Skip to content

Add Hawaii's employee temporary disability insurance contribution - #10677

Merged
MaxGhenis merged 12 commits into
mainfrom
hi-tdi-employee-contribution
Nov 14, 2026
Merged

MaxGhenis merged 12 commits into
mainfrom
hi-tdi-employee-contribution

Conversation

@MaxGhenis

@MaxGhenis MaxGhenis commented Nov 6, 2026 •

Copy link
Copy Markdown
Contributor

Adds Hawaii's employee temporary disability insurance (TDI) contribution and counts it in employee payroll tax and household net income, following the New Jersey TDI pattern. Because Hawaii lets the employer choose whether to withhold, the contribution also follows main's pattern for New York's contributions and reads the employer's deduction share.

What Hawaii's law says

Question Answer Source
Employee contribution An employer may withhold “one-half the cost but not more than .5 per cent of the weekly wages”. So 0.5% is a ceiling, and half the cost can bind below it. The law authorizes withholding and doesn't require it. HRS 392-43(a), Form TDI-15
Weekly wage cap The State's average annual wage for the preceding year ÷ 52, “multiplied by the factor 1.21”, where the average is the one HRS 383-61(b) computes, rounded to the nearest $100. The cap reaches $1,500.87 in 2026. DLIR's notice publishes $1,500.21, from the average before rounding (see "Reading to confirm: which average sets the cap"). HRS 392-43(b)(2), HRS 383-61(b), 2026 notice
Yearly change The Director publishes a new amount each year. It follows the wage, so it fell in 2011 and 2021. Yearly DLIR notices, which give each year's average annual wage; maximum_weekly_wage_base.yaml cites each one
Who counts Wages from a payroll period count only if, in the prior 52 weeks, the employee earned at least $400 and worked at least 14 weeks of 20 or more paid hours. HRS 392-43(b)(1)
Employer share The employer pays the balance of the cost. The law sets no employer rate, so this PR models no employer dollars. HRS 392-43(a)

What this PR changes

  • Parameters under gov/states/hi/tax/payroll/temporary_disability_insurance/: the rate limit, 23 yearly wage bases, the three parts of the wages test, and the payments the statute leaves out of wages.
  • Variables under gov/states/hi/tax/payroll/: wages, the wages test, taxable wages, the employee contribution and hi_employee_state_payroll_tax.
  • employee_state_payroll_tax now adds hi_employee_state_payroll_tax. From there the amount reaches employee_payroll_tax, household_tax_before_refundable_credits and household_net_income, as New Jersey's does. It also reaches spm_unit_payroll_tax, spm_unit_taxes and spm_unit_net_income, and so SPM poverty.
  • The contribution is the 0.5% limit times state_payroll_premium_deduction_share, which clips the employer_state_payroll_premium_deduction_share input (default 1) to between 0 and 1. #10285 set this up on main for contributions the law lets the employer withhold without requiring it, such as New York's. The documentation of employee_state_payroll_tax, employer_additional_state_payroll_tax and that input now lists Hawaii TDI with the programs whose employer cost the model doesn't hold.
  • Federal parameters: the 401(k) elective deferral limit, the catch-up amounts and the catch-up age now start in 2004, from the IRS cost-of-living table. Hawaii's wages need them because the statute takes deferrals out in every year. The backfill also corrects 2015–2017: the first value on file, $18,500 for 2018, had filled those years, and the IRS table sets $18,000.
  • programs.yaml gains a hi_tdi entry with status: partial.

Reading to confirm: what counts as wages

HRS 392-3 counts “all remuneration for services” and leaves out what HRS 383-11 lists. This PR builds a Hawaii wage definition from those two sections instead of reusing payroll_tax_gross_wages.

