US Medicare IRMAA: enrollee-gated Part B premium target and held-out IRMAA validation rows (#1206)

[MaxGhenis]

Implements the Microcosm changes Max approved in d1372 (2026-10-12) for #1206: changes 1 and 3. Change 2 waits on PolicyEngine/policyengine-us#10115, and change 4 is out of scope.

1. Part B premium target binds the enrollment-gated premium

The CMS premiums-from-enrollees target (2025 Trustees Report Table III.C3, CY2024 $139.837B, Chronicle package cms-medicare-trustees-report-2025-part-b-premium-income) moves from gross_medicare_part_b_premium to gross_medicare_part_b_premium_if_enrolled.

• The old variable is defined for is_medicare_eligible, so it also charged eligible people who never enrolled.
• The new one is defined for medicare_enrolled, and pe-us documents it "for CMS premiums-from-enrollees calibration targets".
• Nothing else about the target changes: same name, value, period, family and metadata.

pe-us version. Microcosm locks policyengine-us 2.2.1 (uv.lock; the extras allow >=2.2.1,<3), which does not include PolicyEngine/policyengine-us#10106.

• #10106 merged on main as 8df4dffb2f on 2026-10-12T01:03Z.
• The first PyPI release that includes it is 2.40.0 (uploaded 2026-10-12T06:49Z). I checked the wheels: 2.40.0 and 2.40.1 ship medicare_irmaa_magi and the same-year switch; 2.39.1 does not.

Microcosm locks the engine its certified default was built with (#1086), so moving to ≥2.40.0 is a separate step, not done here. Until then the target reads without #10106.

What the target reads on the certified file (populace-us-2024-spm-20260915, 2024, uncalibrated weighted sums). These come from real microsimulations in the #1206 diagnosis (crfb-m4a/research/irmaa-run/diagnosis/part_b_target_check_*.json):

pe-us	old: gross_medicare_part_b_premium	new: ..._if_enrolled
2.2.1, locked (2024 IRMAA is $0)	$128.23B (−8.3%)	$120.58B (−13.8%)
2.2.1 + #10106	$148.61B (+6.3%)	$137.65B (−1.6%)

Under 2.2.1 the old mapping misses by less only because it charges eligible non-enrollees the base premium.

Caveat for #1206 (not changed here). The $139.837B total includes $14.410B from disabled enrollees under 65; enrollees aged 65 and over paid $125.427B. Source: 2025 Trustees Report Table III.C1, CY2024, PDF p. 86. The model has no Medicare enrollees under 65 (#876), so about 10.3% of this target is out of the model's reach. Whether to target the 65-and-over row instead is a methodology call, raised on #1206.

3. Held-out IRMAA validation rows

A new file, us_medicare_irmaa_holdout.json, is written beside us_ssi_social_security_holdout.json and follows the same pattern as #1178.

• It is scored on the written release H5 through the household-batched post-export scorer.
• It is listed in the release manifest as a diagnostics artifact.
• It is never a calibration row and never a release refusal. An unavailable measurement is reported as status: unavailable.

It compares the model with the 2026 Medicare Trustees Report. These are the CY2024 historical rows, checked against the report PDF (sha256 ffa56b91…, pinned in the module):

	Table	Page (printed / PDF)	Beneficiaries affected	Aggregate premiums
Part B	V.E3	209 / 215	4.8M	$12.8B
Part D	V.E4	212 / 218	4.2M	$2.2B

Model measures, per person, weighted by the calibrated household weight. Both Parts count medicare_enrolled people only (Part A and/or Part B):

• Part B: gross_medicare_part_b_premium_if_enrolled − base_part_b_premium, because base_part_b_premium is defined for every eligible person. This matches the table's footnote 2: amounts "in excess of the Part B standard premium".
• Part D: income_adjusted_part_d_premium_surcharge. pe-us charges it to every Medicare enrollee above the thresholds until pe-us#10115 gates it on Part D enrollment. Any surcharge outside Medicare enrollment is excluded and reported as outside_enrollment_weighted_amount.
• Payer: an enrollee with a positive annual amount. The tables don't say whether "beneficiaries affected" is an average-month, year-end or ever-in-year count, and the payload says so.
• Status: available when some weighted enrollee pays. Otherwise it is no_payers or no_medicare_enrollees: the model totals are still reported, the Trustees comparisons are withheld rather than shown as −100%, and the payload gives the reason.

Source registration. The rows are cited constants in the module: title, URL, tables, printed and PDF pages, PDF hash, footnote definitions. That is how the SSI holdout registers SSA Table 9. I did not add a Chronicle package, because Chronicle facts feed the calibration compiler and these rows must stay out of it. The pinned feed carries no IRMAA record set, and a test pins that.

What the holdout will read. On locked 2.2.1 it reports no_payers for 2024, because 2024 IRMAA is effectively $0 on the certified file. Verified end to end: the builder's own _write_medicare_irmaa_holdout, run through the real household-batched scorer on the certified H5 (sha256 6496cc43…), wrote status: no_payers. It found 57,519,410.6 weighted Medicare enrollees and no payer in either Part, in 46.7 s (crfb-m4a/research/irmaa-run/holdout_check/scorer_pinned.json). In the diagnosis run (irmaa_aggregates_pinned-pe-us-2.2.1.json), the weighted count of tax units with a nonzero two-years-prior MAGI rounds to 0.00M, and both IRMAA totals round to $0.000B. pe-us 2.2.1 takes that MAGI from 2022 AGI and income, and the H5 carries 2024 inputs only. With #10106 the diagnosis measured Part B IRMAA among enrollees at $17.067B and the Part D surcharge at $3.278B for 2024 (irmaa_aggregates_patched-2.2.1-pr10106.json).

