Count capital gains once in Arizona property tax credit income

[MaxGhenis]

Arizona Form 140PTC household income now counts capital gains and qualified dividends once, includes dependents' signed income, applies the $1,500 capital loss limit separately to each member, and uses zero when looking up the credit schedules for negative household income. #9770 is folded into this PR. The PR remains draft.

Methodology ruling: terminating estates and trusts

Max's 2026-10-09 ruling d1107 resolves the sole P2: keep as carryover. A terminating trust's loss passed to a beneficiary through Form 1041 K-1 box 11 code D remains in long_term_capital_loss_carryover and is added back when calculating Arizona line D. Reg. 1.642(h)-1(a) transfers the excess as a capital loss carryover; no cited Arizona source provides a different treatment. This round records the ruling and pins the existing calculation without changing model formulas.

Form 1041 instructions, box 11 codes C/D use the carryover classification. Arizona ITR 12-1, items (5) and (7) counts current-year K-1 income and excludes prior-year capital losses. R15-2C-502(C)(3) and the 140PTC line D instructions do not expressly distinguish terminating-trust carryovers from a beneficiary's own carryovers.

The ruled regression represents a 70-year-old Arizona renter paying $5,000 rent, with $4,000 personal gains and a $3,000 terminating-trust loss already netted into federal long-term gains of $1,000. The $3,000 memo is added back: line D and household income are $4,000, and the credit is $0. The previous review's alternate $1,000-income/$502-credit expectation is replaced with this ruled contract, with a comment citing Reg. 1.642(h)-1(a) and d1107. No deliberately failing or xfailed alternate remains.

Separate follow-up, tracked in #10021: split long_term_capital_loss_carryover into own-carryover and trust-passthrough leaves. The current memo cannot distinguish those origins. That input refinement is outside this ruling round and does not alter the carryover treatment adopted here.

Design and sources

• ARS 43-1072(H)(5)-(6), (I), R15-2C-502, ITR 12-1 and Form 140PTC govern household income. Every member's separately determined income counts; Social Security is excluded.
• Person-level line D combines capital_gains, non_sch_d_capital_gains and the positive long_term_capital_loss_carryover memo, then floors the result at minus the member's $1,500 limit. Long-term inputs already net the memo. Capital gain distributions are combined with sale gains/losses by inference from 26 USC 852(b)(3)(B).
• Non-dependent income starts from federal AGI, removes taxable Social Security and the capital amounts actually included by the federal source/deduction lists, and adds those members' line D amounts. The capital deduction reversal clips loss_ald - limited_business_loss between zero and the two capital deduction components. Reform guards remain intact.
• Dependent income includes signed federal sources other than taxable Social Security and the capital sources, then adds each dependent's line D. Business/rental/estate losses remain signed. Tax-exempt interest and pensions count for all members.
• Household income is not floored; credit schedules receive max(income, 0) as the line J instructions require.
• The PR extracts limited_business_loss from federal loss_ald, which then adds that component and its two capital components. The d1107 round does not change these formulas.

Delta review P3 work and remaining precision limitation

The previous repair consolidated the two release fragments into one, preserving the capital loss limit, dependent income/loss and negative-income release notes. It also documented the Arizona input convention: short-term amounts contain the current year's net sale gains/losses and exclude Schedule D line 6 carryovers. The form includes those carryovers in line 7; that net amount cannot be supplied unchanged for Arizona. No short-term carryover memo exists; adding one remains outside this PR.

The federal extraction differential uses production variable classes in one minimal system and fresh vectorized simulations for 2021, 2025 and 2026. An independent sections 461(l)/1211(b) reference casts the complete deduction once, as in the unextracted main formula after #9788. Each year covers 120 seeded units and 15 all-status threshold cases, including dependents and opposite-signed capital sources. Exact whole-dollar parity is explicitly bounded to every intermediate total below 2**24.

Federal precision P3 remains open: the extraction affects federal AGI and does not guarantee universal whole-dollar or fractional-dollar parity. Above float32's exact-integer range, sequential component storage/addition can produce dollar differences. Business loss $3, capital losses $16,777,218 and offsetting distributions $16,777,216 produce a once-cast unextracted-main reference result of $16,777,220 and an extracted result of $16,777,222. This comparison uses the main formula after #9788, rather than the earlier PR predecessor that lacked its against-gains component. Another example with business losses totaling $16,777,219 shows why changing only the addition dtype cannot restore universal parity. A wider redesign remains open; the bounded differential does not establish parity outside its stated domain.