Amount Federal payroll base This PR Why
Pre-tax health premiums, payroll HSA contributions Out Out HRS 383-11(1)(C), “Medical or hospitalization expenses”
401(k) deferrals In Out HRS 383-11(4)(A): payments “by an employing unit to, or on behalf of, an individual” to a “trust described in section 401(a)”
403(b) deferrals In Out HRS 383-11(1)(A), “Retirement”
Dependent care benefits Out In HRS 383-11 lists no such payment
Military basic pay In Out HRS 392-5(6) leaves service in the employ of the United States government out of employment

Only paragraph (5) stops at payments made “without deduction from” pay. Paragraphs (1) and (4) carry no such limit. The federal base differs because of IRC 3306: (r)(1)(A) keeps 401(k) deferrals in, (b)(5)(D) keeps 403(b) salary reductions in, and (b)(5)(G) and (b)(13) take cafeteria-plan amounts and dependent care out. Hawaii enacted none of these. The treatment of Roth deferrals follows from the text, which sets no income tax condition. The temporary disability rules take wages from HRS 392-3. They and the unemployment insurance rules address pay in kind given rather than deducted from money wages. Neither set of rules names a plan contribution or salary reduction. The employer handbook lists payments to welfare plans among its exclusions but doesn't address salary reductions, so no agency reading breaks a tie.

The effect stays small: $60,000 with a $4,000 401(k) deferral gives $280 here and $300 on the federal base. source_conflicts/candidates.yaml records this as us-hi:tdi-wages-salary-reductions, with excerpts in source_conflicts/sources/us-hi/. The register entry and its marker follow after merge, as CONTRIBUTING.md sets out.

Reading to confirm: which average sets the cap

HRS 392-43(b)(2) takes the average annual wage as HRS 383-61(b) determines it, and 383-61(b) ends its computation by rounding that average to the nearest $100. DLIR's temporary disability insurance notices apply 1.21 / 52 to the average before rounding. This PR follows the statute: each year's base is the notice's average rounded to $100, times 1.21 / 52, rounded once to the cent. For 2026, $64,471.82 rounds to $64,500, which gives $1,500.87 against the notice's $1,500.21. Across 2004–2026 the statutory base runs from $0.84 below to $1.15 above the published one, so a full-year worker at the cap pays from $0.22 less to $0.30 more a year than the notice implies.

DLIR computes one average for both programs. Its unemployment insurance taxable wage base equals the TDI notice's average rounded to $100 in each year from 2003 to 2026 but four. The law fixed it at $13,000 in 2008 and 2009, and in 2010 and 2011 DLIR took 90% of the average and rounded again to $100. Either reading gives that base, since the first sentence of 383-61(b) also rounds it to $100. The definition settles the reading, so the principle that the agency's reading breaks ties doesn't reach it. source_conflicts/candidates.yaml records this as us-hi:tdi-wage-base-rounded-average, with excerpts of both sections and the 2026 notice in source_conflicts/sources/us-hi/. The register entry and its marker follow after merge. Employers who withhold against DLIR's published cap will differ from the model by the amounts above.

Within the formula the parameter rounds once, because the statute sets no intermediate rounding. Every notice except 2010's rounds the weekly wage to the cent first, and that convention would move 14 of the 23 values by a cent.

Invariants

These hold for every input, in the years each variable can run (see "Earlier years"):

  1. 0 ≤ contribution ≤ 0.5% × min(wages, base × counted weeks), where wages exclude the listed payments and military basic pay, and counted weeks default to 52.
  2. The contribution never falls when wages rise.
  3. The contribution is zero outside Hawaii.
  4. hi_employee_state_payroll_tax equals each person's contribution, and employee_state_payroll_tax equals the sum across a Hawaii tax unit.
  5. Under current law, removing the contribution lowers employee_state_payroll_tax, employee_payroll_tax and household_tax_before_refundable_credits by the contribution, and raises household_net_income by the contribution. The test runs this identity on three 2026 workers. It should hold beyond them because no Hawaii tax or benefit formula reads these totals: the only such formulas I found sit in Alabama, Delaware and two California counties.
  6. For one employer and a worker who passes the wages test, the annual amount is at least the sum of the weekly statute's amounts, and the two match for even wages.
  7. Each wage base parameter equals its notice's average annual wage, rounded to $100, times 1.21 / 52, rounded to the cent, computed from a stored transcription of the notice. Each notice's own base follows the same formula from the average before rounding, and its deduction equals 0.5% of its base, rounded to the cent. The two bases stay within $1.18 of each other: rounding moves the average by at most $50, or $1.16 of base, and rounding to cents adds at most $0.02.
  8. Previous-year history only adds workers. Take a Hawaii worker with the simulation year's pay, excluded payments, weeks and hours on file, previous-year pay on employment_income_before_lsr (where Simulation moves situation employment_income), and any deferral before 2015 entered as the contribution itself. With history, that worker passes the wages test if and only if they pass without it or the previous year shows at least $400 of wages under Hawaii's definition, 14 weeks and 20 usual hours. If a household enters history but leaves out any of those for the simulation year, the model carries the history into that year, uprating pay and excluded payments, where it can change that year's test either way. A YAML case pins the weeks-and-hours case.
  9. The contribution equals the clipped deduction share times its amount at the default share, and the employer amount stays at zero at any share. test_state_payroll_premium_deduction_share_invariants.py checks both, now with Hawaii among the optional programs.

