Separate the dependent portion of 11 states' personal exemptions

[DTrim99]

What this does

Adds Child Poverty Impact Dashboard contrib reforms that separate the dependent portion of each state's per-person personal exemption (or credit) into its own variable + parameters, for 11 states: AR, HI, IN, MD, MI, NE, OH, OK, VT, WI, WV. This mirrors the existing RI/DE/OR/VA reforms and lets the per-dependent amount be adjusted or eliminated independently of the head/spouse exemption (which most of these states bundle together).

Each reform:

• Defaults to a no-op — in_effect=false, and when on, the default amount reproduces current law.
• Exposes gov.contrib.states.{st}.dependent_exemption.{in_effect, amount, age_limit/*} (AR uses dependent_credit; MD/OH also get a phaseout group).

How each separates

Group	States	Approach
Bundled per-person exemption → head/spouse + dependent	HI, IN, MI, NE, OK, VT, WI, WV	Recompute the personal count excluding dependents; add a separate {st}_dependent_exemption. HI preserves its aged add-on; IN leaves the existing per-child additional exemption untouched.
AGI-stepped per-person amount	MD, OH	Dependent amount defaults to the baseline income-stepped schedule via a negative-amount sentinel (so default = no-op); a non-negative value applies a flat per-dependent amount, plus an optional income phase-out.
Shared credit base	AR	Redirects the existing person-level dependent slice of the personal credit to a contributed amount (its base is otherwise shared with head/spouse). No age limit — it's summed via adds, which only reflects uniform per-dependent amounts.

Indiana's module directory is the Python keyword in, so it is imported dynamically in reforms.py and uses getattr for parameter access.

Tests

Each state has YAML tests (tests/policy/contrib/states/{st}/) covering: default no-op for a family with children, amount reduction / elimination (amount: 0), childless-filer no-op, and (exemption states) age-limit fallback. All pass. The system smoke-imports cleanly with all 11 registered in create_structural_reforms_from_parameters.

Follow-up

Once released, the dashboard flips these states to amount-editable (separate PR). Companion: #8691 (SC dependent-exemption parameter).

🤖 Generated with Claude Code

[PavelMakarchuk]

Program Review — Separate the dependent portion of 11 states' personal exemptions

Scope: Child Poverty Impact Dashboard contrib reforms, 11 states (AR, HI, IN, MD, MI, NE, OH, OK, VT, WI, WV) · three separation patterns (bundled split; AGI-stepped sentinel + phaseout; shared credit-base redirect)

A large but disciplined PR. All four review dimensions cleared the implementation: every state defaults to a verified no-op, cross-state isolation holds (no copy-paste/wrong-state references — the main risk in a multi-state PR), variable overrides match their baselines exactly, and all 55 YAML cases pass. Requesting changes to tighten the test suite before merge — the gaps are in what the tests prove, not in the code.

Requested changes

1. Add an end-to-end tax/net-income assertion (systemic — affects all 11 states). Every test currently asserts the reform's own exemption/credit-total variable (e.g. md_total_personal_exemptions, hi_regular_exemptions, ar_personal_credits) — one level above the per-dependent variable, which does prove the head/spouse split arithmetic. But no test asserts state taxable income, state tax, or net income, so the exemption→tax wiring is unproven end-to-end. AR is most exposed (it asserts only a credit total, never a resulting AR tax/net-income figure). Please add at least one downstream-outcome case — ideally for AR plus one bundled-split state and MD/OH.

2. Add the joint 2-adult case to NE, OK, VT, WI, WV. These five states have 4 cases and omit the joint two-adult + dependents case that HI/IN/MI/MD/OH/AR include. That case is precisely what catches off-by-one / sign-flip errors in the personal_count = size − dependents + over-age dependents formula — the risk the PR description itself flags. Add a joint 2-adult + 2-dependent case to each, mirroring HI/IN/MI.

