[MaxGhenis]

ILLUSTRATIVE HOUSEHOLDS, NOT SURVEY DATA. Applied to 2026 law and prices: the cut would not happen in 2026, and this is not a projection.

This PR shows what an across-the-board Social Security cut at trust fund depletion does to a household once federal and state taxes and benefits respond, and who pays the part that other programs offset. It is the second use of the PolicyEngine-US bridge from #496. Five illustrative households are each run in California, Montana and Florida through PolicyEngine-US 2.18.0 for 2026, installed from PyPI and checked against the wheel's RECORD.

The cut: 2026 Trustees Report

Source. The 2026 Annual Report of the Board of Trustees of the Federal Old-Age and Survivors Insurance and Federal Disability Insurance Trust Funds, section II.A (Highlights), intermediate assumptions.

• URL: https://www.ssa.gov/oact/TR/2026/II_A_highlights.html
• Retrieved 2026-10-01T02:35:26Z; SHA-256 8447f9cae65e561cb84239350e6a037f9b2572d18bf457212130d682c0003551.
• The page is committed with its fetch record under docs/analysis/pe_us_depletion_cut_20261001/sources/, byte-pinned in .gitattributes.

The OASI Trust Fund is projected to become depleted in the fourth quarter of 2032, one quarter earlier than projected in last year’s report. Upon reserve depletion in 2032, projected income is sufficient to pay 78 percent of scheduled benefits.

The combined OASDI fund is projected to become depleted in the third quarter of 2034, the same quarter as in last year’s report. Upon reserve depletion in 2034, projected income is sufficient to pay 83 percent of scheduled benefits.

Both quotes are on printed page 5 of the full report.

• Primary: a 22 percent cut in each OASI benefit, leaving the 78 percent payable share. The report's introduction (printed page 1) says retired workers, their families and survivors are paid under OASI, and Table III.A5 (printed page 40) lists retired workers and spouses under OASI benefit payments. Every benefit in these households is a retired-worker or spouse's benefit.
• Sensitivity: a 17 percent cut, leaving the 83 percent payable share for OASI and DI on a combined basis. The Trustees note that this figure "implicitly assumes that the law will have been changed to permit the transfer of funds between OASI and DI as needed" (printed page 28).
• First-year cuts. Both shares apply in the first year of depletion. The Highlights project the share to decline "gradually to 62 percent by 2100" for OASI, and to 65 percent for OASDI.
• How the numbers are taken. The script parses them from the quoted sentences and requires the page's Table II.A1 to state the same numbers. The full report, tr2026.pdf, is not committed; its SHA-256 is fb4e1556a00d925efa03e6355f64f911656a2626eaaceeb16062687008feb683.
• The across-the-board cut is an assumption. The Trustees project the payable share. Once reserves are depleted, "scheduled benefits could not be paid in full on a timely basis, and actual amounts paid would be less than the scheduled benefits" (Table IV.A1 note, printed page 46). They do not say how a shortfall would be spread across beneficiaries.
• Rounding. Each beneficiary's payable monthly benefit is floor(scheduled whole-dollar monthly benefit × share). That uses 42 USC 415(g)'s rule of rounding down to the dollar.
  • 415(g) rounds a benefit after the reductions and deductions of sections 403 and 424a and the Medicare Part B deduction; this analysis floors the gross benefit.
  • Applying the share to the unrounded amount would pay D and E a dollar a month more at 78 percent.
  • A couple's two benefits are cut and rounded separately.

Households

All five households share #496's common facts: $800 monthly rent, $1,500 in the bank, and housing take-up off.

Household	Social Security, monthly (scheduled → after 22% cut)	Other income	Derivation
A, B, C	$743 → $579	none / $4,800 / $31,200 pension	#496's low earner, worker_benefits() unchanged; cross-checked against #496's committed JSON
D	$2,351 → $1,833	$31,200 pension (C's)	Medium earner: 100% of the average wage index, 1980–2014, full retirement age 66 in 2020. history_pia gives AIME $3,873 and PIA $1,735.80; statutory COLAs carry it to $2,351.50
E	$2,351 + $1,175 → $1,833 + $916	$42,000 pension, entered on the worker	D's worker plus a spouse with no covered work. The spouse's benefit and the family maximum (not binding) come from Track B's gross-benefit layer, which is checked against D's derivation

Headline: the 22% (OASI) cut, 2026 law, annual dollars

• Offset share = 1 − (net change ÷ Social Security change). The federal and state columns split that share by who pays.
• Health coverage is outside PolicyEngine-US's default net income. The last column counts it, as Bridge to PolicyEngine-US: federal and state tax-benefit effect of the minimum benefit on sample households #496 does, valuing Medicaid at average program cost.