The test module labels three intended differences from the published figures:

  • The parameter starts from the rounded average, so it runs from $0.84 below to $1.15 above each notice's base (us-hi:tdi-wage-base-rounded-average).
  • The parameter rounds the base once, to the cent, as the 2010 notice does. Rounding the weekly wage first, as the other notices do, would move 14 of the 23 values by a cent.
  • A full-year worker over the cap pays 52 × 0.5% × base here. DLIR rounds each week's maximum to the cent, so 52 weeks of it differ by at most $0.26 a year.

Assumptions and limits

  • Statutory limit. The model holds no plan cost, so it takes the 0.5% limit as the permitted amount and multiplies it by the employer's deduction share, which defaults to 1. Half the cost can bind first, so at the default the amount is an upper bound. DLIR's own example, with a premium of 80 cents per $100 of wages on $300 of weekly wages, gives $1.20 against a $1.50 limit, which a share of 0.8 reproduces.
  • Annual cap. The statute caps each week. The model uses base × weeks worked, or 52 when weeks went unreported. This is an upper bound for uneven wages. It can't separate two concurrent employers.
  • Wages test. The statute looks back 52 weeks from each payroll period, and the model approximates that with annual inputs. A worker passes on the simulation year's wages, weeks worked and usual weekly hours, or on the previous year's. So a worker who retires in January after a full prior year pays on January's wages, and new hires' first weeks stay in once the simulation year passes. The previous-year route needs all three values on file, and its wages must reach $400 after the same exclusions, so a prior year of active-duty pay or self-employment doesn't pass. A single year of inputs leaves the previous year at zero, so a short year with no history on file pays nothing. I kept that on purpose: the statute counts wages only when the employee has that history, and dropping the weeks part would charge new entrants and seasonal workers instead. The previous-year route has four limits. It skips the state test, because households often enter a state for the simulation year alone, so a worker new to Hawaii can pass on wages earned elsewhere. Its weeks and hours count all work, self-employment included. It reads employment_income_before_lsr, as SNAP's earnings test reads pre-response earnings, so it calculates no labor supply response for the earlier year. I checked that under a reform with nonzero elasticities: calculating the 2026 contribution and household net income, in either order, computed a response for 2026 alone, while a copy that read employment_income also computed one for 2025. In that case both copies gave the same contribution and net income, so there the read saved work rather than changing an answer. No test pins it, because a reform simulation with labor supply responses would add its extra branches to CI's core group. If set_input adds pay on employment_income after the simulation exists, that pay stays out of the previous year's wages, as it stays out of SNAP's earnings test, which reads the same variable; enter it on employment_income_before_lsr. Before 2015 it reads only inputs for that year, so it calculates and caches nothing before the model begins. A value the simulation calculated for that year first stays out, and a deferral counts only when entered as the contribution itself, not as the desired amount. For four households, every order of calculating 2024, 2025 and 2026 in one simulation gave the same answers. In a microsimulation the dataset year has no previous year on file, and later years read carried weeks and hours and uprated wages. Usual weekly hours stand in for each week's hours, so a worker with 14 qualifying weeks and lower average hours pays nothing.
  • Unreported values. Zero weeks or zero hours count as unreported, so that part of the test passes. The documentation of weekly_hours_worked_before_lsr has a missing value fail work-requirement tests. This PR departs from that for a tax: a worker with wages must have worked, and a household calculator that enters income alone would otherwise show no contribution. A reviewer may prefer the other default.
  • Excluded employees. HRS 392-5(6) leaves out service for the United States government. The model subtracts military_basic_pay from employment_income and keeps a service member's civilian wages. That reads basic pay as part of employment_income, as federal income tax does, since no federal tax formula reads military_basic_pay, and as Virginia's and Indiana's subtractions do. The model has no single convention, though. Hawaii LIHEAP's income list and 15 others add military_basic_pay to employment_income, and Hawaii LIHEAP's parameter calls it an input apart from employment_income. A household that enters basic pay that way has it subtracted from its civilian wages here. Microcosm's concept-coverage fixture lists military_basic_pay as uncovered, so in a microsimulation active-duty pay recorded as wages stays in. Other military pay, such as special pays, stays in too. The model identifies no other federal employees, so it still charges them. HRS 392-5 also leaves out some domestic and family employment, commission-only insurance and real estate agents, majority owners, partners and sole proprietors, and state and county sick leave serves as TDI. The model can't observe those groups, so it treats their Hawaii wages as covered.
  • No uprating past 2026. No 2027 notice exists yet, and no state payroll parameter in the repo uprates.
  • Earlier years. The Hawaii variables and employee_state_payroll_tax ran in each year I tried between 2003 and 2026. The contribution stays at zero before 2004, the first year this PR models, and in 2003 the Hawaii wage doesn't read deferrals, so it needs no federal limit. Further down the chain, existing limits that this PR leaves alone decide which years run. For single adult workers at $10,000 and $100,000 in Hawaii and Texas, employee_payroll_tax raised a missing-parameter error before 2013, and household_net_income and spm_unit_net_income raised one before 2015. With a 15-year-old in the tax unit, employee_payroll_tax raised in 2013 and 2014 as well, in Texas too. The 401(k) limits now start in 2004, so a deferral computed from a desired amount raises in 2003.

