	2,220	0
  12,000	True	2,220	0
  13,563	True	2,220	0
  13,564	True	2,220	0

  So the premium decides for 11,344 ≤ x < 13,564. At 13,564, 23,580 − 13,564 = 10,016 qualifies without the premium. The new case's 12,000 sits inside.

• The new case asserts the claim. It pins countable income, medically needy expenses, the 209(b) decision, the buy-in and the premium, so it asserts the decision, not only the expense line.

• Mutants (scratch worktree, one file per process, restored after each; git status is clean in every worktree):

  Mutant	Buy-in	Medically needy	CA countable	is_209b…	Premium
  Review's: Part B term in medicaid_eligibility_health_insurance_premiums × 0	1 failed (new case, medicaid_medically_needy_medical_expenses@2025: [12000.] differs from 14220.0)	1 failed	2 failed	18 passed	19 passed
  Mine: that term × 0.5	1 failed (new case, [13110.] differs from 14220.0)	1 failed	2 failed	18 passed	19 passed
  Mine: spend-down <= → <	8 passed	3 passed	42 passed	18 passed	19 passed (P3-2)

  Round 2's mutant survived the buy-in file. Now the buy-in file kills it, through the new case alone, which matches the hand-off.

• The corrected SSI row. TX, 70, $30,000 interest, 2026, receives_ssi entered for all 12 months with one month true:

  	Main 25c42534	Head
  December only	premium 2,434.8, buy-in doesn't exist	buy-in 202.9, premium 2,231.9
  January only	premium 2,434.8, buy-in doesn't exist	buy-in 202.9, premium 2,231.9

  The math: 2,434.80 / 12 = 202.90, and 2,434.80 − 202.90 = 2,231.90. The year's receives_ssi is True for December-only and False for January-only, so the gap cell's "an annual SSI read would buy in December-only SSI for all 12 months and January-only SSI for none" is right. The middle and gap cells match the brief's text exactly.

• The readers paragraph.

  • medicare_part_b_premium_before_buy_in has two readers: medicaid_eligibility_health_insurance_premiums and medicaid_optional_senior_or_disabled_countable_income.
  • medicaid_eligibility_health_insurance_premiums has two readers: medicaid_medically_needy_medical_expenses (line 25) and the countable-income file.
  • Both take the Part B gate:
    • the before-buy-in premium is defined_for Part B enrollment;
    • medicaid_eligibility_health_insurance_premiums.py:28-31 multiplies by medicare_part_b_enrolled;
    • the countable-income file multiplies paid_part_b and full_part_b by it.
  • The two Medicaid items are gone from the Benefits bullet.

  The paragraph is true.

• The Limit SSI spousal-deeming child allocations to the couple's co-resident children #10146 cycle. Limit SSI spousal-deeming child allocations to the couple's co-resident children #10146 is open, still at 0b6e566ff4. git merge-tree of the head with it flags only the annual countable-income file. With -X theirs, the buy-in file gives 7 failed and 1 passed (8 cases now). The errors are CycleError: Circular definition detected on formula for medicaid_part_b_buy_in_premium_coverage@2025 and @2026, is_209b_ssi_recipient_for_medicaid@2025 and medicaid_optional_senior_or_disabled_countable_income@2025. That confirms the hand-off's finding. Whichever PR lands second must make the P2-1 changes.

• Changed and related files at the head, one per process:

  • buy-in 8/8, premium 19/19, enrolled 4/4, medically needy expenses 3/3;
  • CA countable income 42/42, is_209b_ssi_recipient_for_medicaid.yaml 18/18;
  • test_part_b_msp_offset.py 15 passed.

  These match the body's counts at aa46e20cd1.

• Brief items 14-18.

  • Item 14: the case is added exactly as given.
  • Item 15: both replacements and the paragraph are verbatim.
  • Item 16: the section sits above impact:, which sits directly above axiom:. The bullet is verbatim, but see P2-1.
  • Test counts: the buy-in file is listed at 8, and the "Local results" sentence names its heads.
  • Item 18: no Python file changed this round, so there's no ruff run to report.

  Nothing the brief asked for is missing.