Household	State	Social Security	Net income	Offset share	Federal	State	Net income, health counted
A	CA	−1,968	0	100%	100%	0%	0
A	MT	−1,968	0	100%	100%	0%	0
A	FL	−1,968	0	100%	100%	0%	0
B	CA	−1,968	0	100%	21%	79%	0
B	MT	−1,968	−1,548	21%	21%	0%	+3,916
B	FL	−1,968	−936	52%	52%	0%	−936
C	CA	−1,968	−1,776	10%	10%	0%	−1,776
C	MT	−1,968	−1,654	16%	10%	6%	−1,654
C	FL	−1,968	−1,776	10%	10%	0%	−1,776
D	CA	−6,216	−5,899	5%	5%	0%	−5,899
D	MT	−6,216	−5,775	7%	5%	2%	−5,775
D	FL	−6,216	−5,899	5%	5%	0%	−5,899
E	CA	−9,324	−8,928	4%	4%	0%	−8,928
E	MT	−9,324	−8,741	6%	4%	2%	−8,741
E	FL	−9,324	−8,928	4%	4%	0%	−8,928

Sensitivity: the 17% (OASDI) cut

Household	State	Social Security	Net income	Offset share	Federal	State	Net income, health counted
A	CA	−1,524	0	100%	100%	0%	0
A	MT	−1,524	0	100%	100%	0%	0
A	FL	−1,524	0	100%	100%	0%	0
B	CA	−1,524	0	100%	0%	100%	0
B	MT	−1,524	−1,524	0%	0%	0%	+3,940‡
B	FL	−1,524	−912	40%	40%	0%	−912
C	CA	−1,524	−1,373	10%	10%	0%	−1,373
C	MT	−1,524	−1,277	16%	10%	6%	−1,277
C	FL	−1,524	−1,373	10%	10%	0%	−1,373
D	CA	−4,800	−4,555	5%	5%	0%	−4,555
D	MT	−4,800	−4,459	7%	5%	2%	−4,459
D	FL	−4,800	−4,555	5%	5%	0%	−4,555
E	CA	−7,200	−6,894	4%	4%	0%	−6,894
E	MT	−7,200	−6,750	6%	4%	2%	−6,750
E	FL	−7,200	−6,894	4%	4%	0%	−6,894

‡ ENCODING ARTIFACT, NOT STATE POLICY. See household B in Montana below.

Each dollar cell is rounded on its own, so a row can miss its own sum by a dollar.

What the table shows.

• Household A (the low earner, no other income). In all three states, SSI replaces the cut dollar for dollar: a 100 percent federal offset, and no change in net income.
• Household B (a $4,800 pension).
  • California: the cut is fully offset, 21 percent by SSI and 79 percent by the state's SSI supplement.
  • Florida: 52 percent is offset, by SSI and SNAP.
  • Montana: 21 percent is offset, by SSI; SNAP does not change.
• Household C (a $31,200 pension). 10 percent is offset by lower federal income tax. In Montana it is 16 percent, because Montana's income tax and refundable credits add 6 points.
• The medium earner and the couple (D and E). They absorb 93 to 96 percent of the cut. The 4 to 7 percent offset comes entirely through income taxes.
• The 17 percent (OASDI) cut follows the same pattern, with two differences:
  • Florida: household B's offset is 40 percent.
  • Montana: household B's SSI countable income is $996 a month, $2 above the 2026 federal benefit rate, so PolicyEngine-US pays no SSI and the offset is 0 percent.

Household B in Montana, health coverage counted.

Under the 22 percent cut:

• SSI begins at $420 a year, and PolicyEngine-US gives SSI recipients in Montana Medicaid (is_ssi_recipient_for_medicaid.py:20-34, checked before any other category in medicaid_category.py:44-47). Medicaid, valued at average cost, adds $10,799.
• The Medicare Savings Program value of $5,335 ends:
  • $2,434.80 is Medicare premiums.
  • $2,900 is QMB cost sharing, which PolicyEngine-US approximates as 20 percent of an average Medicare cost per enrollee of $14,500, an entry its parameter file marks "Estimated".