Coverage

The notices' averages run from 2004 to 2026, one DLIR notice per year. Live DLIR files cover five of those years, so the other 18 link pinned Internet Archive captures. Years before 2004 hold a zero that marks them as unmodeled, as New Jersey's and Rhode Island's parameters do. #10771, a draft that builds on this branch, adds DLIR's 2003 notice and takes 2000–2002 from DLIR's unemployment insurance taxable wage base under the same reading of 383-61(b). Act 30 of 1999 set the wages test's 52-week window, and the section's one later amendment, Act 100 of 2013, renames a fund, so the test holds across 2004–2026.

I read every notice. A later attempt to save copies hit the Internet Archive's rate limit, so for the pinned years the record is the text I captured on the first read. The 2004, 2012 and 2014 notices came as images, so for those three the record is the figures I transcribed.

Tests

Pull request run 43555976963 at 9f8a944722 passed all 36 jobs on GitHub's merge with main 93444c3052.

  • YAML: 68 Hawaii cases with hand-computed values, across 2003–2026, in hi_employee_temporary_disability_insurance_contribution.yaml and hi_temporary_disability_insurance_wages.yaml. Three of them set the employer's deduction share to 0.5, 0 and 1.5. The previous-year cases enter that year's pay as employment_income_before_lsr, the variable the route reads, as snap_work_requirement_weekly_earnings.yaml does: Simulation moves employment_income inputs there, and the YAML runner builds its simulation without that move. One case enters the previous year's pay on employment_income alone and checks that the route leaves it out, the limit the variable's documentation states. The Python tests enter employment_income, so they cover the move. Other cases cap 53 weeks worked at 52, leave provided 2014 military basic pay out of a short 2015's history, and take premiums, HSA contributions and military basic pay out of 2003 wages. 32 federal cases in elective_deferral_limit.yaml and k401_catch_up_limit.yaml take their expected values from the IRS table. Each year from 2004 to 2018 in which the 402(g)(1) or 414(v) amount changes has a case, and the catch-up cases check ages 60 and 64 at each change from 2004 to 2015.

  • Python: tests/core/test_hi_temporary_disability_insurance.py holds what YAML can't state:

    • the stored notice table, with the statutory formula checked against each parameter and DLIR's formula against each notice;
    • the Hypothesis properties over mixed-state batches;
    • the previous-year differential over wages, weeks and hours;
    • the annual form against the weekly statute;
    • the net income identity;
    • a 2015 check that provided 2014 history counts while the simulation calculates and caches nothing for 2014;
    • a 2015 check that a 2014 deferral calculated first from its desired amount stays out of that history.

    tests/core/test_payroll_contributions.py gains a Hawaii row, and tests/core/test_state_payroll_premium_deduction_share_invariants.py now draws Hawaii workers. All three run in CI's core group.