Validation

This d1107 round

All nine requested single-file foreground processes exited successfully at 3603ddc030e5d0bae632930ade0f1a28fd7150b6: 224 YAML cases and 10 Python cases passed. The federal differential covers 405 tax units across 2021, 2025 and 2026.

Single-file verification	Result
policyengine_us/tests/core/test_loss_ald_extraction_parity.py	3 passed; 405 federal tax units
policyengine_us/tests/policy/baseline/gov/states/az/tax/income/credits/az_property_tax_credit_income.yaml	38 passed, including d1107's $4,000 income/$0 credit case
policyengine_us/tests/test_az_property_tax_credit_income.py	7 passed; 6 existing warnings from division in unemployment compensation calculations
policyengine_us/tests/policy/baseline/gov/states/az/tax/income/credits/az_property_tax_credit_capital_gains.yaml	6 passed
policyengine_us/tests/policy/baseline/gov/states/az/tax/income/az_property_tax_credit_agi.yaml	8 passed
policyengine_us/tests/policy/baseline/gov/states/az/tax/income/credits/az_property_tax_credit.yaml	12 passed
policyengine_us/tests/policy/baseline/gov/irs/income/taxable_income/adjusted_gross_income/above_the_line_deductions/loss_ald.yaml	11 passed
policyengine_us/tests/policy/baseline/gov/states/az/tax/income/credits/az_property_tax_credit_eligible.yaml	7 passed
policyengine_us/tests/policy/baseline/gov/states/az/tax/income/credits/az_property_tax_credit_schedules.yaml	142 passed

make format passed before commit 3603ddc030e5d0bae632930ade0f1a28fd7150b6: 7,039 files unchanged and Ruff checks passed. Formatting used Python 3.13 with --no-sync, despite the .python-version 3.14 warning. Git whitespace checks passed for both the committed change from c6ad076aaca4507ba288071155295ffbffb44d51 and the staged diff.

Current verification environment: macOS, the user-provided Python 3.13.9 interpreter, policyengine-core 3.32.15 and pytest 9.1.1.

Current-source provenance: the untracked source-read cache serves the committed Git archive at 3603ddc030e5d0bae632930ade0f1a28fd7150b6 (SHA256 6a07b309a859513efaef88a505fb476b9204f31ac1f577909bfc2c12291e5b45) without altering model source or calculations. The first two processes used the verified five-package dependency cache. The remaining seven used its verified 44-package extension, containing 10,355 cached entries. Cached wheel bytes were checked against installed RECORD hashes; three entries with different RECORD hashes were omitted from the extension and were not overridden. This does not claim fresh byte comparisons against every installed file. Installed dependency files and native extensions were unchanged.

Each test file ran separately in the foreground. No new microsimulation, full-suite result, CI pass, or performance benchmark is claimed.

Each YAML file is invoked separately with python -m policyengine_core.scripts.policyengine_command test <file> -c policyengine_us; each Python file is invoked separately with python -m pytest -q <file>. This round runs every PR-changed test file and the Arizona property tax credit YAMLs. No folder suite, full suite or new microsimulation is run. Existing Arizona property/differential coverage is preserved. No partner contract test is edited.

Previous repair at c6ad076

The preceding round reported 108 YAML cases and 10 Python cases passing, plus two upper-range precision evidence cases. These are previous results, not new d1107 verification:

Previous single-file verification	Previous result
Original Arizona property/differential module	7 passed, 6 existing divide warnings
Bounded federal extraction differential	3 passed, 405 tax units across three years
Arizona household-income YAML	37 passed, 1 warning
Arizona capital-gains YAML	6 passed, 1 warning
Arizona AGI and credit-schedule YAML files	8 and 12 passed; one pytest plugin warning per invocation
Federal loss, dependent-loss and threshold YAML files	11, 11 and 23 passed; one pytest plugin warning per invocation
Upper-range precision evidence outside CI	2 passed, each reproducing a once-cast reference of $16,777,220 versus the extracted model's $16,777,222
Review's alternate terminal-trust expectation	Historically reproduced $4,000 rather than the alternate $1,000; superseded by d1107's $4,000/$0 contract
Required formatting and whitespace checks	make format passed before each change commit; Git whitespace checks passed

The federal differential passed against unchanged production calculations at 57eceac99d6fb0561daad75638a341c30e868b76; adding that test was the only tree difference from c6ad076aaca4507ba288071155295ffbffb44d51. Its source matched the committed test byte-for-byte (SHA256 9cffe62437b81e7cb119f74c9cb02d5f31fd8dc7b85d3bdc8eeb708979d00f6a). Remaining preceding-round verification used the final-head source archive (SHA256 c2059444156c9dad955d75babbd3904823477c45ceba7a4e371e537fa8764568). The original Arizona property module was preserved byte-for-byte (SHA256 93391bf1641b21c78d99f85f4246fb61dd54a2ffd1094e74281b8726216e543f).