Under the 17 percent cut, the +3,940 is an encoding artifact, not state policy:

• SSI stays $0, but PolicyEngine-US 2.18.0 starts Medicaid through its optional aged and disabled pathway. Its limit for Montana is 75 percent of the poverty guideline, $11,970 a year (senior_or_disabled/income/limit/individual.yaml:80-81), and B's countable income is $18 under it.
• Montana's own rules set a different standard:
  • Montana covers aged, blind and disabled people who meet SSI's income standard without receiving SSI (State Plan Attachment 3.1-F, TN 24-0002, row 7, 42 CFR 435.210).
  • Its limit is the SSI payment rate in dollars, $994 a month for one person in 2026 (DPHHS Medicaid manual ABD 008).
  • The same plan page marks the poverty-level aged or disabled group (1902(m)) N/A.
• The parameter's 0.75 matches its cited source, KFF, which writes the SSI rate as a share of the poverty guideline.
• Reported upstream as Medicaid aged/disabled income limit is 75% of poverty where the state standard is the SSI rate (MT, CO, IA, LA, OH, OR, WA) policyengine-us#9733, with a fix proposed in #9735. Under that fix, B gets no Medicaid under the 17 percent cut.
• Every artifact labels this figure; the tables mark it ‡.

Outputs. In docs/analysis/pe_us_depletion_cut_20261001/:

• JSON with full provenance: package versions, the source and RECORD check, the Trustees and statute citations, inputs, and each component;
• a Markdown table;
• one waterfall chart per household (PNG and SVG);
• offset_share_summary, the offset share by household and state.

Every artifact carries the ILLUSTRATIVE label and the 2026-law note. Each chart's visible title starts "Illustrative households:", and its file metadata carries the label. The design note is docs/design/pe_us_depletion_cut.md.

Who pays the offset

depletion_cut.LEAF_LEVELS assigns each reviewed leaf of the bridge's decomposition to a payer group. Every statute phrase quoted below is checked against the saved Legal Information Institute page and its SHA-256.

• Federal.
  • Federal income tax and refundable credits.
  • SSI: "there are authorized to be appropriated sums sufficient to carry out this subchapter" (42 USC 1381).
  • SNAP: benefits are "redeemable at face value by the Secretary through the facilities of the Treasury of the United States" (7 USC 2013(a)(1)). The State cost share in 2013(a)(2)(B) starts in fiscal 2028 at the earliest.
• State.
  • State income tax and refundable credits.
  • California's SSI supplement. 42 USC 1382e(a) describes such supplements as cash payments made by a State. Under 1382e(d)(1), where SSA makes the payments on a State's behalf, the State pays SSA back.
• Joint. Medicaid, and the Medicare Savings Programs other than QI. These appear in the health sensitivity only.

The script refuses a run in which any unreviewed leaf or any market-income leaf changes.

Invariants

Hypothesis properties in tests/bridge/test_depletion_cut.py:

• A zero cut (payable share of 1) changes nothing. Identical runs decompose to zero, with no offset share.
• The cut is the cut share times the benefit, within the stated rounding: at least that amount, and less than $1 a month more.
• A deeper cut never leaves more Social Security, for one benefit or for a whole household.
• The components sum exactly, in cents, to the net change.
• The government-level groups partition the components (baseline, reform and change).
• Whenever Social Security changes, the offset share is exactly 1 − (net change ÷ Social Security change), and the level shares sum to it.

Before writing, the script also checks every comparison (raising, never asserting):

• the Social Security change equals minus the computed cut;
• each payer group equals the bridge's display categories;
• Bridge to PolicyEngine-US: federal and state tax-benefit effect of the minimum benefit on sample households #496's wrong-way check, mirrored for a cut, finds no fall in SSI, SNAP, CSFP, state benefits or refundable credits, and no rise in federal or state income tax. Health programs are not in that check: household B's Medicare Savings Program value in Montana does fall;
• no joint change appears without health coverage counted;
• no all-federal (QI) Medicare Savings Program value is labelled joint;
• PolicyEngine-US's own net-income change agrees with the decomposition to within a cent;
• Bridge to PolicyEngine-US: federal and state tax-benefit effect of the minimum benefit on sample households #496's float32 guard finds no change without a cause;
• Medicaid at cost equals a whole number of enrollees.

Artifact test. tests/bridge/test_pe_us_depletion_cut_artifact.py rechecks on the committed JSON:

• the identities;
• the Social Security change against the cut;
• that nothing changes outside the payer groups;
• the wrong-way categories;
• the payer groups against the display categories;
• that no joint change appears without health coverage;
• the float32 guard.

It also checks:

• the labels, both the visible titles and the file metadata;
• the citation, and that the payable shares used are the ones parsed from the quotes;
• the encoding-artifact labels;
• every column of the summary tables;
• the low earner's baseline against Bridge to PolicyEngine-US: federal and state tax-benefit effect of the minimum benefit on sample households #496's committed benefit;
• that the run was fresh, and that its recorded commit is an ancestor of HEAD with the same src/ and scripts/.

Oracle-tier tests, run live:

• the benefit derivations;
• a zero cut: household A in Florida, built with a payable share of 1 and run as its own case against the baseline;
• SSI offsetting household A's cut dollar for dollar;
• taxes offsetting part of household D's cut.