  • Changed existing tests: test_first_modeled_year_is_the_backdating_floor used the 401(k) limit as its parameter first dated after 2015. The backfill moves that date to 2004, so the test now scans the parameter tree instead of naming one parameter: no leaf starts after the floor, and a leaf that starts there has no value the day before. The scan of 127,034 leaves took 0.18 s here. The Hawaii case in marginal_tax_rate_including_health_benefits.yaml now expects −0.0256 instead of −0.0306. The model gives −0.0256 with the contribution and −0.0306 with it neutralized. The case's 0.01 margin would have passed either value; I updated it so that the file records what the model computes. test_state_payroll_premium_deduction_share_invariants.py adds Hawaii to its optional states and programs and to the states whose employer amount stays at zero, so each example also forces a $60,000 and a $250,000 Hawaii worker.

  • CI's load (Linux, run 42971554028 at c1e0406): the Hawaii module runs in the first Python shard of the core group, where its 51 cases took 14.8 s of the shard's 29:00. In the run at 23f238d, cancelled by the next push after that shard finished, they took 12.6 s of 23:56. The shard peaked at 8.35 GiB resident in both runs. Here the three core shards peaked at 6.78 to 10.86 GiB, against 6.84 to 11.08 GiB in four other PRs' runs from the same day, where the planner put the same files in different shards. In this run the files after the module in its shard took a median 12% longer than in those four runs, and the files before it 5% longer. In the run at 23f238d, the files after the module ran no slower than the ones before it. So I read the gap as runner variance. The Hawaii row in test_payroll_contributions.py and the Hawaii workers in the share invariants test run in the second core shard, where those files took 121.8 s and 1.3 s.

  • Local cost, not Linux CI (macOS): at a host load near 160, local times say little. Together in one process, the Hawaii module and the IRMAA module peaked at 3.51 GB resident. At load about 30, round 1 measured the Hawaii module at 45.3 s real and 1.88 GB with its property test, against 46.2 s and 1.80 GB without it. Before main modeled Minnesota's employer premium, the share invariants test's call took 8.8 s with Hawaii and 10.0 s without, and the file's peak resident size went from 1.65 GB to 1.68 GB, at load 15 to 20.

Microsimulation impact

We estimate that this branch raises employee_state_payroll_tax by $161.6m in 2026 on the default dataset, for 975 sampled tax units (505,633 weighted). We compared main at 3f6e604b14 with this branch at 9f8a944722.

Output Main total Change Records changed (weighted)
employee_state_payroll_tax $36.63bn +$161.6m 975 (505,633)
state_income_tax $577.2bn $0 0 (0)
income_tax $2,594.9bn $0 0 (0)
household_benefits $2,060.4bn $0 0 (0)
household_net_income $17,199.2bn -$161.6m 618 (373,758)

impact: employee_state_payroll_tax +$161.6m in 2026 (975 tax units, 505,633 weighted)

axiom: TheAxiomFoundation/rulespec-us#2052 queued, extended with a comment (HRS 392-43, 392-3, 392-5, 383-11 and 383-61(b), with DLIR's yearly notices; proposed corpus citations us-hi/statute/392-43, us-hi/statute/392-3, us-hi/statute/392-5, us-hi/statute/383-11 and us-hi/statute/383-61). It also covers the federal limits under 26 USC 402(g) and 414(v), proposed us/statute/26/402/g and us/statute/26/414/v. Missing corpus text blocks the signed encoder for Hawaii and, in the corpus copy I could read, for 402(g) and 414(v). Each billed encoder run needs separate approval.

🤖 Generated with Claude Code

MaxGhenis and others added 2 commits November 6, 2026 06:30
HRS 392-43 lets an employer withhold half the cost of coverage, up to 0.5% of
weekly wages, and caps those wages at a maximum weekly wage base the Director
publishes each year. Model the contribution at that limit from the published
2004-2026 wage bases, and add it to employee_state_payroll_tax so it reaches
employee_payroll_tax and household net income.