The preceding environment was macOS, Python 3.13.9 and policyengine-core 3.32.15. Stalled filesystem reads required an untracked source-read facade serving the exact committed Git archive with matching directory inventories and frozen source introspection. A dependency cache served pinned wheel bytes matching all 6,181 installed RECORD hashes, with zero mismatches; that does not claim fresh byte comparisons against every installed file. It changed source delivery, not calculations. Later process starts reused compiled code for immutable archived source while executing fresh module globals and classes, and disabled pytest disk-cache bookkeeping. Interrupted imports and a diagnostic-timeout crash were infrastructure evidence, not policy test passes or failures. The historical terminal-trust assertion completed after those startup problems were resolved.

Previous local test-cost measurements included setup and source-cache initialization. Collection of the new federal module took 32.51 seconds, with peak RSS 1,279,049,728 bytes; executing its three minimal-system cases took 60.55 seconds, with peak RSS 1,287,880,704 bytes. This is a collection-versus-execution comparison, not a CI-group benchmark or reliable estimate of incremental CI cost. The original Arizona property module took 67.29 seconds, with peak RSS 1,446,739,968 bytes. The income YAML's 37 assertions took 356.64 pytest seconds, 1,081.05 full-process seconds and 817.71 facade seconds, with peak RSS 5,913,198,592 bytes. The capital-gains YAML's six assertions took 37.87 pytest seconds, 219.11 full-process seconds and 201.79 facade seconds, with peak RSS 2,449,096,704 bytes. Facade timing started before cache initialization and stopped during atexit, so it did not measure the full process lifetime. Different cache settings prevent inference of incremental test cost. These isolated macOS measurements do not establish Linux CI group memory or timing. The new module routes to the existing rest-python core group, constructs one minimal read-only system and uses fresh simulations per year; no CI jobs or runner budgets change.

Invariants

The existing seeded Arizona population and every original property assertion remain unchanged. Arizona comparisons allow the original numerical tolerances; the bounded federal differential requires exact array equality:

1. Household income equals an independent Form 140PTC line A+B+D+E+F sum over every member, excluding Social Security, netting rental losses and applying the member-level line D limit.
2. Raising a head's or dependent's gains changes income by exactly the change in that member's line D; the credit does not increase beyond the original numerical tolerance.
3. Further losses below minus $1,500 do not change that member's line D or household income.
4. Credit lookup depends on household income through its maximum with zero.
5. Arbitrary changes to the federal preferential-rate amount do not change Arizona household income.
6. The federal differential verifies exact agreement with a once-cast reference for whole-dollar inputs when all intermediate totals remain below 2**24, with dependent exclusions, filing-status limits and source offsets preserved. It does not claim parity outside that domain.
7. Under d1107, a terminating-trust K-1 box 11 code D loss remains a carryover and is added back to the beneficiary's net long-term gains for Arizona line D.

Previously measured hub microsimulation impact

The hub ran a real microsimulation on the default dataset comparing main at merge base 7b956eb4014e3a189075c79f9b1f0a8f1d3b02ab with PR head c6ad076aaca4507ba288071155295ffbffb44d51, for 2025 and 2026. Runs proceeded one at a time under the shared heavy-job lock. Sums are weighted. These are measurements of those exact heads; no new microsimulation is run in the d1107 round, which changes tests and documentation without changing model formulas.

Output	2025 change	2026 change	2025 records changed (weighted)	2026 records changed (weighted)
az_property_tax_credit	+0	+0	0 (0)	0 (0)
az_income_tax	+0	+0	0 (0)	0 (0)
state_income_tax	+0	+0	0 (0)	0 (0)
income_tax	+0	+0	0 (0)	0 (0)
household_net_income	+0	+0	0 (0)	0 (0)
household_benefits	+0	+0	0 (0)	0 (0)

Hub comparison artifacts are retained separately. These zero changes do not establish universal float32 parity or exercise supplied terminating-trust K-1 losses. This round does not obtain a new partner-output comparison, full CI result or Linux CI performance measurement.