Caveats

• No behavioral response. Nothing changes in work, claiming, saving or living arrangements.
• Default take-up. PolicyEngine-US assumes full take-up of SSI, SNAP and Medicaid. The Medicare Savings Programs have no take-up input. Housing assistance is switched off, as in Bridge to PolicyEngine-US: federal and state tax-benefit effect of the minimum benefit on sample households #496.
• 2026 law and prices. The cut would not happen in 2026. Some rules already in law differ by the depletion date, for example the SNAP State cost share. The Trustees Report calls the senior deduction temporary; this analysis applies it as 2026 law does.
• Illustrative households, not survey data. They show mechanisms, not how many beneficiaries are in each situation.
• Not a projection.
• Federal grouping. Part of the federal income tax on benefits is credited to the trust funds; it is grouped as federal here.
• Medicaid valuation. Medicaid at cost is valued as Bridge to PolicyEngine-US: federal and state tax-benefit effect of the minimum benefit on sample households #496 does: an average over all enrollees, not uprated.

Review, run and tests

Review rounds.

• Before the PR, three independent read-only reviews checked correctness and test adequacy, claims against PolicyEngine-US source and statute text, and the benefit derivations. A round-2 check of their fixes followed.
• After the PR opened, two independent reviews ran:
  • A reproduce review re-ran the script from a clean checkout. Every number reproduced, and nine deliberate mutations each failed a test.
  • A claims review asked for changes to the Montana Medicaid explanation, the Trustees wording, the wrong-way and test claims, the chart titles, two absolutes and a CalWORKs detail.
• Those are fixed in 7199458, bf7dec1 and d0c6933. The claims re-review then found the generated caveat calling the 78 percent payable share "the cut" (fixed in a360ca3) and noted that household D's subtitle overlapped the state headings and the summary chart's title was clipped (fixed in 5ac5ab5).

The run.

• The outputs come from a fresh run from a clean clone at 5ac5ab5 (committed in 363d9e8). Against the outputs at 9001436, all 10,736 numeric JSON fields are identical; only the recorded commit and the corrected caveat sentence differ.
• It ran on Modal (Linux, 8 cores, 32 GiB), not on the Mac, which was out of memory. The container reproduces the local paths and the package versions frozen from the local environments. policyengine-us 2.18.0 comes from PyPI, and its 17,551 files match the wheel's RECORD.
• The Modal runner and its log are kept with the evidence for this analysis, outside the repository.
• No number changed from the earlier macOS runs; the differences are text, chart titles and provenance.
• The runs cover 90 cases. The float32 guard traced every changed leaf and found no change without a cause.

Targeted tests, Python 3.14:

• tests/bridge/test_depletion_cut.py and tests/bridge/test_pe_us_depletion_cut_script.py: 119 passed. In CI the chart smoke test skips: matplotlib is in the model extra, and CI installs only dev.
• tests/bridge/test_pe_us_depletion_cut_artifact.py: 27 passed against the committed outputs.
• tests/bridge/test_pe_us_depletion_cut_oracle.py: 5 passed live against policyengine-us 2.18.0.
• All five files together (unit and script, artifact, oracle, and tests/estimates/test_birth_evidence_artifact.py) were re-run at 9001436 in the same Modal container: 164 passed, none skipped. Re-run at 363d9e8 on the host, the oracle live against policyengine-us 2.18.0: 164 passed.
• Wiring: the new module is listed in POST_REVIEW_SOURCE_EXCLUSIONS and in bridge_modules, next to Bridge to PolicyEngine-US: federal and state tax-benefit effect of the minimum benefit on sample households #496's bridge.
• Tier counts after merging origin/master (Run PolicyEngine-US over a validated invented population #498): unit 5,824 and oracle_policyengine 220, matching tests/tier_counts.json and tests/README-tiers.md.

black -l 79 and ruff pass on all changed Python.

🤖 Generated with Claude Code

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated
social-security-model	Ready	Preview	Oct 3, 2026 12:01pm UTC

[MaxGhenis]

Merge record:

• Max's go: 2026-10-03 in chat ("yes" on d775).
• CI: gh pr checks 506 exits 0 on 7e082b6; MERGEABLE; not a draft; no CHANGES_REQUESTED.
• Independent review: five claims-review rounds (review-claims-r1 to r5); r5 APPROVE on 363d9e8. 7e082b6 adds only a merge of master, resolving tests/tier_counts.json and tests/README-tiers.md by recounting the merged suite (unit 5,828; artifact 3,331; integration_psid 1,341; reproduction_legacy 520; oracle_policyengine 220). No analysis code or output changed between 363d9e8 and 7e082b6.
• Merged as a merge commit so the commit the outputs record (5ac5ab5) stays in history.