HRS 392-3 takes its wage exclusions from HRS 383-11, the same list Hawaii
unemployment tax uses, so wages start from payroll_tax_gross_wages. That
variable raised a TypeError before 2010 because the FICA pre-tax list started
in 2010, so the list now applies to every year.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The contribution read the federal payroll wage base, which rests on federal
rules Hawaii did not enact. Hawaii's text leaves out what an employer pays to
medical and retirement plans, including salary reductions, and lists no
dependent care payment. Build wages from those two sections, record the
reading as a source-conflict candidate with short excerpts, and restore the
shared FICA list's start date, which only the federal base needed.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
MaxGhenis and others added 4 commits November 9, 2026 08:01
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Leave military basic pay out of Hawaii wages under HRS 392-5(6), let the
wages test pass on the previous year's weeks and hours, and take
retirement deferrals out from 2004. Extend the federal elective deferral
and catch-up limits back to 2004 from the IRS table, which also corrects
2015-2017 to $18,000. Add the program to programs.yaml, split the
source-conflict regulation captures by page, and update the Hawaii
marginal rate case for the contribution.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Main's #10285 rewrote the documentation of employee_state_payroll_tax.
The resolution keeps its text and lists Hawaii TDI with New York's
contributions, which the employer is authorized but not required to
collect.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
HRS 392-43(a) permits the employer to withhold the contribution without
requiring it, like New York's PFL and DBL. Following main's #10285, the
employee amount is the 0.5% limit times state_payroll_premium_deduction_share,
which defaults to 1. Hawaii joins the programs whose employer cost the model
doesn't hold, in the documentation of employee_state_payroll_tax,
employer_additional_state_payroll_tax and the share input, and in the share
invariants test. Three YAML cases cover a share of 0.5, 0 and 1.5.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
MaxGhenis and others added 3 commits November 10, 2026 23:56
The previous-year route of hi_temporary_disability_insurance_qualified now
also needs that year's wages, less the payments the Hawaii wage definition
leaves out, to reach the $400 minimum, so a prior year of active-duty pay or
self-employment no longer passes. The route reads
employment_income_before_lsr, so it calculates no labor supply response for
the earlier year, and before 2015 it reads only values provided for that year.

Also:
- describe the model's two military_basic_pay conventions in the Hawaii wage;
- describe Hawaii's lesser-of limit in the deduction share input;
- add federal limit cases for each year from 2004 to 2018 in which the
  402(g)(1) or 414(v) amount changes and no case existed;
- make the backdating-floor probe scan the parameter tree;
- cite IRC 3306 in the source-conflict candidate and narrow its handbook
  claim;
- enter previous-year pay in YAML as employment_income_before_lsr, because
  the YAML runner builds its simulation without Simulation's input move.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Main now models the Minnesota Paid Leave employer premium (#10294), so the
conflicts keep Hawaii in the lists of states whose employer cost is not
modeled and drop Minnesota from them.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
HRS 392-43(b)(2) sets the maximum weekly wage base from the State's
average annual wage as HRS 383-61(b) determines it, and 383-61(b)
rounds that average to the nearest hundred dollars. DLIR's TDI notices
apply 1.21 / 52 to the average before rounding, so each published base
differs from the statutory one by up to $1.15.

Follow the statute: each year's base is the notice's average rounded to
$100, times 1.21 / 52, rounded once to the cent (2026: 64,500 * 1.21 /
52 = 1,500.87, where DLIR publishes 1,500.21). Record the choice as
source conflict us-hi:tdi-wage-base-rounded-average in
source_conflicts/candidates.yaml with excerpts of both sections and the
2026 notice, and update the YAML cases, the Python property tests and
the changelog to the statutory bases.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
MaxGhenis added a commit that referenced this pull request Nov 11, 2026
#10677's previous-year route nets the payments excluded_payments.yaml lists
for the previous year. This branch starts the deferral exclusions in 2003, so
a 2004 simulation now takes a provided 2003 401(k) deferral out of 2003 wages.
Two 2004 cases cover it: one fails the $400 minimum and one passes at exactly
$400.