Earlier author-reported microsimulations compared branch 2ea9d6bced with main d3978dfc31 using populace_us_2024.h5@populace-us-2024-spm-20260909, policyengine-core 3.32.16:

Earlier historical measurement	2025	2026
Tax units with changed credit	0 of 79,729	0 of 79,729
Weighted credit, both historical heads	$40,666,618	$40,957,022
Units with changed household income	409	409
Weighted household income change	−$32.3bn	−$33.4bn

Those earlier results describe their stated historical heads, separately from the subsequent hub comparison above. The reported earlier sample contained no capital losses, and a later carryover audit found no carryover column in its examined cached snapshots. Neither audit establishes general carryover equivalence or tests supplied terminating-trust K-1 losses.

Remaining scope and review status

The terminal-trust P2 is resolved by d1107. Splitting the carryover memo into own-carryover and trust-passthrough leaves is tracked separately in #10021; the federal float32 precision P3 remains open and disclosed above. Prior repaired findings and coverage for #9647, #9779, #9788, #9801, #9841 and #9633 remain in the branch. This round does not merge main or resolve unrelated core #570/#572, #9859 d1009 or #9621 partner-gate holds. No independent approval or green-CI claim is made.

Existing limitations remain: unemployment has no paying-state input; nontaxable strike benefits and federally nontaxable alimony are omitted; dependent above-the-line deductions are not subtracted; members outside the tax unit are not combined.

axiom: us-az:statutes/43-1072 household income (H)(4)-(6), (I) TheAxiomFoundation/rulespec-us#1483 queued (existing companion tests A-M; d1107 adopted in this round).

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 100.00%. Comparing base (909176a) to head (21cb189).
⚠️ Report is 222 commits behind head on main.

Additional details and impacted files

@@            Coverage Diff            @@
##              main     #9762   +/-   ##
=========================================
  Coverage   100.00%   100.00%           
=========================================
  Files            4         5    +1     
  Lines           76        93   +17     
  Branches         2         5    +3     
=========================================
+ Hits            76        93   +17     

Flag	Coverage Δ
unittests	100.00% <100.00%> (ø)

Flags with carried forward coverage won't be shown. Click here to find out more.

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

[MaxGhenis]

#9762 fails tests when merged with current main. While checking #9779 (Schedule 2 thresholds) against this PR, I found that this head (7491794) merged with current main (6b0e773) fails:

• 3 YAML cases in az_property_tax_credit_income.yaml:
  • "A dependent's capital loss offsets the dependent's wages"
  • "A dependent's capital loss with a claimant's pension"
  • "A dependent's rental loss offsets the dependent's wages"
• all 3 tests in test_az_property_tax_credit_income.py. In test_household_income_matches_form_140ptc_lines, 2 of 60 sampled units are off, for example 7,921.55 vs 7,065.14 desired.

I bisected this to #9647 ("Keep dependents' losses off the filer's AGI", 4e78791). This head merged with 4e78791^ passes 16/16 and 3/3; merged with 4e78791 it fails as above.

#9779 itself doesn't touch these files. #9770's head (0450a28) merged with #9779 passes all 651 Arizona YAML tests and both property test files.

[MaxGhenis]

US + core hub merge audit, us#9762 at 3603ddc030e5d0bae632930ade0f1a28fd7150b6 (squash): Arizona property tax credit income counts capital gains once.

• Ruling: Max's d1107 (2026-10-09). A terminating trust's pass-through loss stays a capital loss carryover (Reg. 1.642(h)-1(a)). The leaf-input split is tracked in Split long-term capital-loss carryover inputs by loss origin #10021.
• Reviews:
  • Review r1 found one P2, which was this methodology call.
  • The d1107 round pinned the ruled behaviour without changing executable formulas.
  • Delta review r2 (GPT-6.1 Sol) APPROVES this exact head.
• Gates checked live at merge:
  • gh pr checks exits 0;
  • the latest Pull request run at the head concluded success;
  • MERGEABLE;
  • not a draft;
  • no CHANGES_REQUESTED review;
  • the semantic-overlap guard;
  • head pinned.
• Impact: a hub microsim, merge base 7b956eb401 against c6ad076aac. Every compared output is unchanged in 2025 and 2026. The delta to this head changes no executable code.
• --admin: used only because a required approving review is the sole block.