Suggestions (non-blocking)

• Re-run the canceled "Quick Feedback (Selective Tests + Coverage)" job — it was canceled by an infra runner-shutdown, not a real failure (the Full Suite - Contrib (states-shard-*) jobs that run these tests all passed).
• Several count variables are typed float / unit = USD (should be int/count); unused instant import in ~7 reform files.
• Most contrib params omit reference — consistent with the merged DE/RI/VA/OR siblings, so optional; a one-line provenance comment on each amount.yaml noting it mirrors the state baseline would make the derivation explicit.
• Minor consistency: inconsistent state-level __init__.py presence (works via namespace packages); WI/WV test values lack numeric underscores; effective-date style varies (0000-01-01 vs 2025-01-01) for inert defaults.

Validation Summary

Check	Result
Regulatory Accuracy	✅ No-op fidelity verified for all 11; bundled-split arithmetic, MD/OH sentinel, and state-specific preservations (HI aged add-on, IN per-child exemption, AR credit base, MI stillborn, OK extras, WV size==0) all correct; isolation confirmed
Reference Quality	✅ All 11 default amounts trace to state baselines; metadata complete & consistent with siblings
Code Patterns	✅ All update_variable overrides match baseline name/entity/period/value_type; non-negative fallback counts; IN dynamic import correct; 5-year gating loop references each state's own parameter — no cross-state leakage
Test Coverage	⚠️ No-op + elimination + special mechanics covered; gaps: no downstream-tax assertion; 5 states lack joint case
CI Status	⚠️ One job canceled by infra (runner shutdown) — not a code failure; contrib state shards passed

Review Severity: REQUEST_CHANGES (approve-leaning — gaps are test-completeness only)

To auto-fix: /fix-pr 8696

🤖 Generated with Claude Code

[DTrim99]

Fixes Applied (full scope)

Addresses PavelMakarchuk's review. CI: the prior "Full Suite - Rest" failure was a runner cancellation, not a test failure.

Blocking review items

• Downstream-outcome assertions — tests no longer stop at intermediate exemption/credit variables; added state taxable-income / tax-reducing assertions for AR (ar_non_refundable_credits), MI (mi_taxable_income), MD (md_exemptions), and OH (oh_taxable_income), confirming the dependent-exemption change flows through.
• Joint 2-adult cases — added married-filing-joint cases for NE, OK, VT, WI, WV (HI/IN/MI/MD/OH/AR already had them), verifying the head/spouse-vs-dependent count split has no off-by-one.

Suggestions

• Age-limit boundary tests — age-17 (eligible) / age-18 (excluded) cases for all 10 age-limit states.
• AR reference — added AR1000F instructions (Line 7A, where $29 appears) plus statute subsections, mirroring the AR baseline parameter.
• WI/WV test literals — added missing numeric underscores.
• Module cleanups — removed unused instant imports (7 reform modules); added the missing ne/wi re-export __init__.py; typed the dependent-count helper variables as int (dropped the spurious USD unit).

Verification

• Reform modules + reforms.py import cleanly (py3.11).
• Local contrib test run for all 11 states was passing (no failures); CI on this PR is the authoritative gate.
• No duplicate test cases; no parameter/reform-logic changes beyond the cleanups above.

[PavelMakarchuk]

Thanks for the updates — the joint 2-adult cases for NE/OK/VT/WI/WV (plus the age-limit boundary cases) and the downstream taxable_income assertions for MI and OH are exactly what was needed, and all 11 suites pass.

One remaining item from the downstream-assertion request: AR still asserts only ar_personal_credits (a credit total) — no resulting AR income-tax / net-income figure. AR is the one state using the shared-credit-base-redirect pattern (the credit is consumed via adds), so it's the most exposed to a wiring issue between the separated dependent credit and the actual tax — and it's the only pattern with no end-to-end assertion.

Requested change: add one downstream AR case asserting ar_income_tax (or net income) so the dependent-credit → tax wiring is proven end-to-end, mirroring the MI/OH taxable_income cases. A no-op default case plus an amount: 0 case showing AR tax rises would be ideal.

Everything else is good to go — this is the last blocker from my side.

Note: the failing "Quick Feedback" check is a transient runner-shutdown (infra), not a real failure; recommend re-running. 🤖

[DTrim99]

AR end-to-end income-tax assertion added

Good call — AR was the one state (shared-credit-base-redirect pattern, credit consumed via adds) lacking a downstream tax assertion. Added a default vs amount: 0 pair on ar_income_tax (mirroring the MI/OH taxable-income cases):

• Default (no-op, credit 87): ar_income_tax = 1346.67
• amount: 0 (credit 29): ar_income_tax = 1404.67

The tax rises by exactly $58 when the dependent portion is removed — matching the credit drop from 87 → 29, proving the dependent-credit → tax wiring end-to-end. All 9 AR cases pass locally (py 3.11).

On the "Quick Feedback" check — agreed, that was a runner-shutdown/cancellation (infra), not a real test failure; it should go green on re-run.

[PavelMakarchuk]

Review: Separate dependent portion of 11 states' personal exemptions

Thanks for this — the separation approach is sound and faithfully mirrors the RI/VA/DE precedent. The split math (head/spouse slice + dependent slice = original per-person total) checks out algebraically and against the no-op tests for all 11 states, and the reform correctly defaults OFF. Requesting changes on a few items before merge, the first being a genuine correctness gap.

Must address

1. MI & NE: the in_effect=true default stops reproducing current law outside 2025.
   mi/dependent_exemption/amount.yaml and ne/dependent_exemption/amount.yaml freeze the per-dependent amount at the single 2025 value (5_800 / 171), but the baseline per-person amounts (mi/.../exemptions/personal.yaml, ne/.../exemptions/amount.yaml) carry uprating: gov.irs.uprating. So when the reform is on, the personal slice uses the live uprated baseline while the dependent slice stays frozen — the "default = no-op" invariant silently breaks for 2026+ and pre-2025. A dashboard sweeping years will hit this. Fix by mirroring VT's uprating on the contrib param (vt/dependent_exemption/amount.yaml), or by adopting the MD/OH -1 sentinel. (AR/HI/IN/OK/WI/WV are safe — flat, non-uprated baselines.)

2. Add references to the 8 value params that lack them.
   hi/in/mi/ne/ok/vt/wi/wv .../amount.yaml have no reference field. Each value exactly matches a baseline parameter that already carries a full authoritative reference — please copy those over so the new file is independently traceable to current law.

3. Cover the reform-activation branch + add downstream tax assertions.
   Tests apply each reform via the bypass=True object, so the non-bypass reform_active = True / return create_…() path never executes (the per-test in_effect: true input is inert — formulas only read p.age_limit.in_effect). Add a parametrized pure-Python test calling each create_<st>_…_reform_fn(parameters, period) with in_effect true and asserting a Reform (not None) is returned. Separately, only AR has an end-to-end income-tax assertion; HI/IN/NE/OK/VT/WI/WV assert the reform variable in isolation. Add a downstream state-tax assertion for those states so the separated dependent portion is shown to reach the tax base.

Should address

4. Missing __init__.py in the ar, md, ne, oh, wi, wv reform dirs — works via namespace packages, but the precedent and the other states ship explicit ones; add for consistency.
5. MD/OH amount.yaml descriptions are two sentences (sentinel explanation) — split to one sentence per the parameter style rule.
6. Inconsistent value-from dates — age threshold uses 2025-01-01 in 8 states but 0000-01-01 in MD/OH. Harmless (gated by in_effect=false) but worth normalizing.

Suggestions

• Adopt the MD/OH where(amount < 0, baseline_per, amount) sentinel across all flat-amount states — it auto-tracks the baseline and removes the MI/NE bug class entirely.
• Consider a boolean toggle instead of the -1 currency sentinel (like the existing age_limit/in_effect), since -1 overloads a currency-USD param.
• No-op tests assert a hand-computed constant; a baseline-vs-reform equality assertion would be more robust.
• Factor out the repeated range(5) lookahead loop (appears 11×).
• AR statute href uses commercial casetext.com; an official source is preferable.

CI note (not blocking)

The failing Quick Feedback (Selective Tests + Coverage) check was an infrastructure runner-shutdown cancel at 10m01s while the selective suite was still passing (only ~27 of 90 tests had run, all green; the coverage step never started and runs --fail-under=0). Re-run the job. If it recurs, the suite is genuinely slow — each structural reform forces a full system rebuild per YAML test — and would benefit from an explicit timeout-minutes or sharding.

---

🤖 Review generated with Claude Code