Also:
- pin a 2003 batch in test_contribution_properties with wages above the
  annual cap and just over it, so the scanned notice's year always runs
  capped;
- say in excluded_payments.yaml that 2003 is the first year the model holds
  the federal limit, not the first year of the limit;
- note on the 2002 wages case why the model keeps deferrals that HRS 383-11
  left out.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
MaxGhenis added a commit that referenced this pull request Nov 11, 2026
#10677 now stores the statutory base: the average annual wage rounded to
$100 under HRS 383-61(b), times 1.21, divided by 52, rounded once to the
cent. Move 2003 to the same reading: 702.73, so the capped 2003 case
expects 36,541.96 of wages and a 182.71 contribution. NOTICES keeps DLIR's
published 2003 figures, and us-hi:tdi-wage-base-rounded-average now runs
from 2003.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
MaxGhenis and others added 3 commits November 11, 2026 16:15
The qualification docs now say the previous-year route reads
employment_income_before_lsr, that pay set on employment_income with
set_input after the simulation is built stays out of it, and that
before 2015 a deferral counts only when entered as the contribution
itself. The source-conflict entry now describes DLIR's 2010 and 2011
UI base as 90% of the average rounded again to $100, which fits
either rounding reading.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Ten 401(k) catch-up cases check ages 60 and 64 at each catch-up change
from 2004 to 2015. Hawaii cases now cap 53 weeks worked at 52, leave
provided 2014 military basic pay out of a short 2015's history, check
that previous-year pay entered only on employment_income stays out,
and take premiums, HSA contributions and military basic pay out of
2003 wages. The qualification docs' last sentence is rewrapped.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Before 2015 the previous-year wages test read any array stored for
2014, so a value the simulation had calculated for 2014 first, such
as a 401(k) contribution from its desired amount, could change the
2015 answer. It now reads a 2014 array only when the holder records
that period as an input, which makes the result independent of
calculation order. A Python test checks both orders.

Also rename the previous-year YAML case so it reads as a model limit,
not a rule of Hawaii law, and rewrap the source-conflict note.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@MaxGhenis

MaxGhenis commented Nov 11, 2026 •

Copy link
Copy Markdown
Contributor Author

Hand-off to the US + core hub.

Head: 9f8a944. Not a draft, MERGEABLE at hand-off. CI: running at hand-off (3 passed, 23 pending, none failed). The previous heads passed 36 of 36 (c1e0406) and 25 with 2 pending and none failed when last read (74ef2d1).

Reviews (all independent of the author session):

  • Round 3, at c1e0406. Two reviews ran as in-session agents served by claude-opus-5-5:
    • verification: VERDICT: AGREE, with five non-blocking findings;
    • mutation: MUTATIONS: NO BLOCKING SURVIVORS, with non-blocking survivors.
  • 74ef2d1 answered them:
    • the qualification docs now state the set_input-after-build and pre-2015 deferral limits;
    • the source-conflict entry describes DLIR's 2010–11 UI base correctly;
    • 14 new YAML cases kill the survivors.
  • Round 4, at 74ef2d1: a re-check that re-ran each mutant gave VERDICT: AGREE: the intended case killed each of the 14 mutants, and every hand value checks. It raised four non-blocking points.
  • 9f8a944 fixes all four:
    • The pre-2015 previous-year route now reads only input periods (Holder.get_input_periods), so a 2014 value the simulation calculated first, such as a 401(k) contribution from its desired amount, no longer changes the 2015 answer. A new Python test covers both calculation orders; it fails on the old code and passes on the new.
    • Invariant 8 in the PR body now covers only the input paths the route reads.
    • One YAML case has a new name, and one note has new line breaks.
  • Round 5, at 9f8a944: VERDICT: AGREE, with no findings. It confirmed that inputs from the YAML runner and from Simulation register as inputs, that branches behave, and that the new test kills a mutant reverting the read. One edge needs no change: before 2015, 2014 pay or exclusions that would carry or uprate from a 2013 input stay out. That is conservative and matches the documentation.

Reports: ~/reviews/hi-tdi-2026-11-06/review_r3fv_out.md, review_r3fm_out.md, review_r4_out.md and review_r5_out.md, with the run log in runs/r3/MANIFEST.md.

Round 4 and round 5 ran on claude-opus-5-5 as an in-session agent. Subfleet had no open lane, and a first attempt on Fable stopped at the account's Fable limit before doing any work.

Local runs (macOS):

  • At 74ef2d1: federal catch-up and deferral YAML, 59 passed.
  • At 9f8a944:
    • Hawaii YAML: 68 passed.
    • The two pre-2015 Python tests: passed.
    • test_source_conflicts.py: 103 passed.

After merge (hub):

Related PRs:

Follow-up offered to Max as a session card, not started: unify PE-US's military_basic_pay convention. This PR reads it as the part of employment_income that is basic pay, as federal income tax and the Virginia and Indiana subtractions do. Hawaii LIHEAP reads it as an input apart from employment_income.