• Baseline movement. This round adds only a test, so it can't move any output. The PR as a whole moves outputs through the buy-in, as round 2 described. The population checks round 2 proposed still apply:

  • buy-in sum and count, split between 1634/SSI-criteria and 209(b) states;
  • the count with fewer than 12 buy-in months;
  • is_medically_needy_for_medicaid unchanged;
  • any medicaid_enrolled change traced to the SNAP work-requirement path.

• Scope. The PR's 19 files include nothing under .github/ or the partner tests. Searching open PRs for each new variable name finds only Charge the Medicare Part B premium only to people who pay it #10634. Searching for medicaid_optional_senior_or_disabled_countable_income finds Apply 416.1163(e)(2) to deemed SSI and identify SSI couples by marriage, not tax roles #10141, Exclude people otherwise eligible for Medicaid from Illinois HBWD and use Illinois's income rules for its disability pathway #10631, Add 2015-2023 Medicare IRMAA brackets and 2015-2020 Part B premiums #10112, Limit SSI spousal-deeming child allocations to the couple's co-resident children #10146 and Include farm and partnership earnings in SSI and correct loss netting #10532. Add 2015-2023 Medicare IRMAA brackets and 2015-2020 Part B premiums #10112 is new since round 2 (P3-1). git merge-tree of the head with origin/main (25c42534) is clean.

What I couldn't check

• Any law or agency source: 42 U.S.C. 1395v, 42 CFR 407.42, 408.28 and 435.121, the 2025 HHS poverty guideline, and Connecticut's 2025 HUSKY C limit and disregard. ecfr.gov and aspe.hhs.gov failed DNS from this VM. govinfo.gov, uscode.house.gov, law.cornell.edu and portal.ct.gov were refused at the proxy. Whether a spend-down landing exactly on the limit qualifies (P3-2) rests on the code.
• Baseline movement on population data.
• The full 105-file sweep the body cites. I ran the 7 files above at the head, 4 on the Add 2015-2023 Medicare IRMAA brackets and 2015-2020 Part B premiums #10112 merge and 2 on the Limit SSI spousal-deeming child allocations to the couple's co-resident children #10146 merge. CI covers the rest, and it passed.

---

Generated by Claude Code

[DTrim99]

PR Review: Charge the Medicare Part B premium only to people who pay it

Summary

The PR makes medicare_part_b_premium a net-of-state-payment amount for Part B enrollees only. It adds a medicare_part_b_enrolled input, which defaults to medicare_enrolled. It also adds a month-by-month Medicaid Part B buy-in for SSI recipients in Section 1634 and SSI-criteria states, and for 209(b) SSI recipients who qualify for Medicaid. Medicaid eligibility tests keep reading the premium before the buy-in, which avoids a cycle.

I checked the law first. The approach matches it:

• Who owes the premium. Every Part B enrollee owes it (42 U.S.C. 1395r). Part B is a separate enrollment that a Part A enrollee can decline (42 U.S.C. 1395o(a); 42 CFR 407.10).
• The buy-in. All states have Part B buy-in agreements (42 U.S.C. 1395v). In a 1634 state, "CMS enrolls SSI beneficiaries in Medicare Part B, on behalf of the State, with the State paying the beneficiary's Part B premiums" (42 CFR 407.40(c)(5)).
• 209(b) states. Buy-in group 1 covers "Individuals who receive SSI or SSP or both and are covered under the State's Medicaid state plan as categorically needy" (42 CFR 407.42(b)(1)(i)). That text has no 209(b) carve-out.
• Timing. Coverage starts and stops by month (1395v(d)(2)(C), (d)(3)(A); 42 CFR 407.47(b)).
• Amount. The State pays the standard premium. Late-enrollment increases are not charged (1395v(d)(1); 407.40(c)(6)), and IRMAA stays with the beneficiary (42 CFR 408.28).

The readers. I grepped every reader. None nets the buy-in or MSP twice, and none adds back an amount it shouldn't, except the two Medicaid eligibility readers, which deliberately read the premium before the buy-in. Every reader that wants "the premium the person pays" or "the premium withheld from the check" now gets the right amount. That covers the LIHEAP/EAP income definitions, SNAP, HUD, the federal Schedule A, the state premium deductions and SPM MOOP. No partner test file changed, and the Household API Partners job passes.

Open findings. Three should-address items:

• the axiom: line is not a valid form, and no rulespec-us issue covers this change;
• the modeled-SSI half of the buy-in's SSI test is unpinned;
• the annual max_() of MSP and buy-in coverage undercounts what the state pays when the two cover different months.

There are no critical findings. CI is 36/36 green.

Critical

None.

Should address

1. The axiom: line is not a valid form, and no rulespec-us issue covers this change.

• What it says. The PR body reads axiom: (hub files pe-parity before landing). That is none of the CONTRIBUTING.md#axiom-parity forms (encoded-correct, encoded, queued, n/a: <reason>). Kansas Limit Kansas itemized deductions to the four K.S.A. 79-32,120(a) categories at each year's shares #10095 and Vermont Phase out Vermont's child tax credit per child #10644 drew the same finding for the same line.
• Search. I searched TheAxiomFoundation/rulespec-us (all states) for "Part B", "buy-in", "1395v", "407.42", "Medicare premium", "medicare_part_b" and "10634". The nearest open pe-parity issues are Refactor SNAP emergency allotment to decompose accurately in the app #1677 (IRMAA, for Add 2015-2023 Medicare IRMAA brackets and 2015-2020 Part B premiums #10112), Fix inaccurate NY supplemental tax calculations #1679 (Part A premiums), Virginia income tax rate #2009 (Medicare/MSP immigration tests, for Apply immigration tests to Medicare and Medicare Savings Programs #10632) and NJ CDCC #1933 (optional ABD Medicaid disregards). None covers Part B enrollment or the SSI buy-in.
• Fix. File a dispatch-ready, pe-parity-labelled rulespec-us issue and change the line to axiom: <issue> queued. The issue needs:
  • the module path;
  • the corpus citations (42 U.S.C. 1395o(a), 1395v(a), (b), (d)(1)-(3); 42 CFR 407.40(c)(5)-(6), 407.42(b)(1)(i), 407.47(b), 408.28);
  • the verbatim law;
  • the required outputs (Part B premium paid, buy-in coverage, Part B enrollment);
  • companion tests from the same external source as the YAML.

2. The buy-in's modeled-SSI test is unpinned (medicaid_part_b_buy_in_premium_coverage.py:54).

• The term. The buy-in month test is (person("ssi", month) > 0) | person("receives_ssi", month).
• The mutant. I replaced it with person("receives_ssi", month) (dropping the modeled-SSI term) and ran the head's tests. It passes all of them:
  • medicaid_part_b_buy_in_premium_coverage.yaml 8/8;
  • medicare_part_b_premium.yaml 19/19;
  • test_part_b_msp_offset.py 15/15.
• Why no test catches it. Every new buy-in case enters receives_ssi. Every existing case that reached the modeled-SSI path (11 premium YAML cases and the order-dependence Python test) now enters ssi: 0. So nothing asserts a buy-in for a person the model itself finds to be an SSI recipient. In a microsimulation that is the main path, because receives_ssi is rarely populated.
• Probe at the head (Python API). CA, age 70, no income, MSP forced off: ssi 11,604 (967 × 12), buy-in 2,220, premium 0. The same in TX.
• Fix. Add a case to medicaid_part_b_buy_in_premium_coverage.yaml:

  - name: A modeled SSI recipient gets the buy-in without reported SSI.
    # No income, so SSI pays the 2025 federal benefit rate: 967 x 12 = 11,604.
    # MSP is switched off to isolate the buy-in: 185 x 12 = 2,220.
    period: 2025
    input:
      age: 70
      msp_income_eligible: false
      msp_asset_eligible: false
      state_code: CA
    output:
      ssi: 11_604
      medicaid_part_b_buy_in_premium_coverage: 2_220
      medicare_part_b_premium: 0

3. The annual max_() of MSP and buy-in coverage undercounts the state payment when the months don't nest (medicare_part_b_premium.py:23-27).

• The problem. Both coverage variables are annual sums of monthly standard premiums. max_(msp, buy_in) equals the months covered by either one only when one set of months contains the other.
• Example. Take MSP in January-June and SSI in July-December. The state pays all 12 months: MSP under 42 U.S.C. 1396a(a)(10)(E), then the buy-in under 1395v(d)(2)(C). The model charges the person for six of those months.
• Probe at the head (Python API). CA, age 70, $30,000 of interest, 2025:
  • msp_income_eligible true January-June and receives_ssi true July-December gives msp_part_b_premium_coverage 1,110, medicaid_part_b_buy_in_premium_coverage 1,110 and medicare_part_b_premium 1,110.
  • The law gives 0: (6 + 6) × 185 = 2,220 paid by the state.
• Reach. I found no path from annual inputs alone, so this is a period mistake with limited reach.
• Why it still matters. The PR's own point is month-by-month accuracy (its December-only and January-only cases), and the comment "a person covered by both is covered once" holds only month by month.
• Fix. Take the union by month. For example, add monthly booleans (MSP pays Part B; buy-in pays Part B), have the two annual coverage variables sum them, and compute state_paid as the sum over months of base_part_b_premium / 12 × (msp_m | buy_in_m).
• Pin it. Add the disjoint-month case above.

Suggestions

1. The Connecticut cases sit inside Connecticut's own MSP limits, which the model doesn't encode.
   • Connecticut's rules. QMB is 211% FPL, SLMB 231% and ALMB 246%, with no asset test (CT DSS, Medicare Savings Program eligibility).
   • The model. It uses the federal 100/120/135% floors for every state (savings_programs/eligibility/income/*/fpl_limit.yaml).
   • The December-only case. For its household, msp_countable_income is 29,760 a year, about 190% of the 2025 guideline. Under Connecticut's rules the person is a QMB and pays nothing all year. The case pins medicare_part_b_premium: 2_035 (medicaid_part_b_buy_in_premium_coverage.yaml:66).
   • Not a regression. This is a pre-existing gap: main gives 2,220.
   • Fix. Raise that case to $40,000 of interest. That is 39,760 of countable income, above 246% of 15,650 = 38,499. The 209(b) spend-down still passes (countable 33,580 less 42,220 of expenses), and the expectations stay 185 / 2,035. Alternatively, add a comment naming the gap and file a follow-up for state MSP income limits.
2. medicare_cost now values Part A-only coverage above Part B coverage.
   • Probe. ID, age 70, $12,000 of Social Security: medicare_cost is 17,751 for a Part A-only enrollee and 15,531 for a Part B enrollee.
   • Why. The per-capita cost includes Part B spending, and it is no longer offset by any Part B premium.
   • Fix. Either leave medicare_cost's offset on the medicare_enrolled gate, or drop Part B spending for a Part A-only enrollee. This reaches only the new non-default input.
3. msp_part_b_premium_coverage still gates on medicare_enrolled. A Part A-only enrollee who is MSP-eligible therefore shows 2,220 of MSP Part B coverage (probe), though no reader uses it. Gate it on medicare_part_b_enrolled once Add MSP take-up gate and any-month Kansas CSFP categorical eligibility #9495 and Apply immigration tests to Medicare and Medicare Savings Programs #10632, which edit that file, have landed.
4. Correct the PR body's IRMAA rationale.
   • The misreading. 1395v(d)(1)'s "without any increase under subsection (b)" refers to 1395r(b), the late-enrollment surcharge, not IRMAA (1395r(i)).
   • Better support for the standard-premium-only reading:
     • POMS HI 00815.001 ("States and territories pay the standard premium for individuals covered under state buy-in");
     • 42 CFR 407.40(c)(6);
     • 42 CFR 408.28.
   • Fix. Consider adding the POMS section to the variable's references.
5. Follow-ups are not tracked.
   • The gap. The body says "follow-up below" and "see follow-ups in the hand-off", but it lists none.
   • Suggested issues:
     • buy-in for other full-benefit Medicaid groups in states whose agreement uses 42 CFR 407.42(b)(3) (all Medicaid eligibles), where coverage starts the second month after the eligibility determination (407.47(d));
     • SSP-only and deemed SSI recipients (407.42(b)(1)(i) "SSI or SSP or both", and (b)(1)(ii));
     • Illinois SMIB netting;
     • state MSP income limits;
     • the Part B hold-harmless (42 U.S.C. 1395r(f)), which is not modeled.
6. Stale wording.
   • spm_unit_medical_out_of_pocket_expenses still says Part B is "net of Medicare Savings Program coverage where modeled".
   • About 16 LIHEAP/EAP readers say medicare_part_b_premium "excludes the share a Medicare Savings Program pays": AR, AZ, GA, IA, IN, KY, MD, NH, NJ, NY, ND, PA, SD, VA, WA, WY.
   • Their logic is still right, and the buy-in netting is what their rules want (ND's manual expressly excludes premiums for anyone "QMB/SLMB, QI-1 or Buy-In eligible"). Updating the comments is optional.
7. The 209(b) buy-in reads Medicaid eligibility, not enrollment.
   • The rule. 407.42(b)(1)(i) requires the person to be "covered under the State's Medicaid state plan". The changelog says "209(b) SSI recipients on Medicaid".
   • Why the code differs. Reading medicaid_enrolled would probably create a cycle through SNAP.
   • Fix. Say so in the variable's documentation.
8. The Python property test.
   • The rule. CLAUDE.md "Test design and CI cost" asks for a reason when a Python test covers behavior YAML can exercise. The YAML overlap, floor and Part A-only cases already pin these bounds.
   • Cost. Measured locally, not on Linux CI: the test body takes about 0.4 s with about 5 MB traced peak, and it reuses the module-level system. So the cost is negligible.
   • Fix. Add a one-line reason, or drop the test. Its annual inputs also can't detect finding 3.

Verified correct

Who is charged after the change. A person is charged max(gross − max(MSP coverage, buy-in coverage), 0) only if medicare_part_b_enrolled is true:

• gross is the standard premium plus any IRMAA;
• MSP coverage is the MSP months × standard/12;
• buy-in coverage is the SSI months × standard/12, for a person in a 1634 or SSI-criteria state or a 209(b) SSI recipient eligible for Medicaid.

A person not in Part B, or not in Medicare, is charged 0 through defined_for. IRMAA stays with the enrollee: in a probe, a CA SSI recipient with $150,000 of MAGI two years prior paid 2,220.

Readers. The PR body's list matches my grep at the head:

• Aggregates: medical_expense_health_insurance_premiums, spm_unit_health_insurance_premiums → SPM MOOP.
• Federal tax: Schedule A (itemized_medical_expenses, itemized_medical_expenses_excluded_premiums) and self_employed_health_insurance_ald_excluded_premiums.
• State tax: ID, MT, OH, and the WA property-tax exemption.
• Benefits: SNAP (and OK LIHEAP through it), HUD, PA CCW, UT HEAT, ME HEAP and NH FAP.
• LIHEAP/EAP: the income definitions for AR, AZ, GA, IA, IN, KY, LA, MD, MN, MO, MT, ND, NJ, NV, NY, OH, PA, RI, SD, TN, VA, WA, WI and WY.

No double-netting and no un-netting.

• No reader subtracts MSP or buy-in coverage again.
• Nothing outside the Part B variables reads either coverage variable.
• SSI and SSP don't read the premium.
• The two Medicaid eligibility readers (medicaid_medically_needy_medical_expenses and medicaid_optional_senior_or_disabled_countable_income) read the premium before the buy-in. So medically needy results don't change from main.
• California's "regardless of how paid" disregard swaps the before-buy-in paid amount for the gross amount, so the two cancel correctly.

LIHEAP readers. Their rules count Social Security net of the Medicare premium withheld from the check. A bought-in premium is not withheld, so netting it is correct. Values change only for people covered by the buy-in whom the modeled MSP doesn't already cover: reported SSI above the MSP limits, 209(b) spend-down cases, or MSP forced off. For a modeled SSI recipient, the modeled MSP almost always already pays.

Partner tests. The PR changes no file under tests/policy/baseline/partners/, and the Household API Partners job passed at the head. So no partner expected value moved, and the partner gate does not apply.

Hand checks (2025 standard premium $185, 2026 $202.90, both matching CMS):

#	Case	Arithmetic
1	CA, full-year SSI	buy-in 185 × 12 = 2,220
2	CA, $30,000 of interest	SSI countable (2,500 − 20) = 2,480 > 967, so no SSI and buy-in 0
3	Part A-only SSI recipient	defined_for gives 0
4	CT, 40,000 of medical expenses	23,580 − (40,000 + 2,220) < 10,016, so 209(b) eligible; buy-in 2,220, premium 0
5	CT, 12,000 of medical expenses	countable (2,500 − 535) × 12 = 23,580; expenses 12,000 + 2,220 = 14,220; 23,580 − 14,220 = 9,360 ≤ 0.64 × 15,650 = 10,016, while 23,580 − 12,000 = 11,580 is over
6	CT, December-only SSI	buy-in 185, premium 2,220 − 185 = 2,035
7	TX, December-only and January-only SSI, 2026	202.90 × 12 = 2,434.80; buy-in 202.90, premium 2,231.90
8	Overlap case	4,440 − max(2,220, 2,220) = 2,220
9	Floor case	max(2,220 − 3,000, 0) = 0
10	CA Part A-only countable income	(1,000 − 20) × 12 = 11,760
11	ID MSP regression	(1,000 − 20) = 980 ≤ 1,304.17 (100% FPL), so QMB, MSP 2,220, premium and subtraction 0
12	TX Part A-only, $40,000	premium, health insurance premiums and gross-if-enrolled all 0; with Part B, 2,220 each; other premiums still count (1,500)

Probes at the head:

• a CT SSI recipient who fails the 209(b) test gets no buy-in and pays 2,220 (correct, since 407.42(b)(1)(i) requires Medicaid coverage);
• modeled SSI in a 1634 state gets the buy-in (CA and TX).

Code checks:

• The buy-in is vectorized, with YEAR and MONTH periods used correctly (ssi and receives_ssi are MONTH).
• No new state-gated int or enum key, so the NH Add New Hampshire state supplementation for 2026 #10156 pattern doesn't apply.
• The new terms are pinned by mutation-style reasoning:
  • the Part B gate, by the Part A-only cases;
  • the max_ against summing, by the overlap case;
  • the floor, by its own case;
  • the 209(b) arm, by the CT cases;
  • the before-buy-in reads, by the CT window case and the cycle that results without them;
  • the direct-premium branch of the new Medicaid aggregate, by the existing "pretax wage exclusion" and CA total-premium cases.
• The references corroborate the code, I checked them against the eCFR and the U.S. Code, and they follow the reference rules.
• Tests are in variable-named YAML files.
• The changelog is a top-level changelog.d/medicare-part-b-payers.fixed.md, with type fixed.
• No conflict-sounding comments were added, so the source-conflict ratchet is unaffected.

Axiom line assessment

axiom: (hub files pe-parity before landing) is not a valid form, and no rulespec-us issue covers Part B enrollment or the SSI buy-in. This is Should address #1.

CI status

36/36 checks pass at aa46e20cd1, re-checked near the end of this review. The run includes the Full Suite - Microsimulation job, so the new dependency chain doesn't create a cycle in baseline computation. Nothing is pending.

Branch status

• Position. 7 commits ahead of main and 4 behind.
• Merge. git merge-tree --write-tree PolicyEngine/main aa46e20cd1 is clean. Main's four commits (Fix EITC eligibility for married separate filers #10307, EITC for separate filers; Store California state supplement parameters as state shares #10600, California state supplement shares) touch none of these files.
• Sibling overlaps (branch notes, not findings):
  • Limit SSI spousal-deeming child allocations to the couple's co-resident children #10146 moves the optional senior/disabled countable-income formula into a monthly file. Whoever lands second must carry this PR's two before-buy-in reads and its medicare_part_b_enrolled gate into that file, or the buy-in cycles.
  • Add 2015-2023 Medicare IRMAA brackets and 2015-2020 Part B premiums #10112 conflicts in comments only.
  • Add MSP take-up gate and any-month Kansas CSFP categorical eligibility #9495 and Apply immigration tests to Medicare and Medicare Savings Programs #10632 edit msp_part_b_premium_coverage.

Verdict: REQUEST_CHANGES

[DTrim99]

Requesting changes: 0 critical, 3 should (the axiom line isn't a valid form and no rulespec-us issue covers Part B enrollment or the SSI buy-in; the modeled-SSI arm of the buy-in is unpinned; the annual max of MSP and buy-in undercounts when the months don't nest, e.g. MSP Jan-Jun plus SSI Jul-Dec charges $1,110 where the law gives $0). The approach matches 42 USC 1395r/1395v, and no reader double-nets. Full report in the comment above.

Answered at 6c5d72e29f. Should 1: the axiom line cites rulespec-us#2032 and comments on #2006, #1931, #1933 and #1625. Should 2: a new buy-in case pins the modeled-SSI path, and dropping that term fails it. Should 3: the body lists the annual netting with your witness under "Not in this PR", since the fix touches msp_part_b_premium_coverage, which #9495 and #10632 edit. Suggestion 1: the December-only case now takes $40,000 of interest. Suggestions 2, 3 and 5: "Not in this PR" lists medicare_cost, the MSP enrollment gate and your follow-ups. Suggestion 4: the body and the buy-in's references cite POMS HI 00815.001 and 42 CFR 408.28 and drop the subsection (b) reading. Suggestion 6: the SPM documentation names the netting, and the LIHEAP and EAP comments stay as they read. Suggestion 7: the buy-in's comment, the changelog and the body say "qualifies for Medicaid". Suggestion 8: the property test gives its reason in a docstring.

[MaxGhenis]

US + core hub merge audit, us#10634 at 6c5d72e29f34e2f5a547469c8b07e0d008b4067b (squash): Charge the Medicare Part B premium only to people who pay it.

• Review: a GPT-6.1 Sol reviewer approved 6c5d72e29f in round 4, with METHODOLOGY_CHOICE yes and no blocking finding. It recomputed the three new or changed cases from CMS's $185 premium, SSA's $967 federal SSI amount and the 2025 poverty guideline. It traced 29 Python readers and three parameter-selected consumers, and it mapped each item in DTrim99's review.
• Known gap, older than this PR: when the Medicare Savings Program and the SSI buy-in cover different months of one year, the model charges $1,110 where the state pays all 12 months. Main charges the same $1,110. The body discloses the gap, and a fix needs coverage counted month by month. The reviewer's one other finding, a stale line on Limit SSI spousal-deeming child allocations to the couple's co-resident children #10146's conflicts, no longer applies: the body now names all three files.
• Mutants:
  • Turning the buy-in's SSI test from OR into AND fails 8 of 10 cases in medicaid_part_b_buy_in_premium_coverage.yaml, among them "A modeled SSI recipient gets the buy-in without reported SSI." (the reviewer's run).
  • Requiring income strictly below the limit in is_209b_ssi_recipient_income_eligible_for_medicaid.py fails "A 209(b) spend-down that lands on the limit qualifies.", 1 of 10 cases in the same file. The hub ran this one at this head; the unmutated file passes 10 of 10.
• Combination: the hub merged main at 43383c0aba into the head and ran the 47 test files that cover the PR's variables and their readers. All 47 pass. Since then, main has changed only the Social Security state backend test and the version.
• Gates checked live at merge:
  • gh pr checks exits 0;
  • the latest Pull request run at the head concluded success (42054053269, 36 of 36 jobs);
  • MERGEABLE;
  • not a draft;
  • no reviewer's latest review requests changes (the hub dismissed DTrim99's review 5776519338 with a note that maps each item);
  • the semantic-overlap guard;
  • the merge pins the head.
• Impact: no change in 2026 on the default dataset to Part B premiums, health insurance premiums, taxes, benefits or household net income, comparing main at 1104c5edb5 with 6c5d72e29f.
• Axiom: Encode Medicare Part B enrollment, SSI Medicaid buy-in and premiums paid by the beneficiary (42 USC 1395o, 1395r, 1395v; 42 CFR 407.40–407.48, 408.28) (PE-US PR 10634) TheAxiomFoundation/rulespec-us#2032 tracks the encoding. Comments on Encode Missouri's Medicaid spend down for aged, blind and disabled people not on SSI (42 CFR 435.121(e)(4)) (PE-US PR 10629) TheAxiomFoundation/rulespec-us#2006, Encode the Medicaid medically needy financial unit for aged, blind and disabled applicants (42 CFR 435.602(a)(2)(i), 435.831, 435.845; 20 CFR 416.1202(b)) (PE-US PR 10370) TheAxiomFoundation/rulespec-us#1931, Encode the optional aged, blind and disabled Medicaid disregards, premium deductions and asset limits for CA, CT, SC, MO, AR and OK (42 U.S.C. 1396a(m), (r)(2); state rules) (PE-US PR 10364) TheAxiomFoundation/rulespec-us#1933 and Encode Idaho Code 63-3022P health insurance premium subtraction and taxpayer-paid scope TheAxiomFoundation/rulespec-us#1625 extend those issues.

🤖 Generated with Claude Code