Invariants (tested)

• The remap changes only the bound variable. A differential test compiles the same ledger fact under the old and new mapping. Only base_variable differs.
• The engine gates hold. An engine-contract test pins that, in locked pe-us, _if_enrolled is a yearly Person variable gated on medicare_enrolled, and the old variable is gated on is_medicare_eligible.
• The holdout never writes its inputs, so it never changes weights. Hypothesis feeds it read-only arrays. It runs after export, on the bytes the manifest pins.
• Totals behave like weighted sums. Payers and premiums scale linearly with the weights, are invariant to person order, and add over any split of persons (Hypothesis). The payload is identical for household batch sizes 1, 2 and unbatched through the real scorer code.
• Totals are bounded. For each Part, 0 ≤ payers ≤ weighted enrollees. Premiums are ≥ 0, and $0 when no weight pays.
• Non-enrollees never reach either Part's measure, whatever premium or surcharge they carry (Hypothesis). A Part D surcharge outside enrollment is reported, not counted.
• Comparisons are reported exactly when some weighted enrollee pays (Hypothesis). Otherwise the status is no_payers or no_medicare_enrollees, never a −100% comparison.
• Validation only. No ledger mapping binds base_part_b_premium or income_adjusted_part_d_premium_surcharge or carries an IRMAA role. No pinned Chronicle record set mentions IRMAA or income-related premiums. Only tools/build_us_fiscal_refresh_release.py imports the holdout module.
• Engine and table agree. An engine-scenario test on pe-us 2.2.1 checks the measures on a household with one enrollee and one eligible non-enrollee, with the MAGI supplied. The engine's 2024 top-tier amounts ($419.30 Part B, $81.00 Part D a month) equal Tables V.E3/V.E4.
• Today's zero is pinned. A second scenario test shows a 2024-only household with $600k of wages pays no IRMAA on 2.2.1, and the holdout reports no_payers. It fails, by design, once Microcosm's engine includes #10106.
• Registry version. The national_state surface version moves from f96be9f41781 to e1f9906bf0e0, because it hashes target metadata. Target counts are unchanged. National concept-budget keys exclude base_variable, so this target's loss weight is unchanged.

Review

Independent review by GPT-6.1 Sol at 42d899c (~/reviews/irmaa-1206/review_42d899c52.md) asked for changes. All findings are addressed in 5570ee8, and the delta review there approved (review_5570ee8b7.md). Its five non-blocking observations are addressed in 8714515: per-Part withholding of comparisons, the verified scorer-path outcome in the guide, conditional reason wording, a robust import guard, and the changelog direction. The final delta review approved at 8714515 with no defects found (review_871451578.md). B1 is the no_payers status plus a locked-engine scenario test. B2 is the caveat in the code comment and changelog. N1 names the gates. N2 masks Part D by enrollment. N3 adds the import guard and drops the exact-set assertion. N4 is a fixture note.

Not in this PR

• Change 2, Part D take-up imputation. It waits on Charge the Part D income-related premium only to Part D enrollees policyengine-us#10115 (open, review required) and a Microcosm pe-us bump that includes it. It is tracked on US Medicare IRMAA: enrollee-gated Part B premium target, Part D take-up input, IRMAA validation rows, SOI age-by-AGI counts #1206.
• Change 4, SOI Table 1.6 age-by-AGI counts. It goes with Bind the SOI Table 1.1 size-of-AGI distribution as calibration targets #960.
• No data was regenerated, released or published. A data release is Max's separate go.

axiom: n/a: Microcosm calibration-target mapping and a validation diagnostic, no policy rule.

🤖 Generated with Claude Code

[MaxGhenis]

Merged at the reviewed head 8714515 as a merge commit. The gates held:

• gh pr checks exit 0 on the latest run: all 10 jobs, including both engine-us jobs and integration-uk.
• MERGEABLE, not a draft, no changes requested.

Independent review (GPT-6.1 Sol, via Subfleet):

• 42d899c: REQUEST_CHANGES. Two blocking findings: the locked-engine zero, and the Part B target caveat.
• 5570ee8: APPROVE.
• 8714515: APPROVE, no defects.

The review files are in ~/reviews/irmaa-1206/.

Real-data runs of the holdout on the certified populace-us-2024-spm-20260915 H5 (sha256 6496cc43…), 2024. Script and outputs are in crfb-m4a/research/irmaa-run/holdout_check/.

1. Release builder path, locked pe-us 2.2.1. The builder's _write_medicare_irmaa_holdout, through the household-batched scorer, gave status: no_payers: 57,519,410.6 weighted Medicare enrollees and no payer in either Part (46.7 s).

2. Direct microsim, pe-us 2.2.1 with the #10106 backport. status: available:

   	Payers	Aggregate	vs Trustees CY2024 (4.8M / $12.8B; 4.2M / $2.2B)
   Part B	5.649M	$17.067B	+17.7% payers, +33.3% dollars
   Part D	5.649M	$3.278B	+34.5% payers, +49.0% dollars

   Part D payers equal Part B payers because the engine charges the Part D surcharge to every Medicare enrollee until Charge the Part D income-related premium only to Part D enrollees policyengine-us#10115. The dollar totals match the US Medicare IRMAA: enrollee-gated Part B premium target, Part D take-up input, IRMAA validation rows, SOI age-by-AGI counts #1206 diagnosis's own 2024 run ($17.067B and $3.278B).

No data was regenerated, released or published.