No Max decision is pending on this PR. The author session owes nothing more.

@MaxGhenis MaxGhenis closed this Nov 14, 2026
@MaxGhenis MaxGhenis reopened this Nov 14, 2026
@MaxGhenis MaxGhenis closed this Nov 14, 2026
@MaxGhenis MaxGhenis reopened this Nov 14, 2026
@MaxGhenis MaxGhenis closed this Nov 14, 2026
@MaxGhenis MaxGhenis reopened this Nov 14, 2026
@MaxGhenis
MaxGhenis merged commit 5768d87 into main Nov 14, 2026
113 of 117 checks passed
@MaxGhenis

Copy link
Copy Markdown
Contributor Author

US + core hub merge audit, us#10677 at 9f8a944722c074c2ce5fe12c4e0d17a3bf00bc6c (squash): Add Hawaii's employee temporary disability insurance contribution under HRS 392-43, from 2004.

  • Reviews:
    • The PR has no GitHub reviews. Five rounds of independent review ran outside GitHub.
    • An independent review requested changes at 683f457d0b, and a second at deca0e3638, with two blocking findings.
    • At c1e0406d83, a verification review agreed, with five non-blocking findings, and a mutation review found no blocking survivors. The verification review also confirmed the earlier blocking findings fixed.
    • 74ef2d1e81 answered them. The wages test's documentation now states its set_input and pre-2015 deferral limits, and the source-conflict entry describes DLIR's 2010–2011 unemployment insurance base correctly. 14 new YAML cases kill the surviving mutants.
    • A re-check at 74ef2d1e81 agreed. It re-ran each mutant, found that the intended case kills all 14, checked every hand value and raised four non-blocking points.
    • 9f8a944722 fixes all four. Before 2015 the previous-year route reads only input periods, so a 2014 value the simulation calculated first no longer changes the 2015 answer. A new Python test fails on the old code and passes on the fix. Invariant 8 in the description now covers only the input paths the route reads, one case has a new name, and one note has new line breaks.
    • A re-check at 9f8a944722 agreed with no findings. It confirmed that inputs from the YAML runner and from Simulation register as inputs, and that the new test kills a mutant that reverts the read. The reviews from c1e0406d83 on ran on Claude Opus 5.5, each as an agent separate from the author.
    • Before merge, the description drops a search for pre-2004 notices that Extend Hawaii's TDI contribution back to 2000 #10771 has since answered. It also presents DLIR's unemployment insurance base as consistent with either reading of HRS 383-61(b), since that section's first sentence rounds the base itself.
  • Gates checked live at merge:
    • gh pr checks exits 0;
    • the latest Pull request run at the head, 43555976963, passed and tested this head merged into main at 93444c3052;
    • MERGEABLE;
    • not a draft;
    • no reviewer's latest review requests changes;
    • the semantic-overlap guard: since 93444c3052, main has changed two files this PR touches, policyengine_us/programs.yaml and source_conflicts/candidates.yaml. In the tree that merges main at a93cc58d7e with the head, the PR's lines in each file equal its own diff, candidates.yaml holds 449 entries with unique ids, and programs.yaml holds 111 programs, also with unique ids;
    • the merge pins the head.
  • Impact (default dataset, 9f8a944722 against main at 3f6e604b14): we estimate that the change raises employee_state_payroll_tax by $161.6m in 2026, in 505,633 tax units (975 sampled), by at most $914.86 each. Household net income falls by the same $161.6m. Income taxes and household benefits don't change.
  • Next:
  • Axiom: TheAxiomFoundation/rulespec-us#2052 queued, now with companion rows for the cases added since c1e0406d83.

🤖 Generated with Claude Code

@MaxGhenis
MaxGhenis deleted the hi-tdi-employee-contribution branch November 14, 2026 05:33
MaxGhenis added a commit that referenced this pull request Nov 14, 2026
The merge has no conflicts. Both sides edit source_conflicts/candidates.yaml
and test_source_conflicts.py in separate places. Main adds five candidates
(#10677, #10783, #10847, #10815) and one KNOWN_UNMARKED line (#10783); this
branch keeps its New Jersey candidate and its three KNOWN_UNMARKED lines.

Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant