Model state-specific SNAP BBCE poverty-guideline update months

[hua7450]

Summary

State SNAP broad-based categorical eligibility (BBCE) standards re-base to a new year's federal poverty guidelines on each state's own schedule. The gross-income screen derived its limit from snap_fpg, which follows the federal October fiscal-year cycle, so states that re-base earlier in the calendar year were held a guideline year behind — understating their cutoffs by roughly 2% for part of each year.

For a one-person Washington household in 2026, the limit stayed at $2,608 from April through September where DSHS publishes $2,660 for April 1, 2026 through March 31, 2027.

Fixes #9231

Research

I verified the update month for all 46 BBCE jurisdictions against state administrative codes and SNAP policy manuals. Four re-base off the federal cycle:

State	Month	Authority
Maine	January, on the guideline publication date	10-144 CMR Ch. 301 §999-3 Chart 4 — six consecutive January eras
Massachusetts	February 1	DTA Helpful Charts and Figures — the 200% column reads "Effective 2/1/2026" beside federal columns marked "Effective 10/1/2025"
Oregon	March 1	OAR 461-155-0180(6) — the 200% standard is "updated every March", while the same rule's 100%/130%/165% standards are "updated every October"
Washington	April 1	WAC 388-414-0001(2)(a)(ii)

The remaining 42 follow the federal October schedule, most evidenced by two consecutive October effective dates in their published charts.

This changes results in four states only. A state still on the previous year's guidelines outside those four is correct, not lagging — that is how the federal fiscal-year cycle works.

Approach

• income_limit/fpg_year_start_month.yaml carries each state's month, defaulting to October. Encoding a newly verified state is a one-line parameter edit — no formula changes.
• tanf_non_cash_fpg selects the guideline vintage from that parameter.
• tanf_non_cash_gross_income_limit applies the state's percentage tier to that guideline and returns the exact result. See "Rounding" below.
• meets_tanf_non_cash_gross_income_test compares dollars against that limit rather than an FPG ratio.
• The federal SNAP gross and net income tests are untouched and keep the October schedule through snap_fpg.

Two percentages the research corrected are also dated: the Virgin Islands rose from 130% to 175% effective 2024-10-01, and Arizona from 185% to 200% effective 2026-03-01.

Arizona

An earlier revision dated Arizona's increase 2026-04-13, inferred from a "What's Changed" entry because azdes.gov sits behind a challenge page. The CNAP manual text resolves it:

Starting the benefit month of 03/2026, the gross income limit for the NA Expanded Categorical Eligibility changed from 185% of the FPL to 200% of the FPL.

The same note confirms Arizona re-bases each October ("increased for FFY 2026 starting 10/01/2025"), so its fpg_year_start_month stays at the default. The increase also applies to gross_hheod, which had remained at 185% and so never conferred the new limit on elderly or disabled households.

Rounding

States publish their BBCE standards as whole dollars, and the conventions differ: some round up, some to the nearest dollar, some down, and several apply the percentage to the whole-dollar monthly guideline rather than the exact one (California, Colorado, DC, and Arizona publish $2,610 where the exact computation gives $2,608.33).

An earlier revision of this PR encoded those conventions as per-state parameter lists. That is removed, for two reasons:

1. No federal rule reaches this limit. 7 CFR 273.9(a)(3) directs that the federal gross and net standards be rounded up, and the model applies that in meets_snap_gross_income_test. But categorically eligible households under 7 CFR 273.2(j)(2) are exempt from the §273.9 standards entirely, and the BBCE threshold is the income limit of the state's TANF-funded non-cash benefit. A published chart's rounding is a property of the chart.
2. The encoding was mostly wrong. Of the states where a published figure was checked, only Washington matched. Shipping a taxonomy at that hit rate asserts more than we know.

The limit is now the exact percentage of the guideline, never more than about $2 from any published figure. Washington's April standard is unaffected: the 2026 guideline divides evenly (15,960 / 12 = 1,330), so the computed figure is exactly the $2,660 DSHS publishes, and the exact computation reproduces all ten household sizes on the DSHS chart.

Whether to model per-state chart conventions is tracked in #9255.

Known limitations

• Guam and the Virgin Islands. Both operate BBCE per FNS (165% and 175%, encoded here), but gov/hhs/fpg.yaml carries zero placeholders for the territories, so their computed standard is $0. Because a $0 limit passes a household with no income, and categorical eligibility waives the asset test, a zero-income territory household with substantial assets is now scored SNAP-eligible where the base commit scored it ineligible. The underlying gap is territory poverty guidelines, which also leave snap_fpg at 0 and the federal gross and net tests degenerate; no territory households appear in the microdata, so this is reachable only through the household API.
• Maine is modeled as re-basing on January 1, though the guidelines are published mid-month (1/13/21, 1/12/22, 1/19/23, 1/11/24, 1/17/25, 1/13/26). The parameter is month-granular, so the standard moves up to 19 days early. A Maine household between $2,608.33 and $2,660 in early January is scored against the new standard before it legally applies.
• snap_gross_test_income_fpg_ratio is no longer consumed by any formula. It remains a public variable and is left in place; removing it is a deprecation decision for API consumers, not part of this fix.

Tests

• tanf_non_cash_gross_income_limit.yaml — guideline vintage by state, the 185% tier, and the Alaska guidelines.
• meets_tanf_non_cash_gross_income_test.yaml — boundary cases at the exact published standard. Maine's January re-basing puts it a guideline year ahead of Washington in the same January, which exercises the state-varying vintage.
• Mid-year boundaries are asserted by keying outputs by month within a whole-year period: Washington holding the 2025 guidelines in March and moving in April, Arizona's February-to-March step, and Colorado holding through September before flipping in October. The Colorado case means the suite cannot pass if the month parameter is ignored.

Test plan

• gov/hhs/tanf/non_cash/ — 56 passing
• partners/analytics_coverage/.../snap/nc.yaml — 4 passing, no partner-facing change
• CI passes

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 100.00%. Comparing base (c1120a4) to head (db88221).
⚠️ Report is 4 commits behind head on main.

Additional details and impacted files

@@            Coverage Diff            @@
##              main     #9243   +/-   ##
=========================================
  Coverage   100.00%   100.00%           
=========================================
  Files            6         3    -3     
  Lines           81        52   -29     
=========================================
- Hits            81        52   -29     

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.

[DTrim99]

Program Review — multi-agent audit

PR #9243 — State-specific SNAP BBCE poverty-guideline re-base months

Scope

This PR corrects when each state's TANF-funded non-cash (SNAP broad-based categorical eligibility, BBCE) 200% income standard re-bases onto the current calendar year's federal poverty guidelines, rather than forcing every state onto the federal October (fiscal-year) cycle. Under 7 CFR 273.2(j)(2) the BBCE limit is the state's TANF non-cash income limit, so keying it to each state's own re-base month is the correct model.

• Off-cycle re-base months (new parameter fpg_year_start_month.yaml): ME = January, MA = February, OR = March, WA = April; default = October (month 10) for the other ~42 states (including AZ).
• Percentage changes: VI 130% → 175% effective 2024-10-01 (gross.yaml); AZ 185% → 200% effective 2026-03-01 in both gross.yaml and gross_hheod.yaml; AK BBCE at 200% Alaska guidelines (2025-07-01, in-comment).
• Rounding removed: the dollar limit is now the exact percentage × guideline / 12, dropping per-chart whole-dollar rounding (argued as not reaching the categorically-eligible standard per 7 CFR 273.2(j)(2)).
• Mechanism: two new SPMUnit variables — tanf_non_cash_fpg.py (picks guideline vintage via where(period.start.month >= start_month, monthly_fpg(year), monthly_fpg(year-1))) and tanf_non_cash_gross_income_limit.py (multiplies the state BBCE percentage tiers by the guideline to yield a dollar limit). meets_tanf_non_cash_gross_income_test.py is rewired to compare snap_gross_test_income (dollars) against the new dollar limit.
• States/years: multi-state, 2026 focus (plus VI 2024-10, AZ 2026-03, AK 2025-07).
• CI: 33/33 pass. Branch: 9 ahead / 1 behind main. Changelog: changelog.d/snap-bbce-guideline-update-months.fixed.md present, .fixed type (patch bump) appropriate.

Source verification (primary-source results)

State / item	Claim	Source	Verdict
WA = April (4)	200% standard re-bases April 1	WAC 388-414-0001(2)(a)(ii)	✓ VERIFIED — "beginning the first of April every year"
MA = February (2)	200% standard effective Feb 1	MA DTA Helpful Charts & Figures PDF	✓ VERIFIED — "Effective 2/1/2026" column vs "10/1/2025" federal columns
OR = March (3)	200% BBCE "updated every March"	OAR 461-155-0180(6)	✓ VERIFIED — verbatim "200 percent … updated every March" (2026 FPL); lower tiers "every October" (2025 FPL)
ME = January (1)	BBCE 185/200% re-bases mid-January	10-144 CMR Ch.301 §999-3 Chart 4	✓ VERIFIED — six calendar-year eras all begin mid-January (1/13/21, 1/12/22, 1/19/23, 1/11/24, 1/17/25, 1/13/26); Charts 1–3 use FFY (Oct)
VI 175% @ 2024-10-01	130% → 175% eff Oct 1 2024	VI DHS Commissioner announcement	✓ VERIFIED (value + date) — "raised … from 130% to 175% … effective October 1, 2024"
VI 1.3 @ 2015-10-01	continuous 130% era 2015–2024	(none)	✗ UNCORROBORATED — source supports 130% immediately before Oct 2024, not a 2015-onward span
AZ 200% @ 2026-03-01	185% → 200% (gross + gross_hheod), Oct re-base retained	AZ CNAP/DES manual	value/date/mapping ✓ (arithmetic verified to the penny); href ✗ — bare in-comment quote, no resolvable URL (des.az.gov 403)
Default = October	~42 states on federal FFY cycle	USDA SNAP FY26 COLA memo	conceptually correct source; memo page returned 403 (format fine)

Note: OR (March) and ME (January) were "unreachable" for the reference-validator (FlateDecode-compressed PDFs), but the external source-verifier subsequently downloaded both and extracted them with pdftotext -layout — both are now primary-source VERIFIED. All four off-cycle months are corroborated by primary sources.

Verified correct (checks that passed)

1. FPG-cycle premise is regulatorily sound — a state whose BBCE 200% standard re-bases before October is understated for part of the year on the federal October vintage; 7 CFR 273.2(j)(2) makes the BBCE limit the state's TANF non-cash income limit, not a §273.9 SNAP standard. Regulatory reviewer agrees keying to each state's re-base month is the right model.
2. All four off-cycle months primary-source corroborated — WA (WAC), MA (DTA), OR (OAR 461-155-0180(6)), ME (CMR Ch.301 Chart 4). See table.
3. Period / vintage selection is correct (tanf_non_cash_fpg.py:32-40) — inclusive >= boundary matches every test: CO flips at 2026-10, WA at 2026-04, ME leads a full guideline year. No off-by-one.
4. AZ value/date + gross_hheod fix — Feb 2026 = 185%×15,650/12 = $2,412.71; Mar 2026 = 200%×15,650/12 = $2,608.33, both off 2025 guidelines (AZ keeps October vintage cycle); fixing the elderly/disabled gross_hheod (which lagged at 185%) alongside gross is in-scope and correct.
5. No hard-coded values — month read from fpg_year_start_month[state], guideline from gov.hhs.fpg, percentages from gross/gross_hheod/ny.*. Only bare literals are n-1 and /MONTHS_IN_YEAR (named constant).
6. Entity / period handling correct and consistent — all four variables are SPMUnit; year-defined geo vars read with period.this_year. The rewire even FIXES a latent MONTH-on-YEAR mismatch (deleted code read state_code_str with a MONTH period; new code uses period.this_year). No stray params arg.
7. Colorado "month-matters" anti-vacuous test — CO (default October) pins Jan–Sep 2026 at the OLD (2025) vintage $2,608.33 and Oct at the NEW $2,660; a month-ignoring formula fails. Paired with WA's two-sided April boundary and the ME-vs-WA same-January divergence ($2,660 vs $2,608.33).
8. WA 10-household-size chart / arithmetic reproduction — every asserted dollar figure reproduces exactly (WA 1p $2,608.33 / $2,660, WA 4p $5,358.33; AZ $2,412.71 / $2,608.33; AK $3,258.33; ME $2,660; CO $2,430; NY tiers; CA Jan-2022 $2,146.67). Dollar assertions carry a tight absolute_error_margin: 0.01.
9. Both new formula variables have direct unit tests — no CRITICAL zero-coverage path; tanf_non_cash_fpg is exercised transitively via 185%-vs-200% cases with known FPG.
10. Changelog present and correctly typed; no TODO/FIXME/placeholder in any changed file.

Critical (must fix)

None. All validators (regulatory/effective-date, source corroboration, code patterns, entity/period, test coverage) returned zero criticals. All four off-cycle months are primary-source verified.

Should address

1. VI 2015-10-01 = 1.3 back-history is uncorroborated — parameters/gov/hhs/tanf/non_cash/income_limit/gross.yaml (VI). The PR rewrites VI from flat 2015-10-01: 1.75 to 2015-10-01: 1.3 → 2024-10-01: 1.75. The 175% value and 2024-10-01 date are verified, but the announcement only establishes 130% immediately before Oct 2024, not a continuous 2015–2024 span. Historical, pre-window, low population impact (no microdata dependence shown, no current-year result flip). Fix: set the 130%-era start date to when 130% is actually documented, or add a source covering the pre-2024 window.
2. AZ percentage reference is a bare quote with no resolvable href — gross.yaml and gross_hheod.yaml (AZ 2026-03-01). Value, date mapping ("benefit month of 03/2026" → 2026-03-01), and October-re-base rationale are all correct; the defect is sourcing/format only (des.az.gov sits behind a 403 challenge page). Fix: add a resolvable/archived CNAP manual URL as a reference title/href, and promote both the AZ and VI justifications from code comments into parameter reference metadata blocks.
3. MA (February) and OR (March) re-base months have no month-boundary test — tests/.../tanf_non_cash_gross_income_limit.yaml. Two of the four off-cycle states the changelog advertises are set in the parameter but never asserted; a regression resetting MA/OR to the October default would pass the entire suite. Fix: add two-sided cases mirroring WA (e.g. MA 2026-01 = 2,608.33 vs 2026-02 = 2,660; OR 2026-02 = 2,608.33 vs 2026-03 = 2,660). This is the single most material coverage gap.
4. VI 175% / 2024-10-01 step is untested — tests/.../*.yaml. The new VI time-step has no case on either side of 2024-10-01. Add one VI case: pre-2024-10 at 130%, post at 175%.
5. AZ gross_hheod test is vacuous on the HHEOD branch — tests/.../tanf_non_cash_gross_income_limit.yaml:47-54. The age: 70 case expects the same value as the general standard ($2,608.33), so it would pass on the general gross path too. Add a month/scenario where gross and gross_hheod diverge, or assert the HHEOD branch explicitly.
6. Rounding-removal caveat for round-up-chart states (accepted-by-design; document only) — tanf_non_cash_gross_income_limit.py. The legal argument is sound: categorically-eligible households are exempt from §273.9 standards per 7 CFR 273.2(j)(2), so 273.9(a)(3) rounding does not reach the BBCE limit — the regulatory reviewer accepts returning the exact percentage×guideline. For states whose published charts round UP, the exact value is ≤ the chart value (≤ ~$2), so a household between the exact and rounded figure scores ineligible slightly early vs. the state's own chart. No test shows the ~$2 gap crossing an actual eligibility threshold (Maine's $2,660 lands exactly on the exact value). Recommend documenting the per-state round-direction in the docstring; do NOT re-introduce per-state rounding.
7. Territory $0 BBCE limit flips a zero-income / high-asset household eligible (documented, do not block) — parameters/gov/hhs/fpg.yaml GU/PR/VI placeholders are 0, so tanf_non_cash_fpg = $0 and the limit = $0; with the BBCE asset-test waiver, income ($0) <= limit ($0) flips such a household SNAP-eligible where the base asset-tested path scored it ineligible. Direction/magnitude: API-only edge, no territory microdata, so it does not move population aggregates; it is pre-existing (the prior ratio path had the same income/0 exposure) and the author already documents it. Report on the record; keep SHOULD ADDRESS, do not prescribe.
8. tanf_non_cash_fpg duplicates snap_fpg arithmetic (maintainability) — variables/gov/hhs/tanf/non_cash/tanf_non_cash_fpg.py vs variables/gov/usda/snap/income/snap_fpg.py. Not a true duplicate (it adds the per-state cutover month snap_fpg cannot express), but both now encode the same guideline arithmetic and can drift. Consider a shared helper parameterized by cutover month, with snap_fpg as the start_month=10 case; at minimum add a cross-reference comment.
9. Orphaned snap_gross_test_income_fpg_ratio — variables/gov/usda/snap/income/gross/snap_gross_test_income_fpg_ratio.py. This PR removes its only in-model consumer; it remains a public, self-tested variable (its own YAML test keeps it green) but has no model caller. Acceptable to keep-and-document or remove in a follow-up; add a one-line note so a future reader knows it is intentionally unconsumed.

Suggestions

1. fpg_year_start_month.yaml description — leads with acronyms ("TANF non-cash (SNAP BBCE)") rather than the standard verb form; consider "Sets the month from which each state's SNAP BBCE income standards begin using the current calendar year's federal poverty guidelines …". House-consistency nit.
2. Validate integer-valued breakdown: [state_code] (fpg_year_start_month.yaml:135-151) — integer values under breakdown with unit: month should validate (matches */age_group/months.yaml shape; no rate_unit), but confirm on a validation run since it is not independently testable in a read-only pass.
3. Frame a rounding-guard test explicitly — the absolute_error_margin: 0.01 assertions already catch re-introduced chart rounding as a side effect, but no test is named as a rounding guard; add a comment or one case documenting exact-vs-published-chart divergence so intent survives future "cleanups."
4. where evaluates both FPG branches (tanf_non_cash_fpg.py:38-40) — correct and idiomatic (vectorized); parameters(f"{year-1}-01-01") for early years relies on held guideline values, which FPG provides back to 1992. Noting as intentional.
5. Keep test-comment arithmetic in sync — the % × guideline / 12 derivations embedded in test comments (e.g. the recurring $2,412.71 185% figure) are good documentation; ensure they track any future guideline backfill.
6. AZ comment duplication — gross.yaml and gross_hheod.yaml both restate the AZ rationale; a shorter cross-file pointer in gross_hheod.yaml would reduce drift risk.
7. WA reference title lacks subsection — file-level title cites bare "WAC 388-414-0001" while the inline comment gives "(2)(a)(ii)"; add the subsection to the title so the citation is self-contained.
8. CO/HHEOD 2016-vintage history cases removed in the meets_ conversion — not a correctness gap for this PR, but historical-vintage coverage at the 2016 boundary is now thinner; note in case a reviewer wants it retained.

Validation summary

Check	Result
Regulatory / effective-date	PASS — premise sound, 0 critical; all four off-cycle months verified against primary law
Source corroboration	PASS with notes — WA/MA/OR/ME/VI-175% verified; VI-2015 back-history and AZ href open (sourcing, not value)
Code patterns	PASS — 0 critical, 3 should (formatting/breakdown-validation/HHEOD-test), 5 suggestion
Entity / period	PASS — consistent SPMUnit; period.this_year on year-defined geo vars; fixes a latent month-on-year bug
Test coverage	PASS with gaps — no zero-coverage formula; MA/OR month tests missing (HIGH), VI step untested (MEDIUM)
CI	PASS — 33/33

Review severity: COMMENT

Zero criticals across all five validators and all four off-cycle re-base months are primary-source verified, so this does not warrant REQUEST_CHANGES; the remaining items (VI 2015 back-history sourcing, AZ href, MA/OR + VI test gaps) are author-judgment and low-impact rather than value errors, so COMMENT is the right call over a bare APPROVE. Credit where earned: the author's own PR body pre-documents the ME mid-January approximation, the rounding-removal legal rationale, and the territory $0-limit edge — several "findings" here are limitations the author already surfaced.

Next steps

• /fix-pr 9243 to apply fixes (add MA/OR + VI boundary tests; add AZ href / promote VI+AZ refs to metadata; re-date or source the VI 130% era; document rounding round-direction + orphan variable + territory edge).

🤖 Generated by /review-program (Claude Code)

[DTrim99]

Address minor nits, then ready

[DTrim99]

Re-review — fix commit 6e996f17c8 verified ✔

Re-checked each finding of the review above against the actual files at head db88221cee (not the commit message), re-doing the arithmetic:

Should-address items resolved:

• MA (Feb) + OR (March) boundary tests (our top gap) → RESOLVED. Both states now have genuine two-sided cases keyed by month within a 2026 period: MA holds 2,608.33 in January and moves to 2,660 in February; OR holds 2,608.33 in February and moves to 2,660 in March — each with a third assertion at October proving the state month is independent of the federal October cycle. Hand-verified against fpg.yaml: 200% × 15,650/12 = 2,608.33 (2025 guideline) and 200% × 15,960/12 = 2,660 (2026). Genuinely two-sided.
• gross_hheod vacuous test → RESOLVED, via Alabama (not AZ). The right call: AZ's general and elderly standards are both 200% post-2026-03 so an AZ test is necessarily vacuous, whereas AL genuinely diverges (general 130% / elderly 200%). The new AL cases at 2026-01 assert 2,608.33 for age 70 vs 1,695.42 for age 30 — the differing values mean the elderly case actually exercises the gross_hheod branch (a formula misreading the general standard would return 1,695.42 and fail).
• VI 130% sourcing → clarified honestly. The vi.gov announcement is now in reference metadata with a resolvable href, and the comment candidly states the 130% is documented only immediately before Oct 2024, with the 2015-10-01 key labeled a file-baseline convention rather than a documented adoption date — exactly the clarification requested (the pre-window value has $0 impact anyway, since VI's FPG is a $0 placeholder).
• AZ 200% reference → RESOLVED. A resolvable SNAP Screener href was added, with a comment explaining the primary CNAP page 403s; the value/date were already source-verified in the first review.
• WA reference subsection + meets_ signature → RESOLVED. Title now cites WAC 388-414-0001(2)(a)(ii); the formula signature gained parameters (cosmetic — the arg is unused in the body, no functional effect).

Documented by design (sound): the VI step has no dollar assertion because a $0-FPG territory makes any boundary test pass vacuously; the chart round-direction caveat and the snap_fpg/tanf_non_cash_fpg sync note are now inline. The orphaned snap_gross_test_income_fpg_ratio is untouched — acceptable, it was flagged as intentionally unconsumed.

New-issue scan clean: every new expected value ties out to fpg.yaml, no hardcoded logic or test short-circuits, no partner-test files touched, CI 33/33 green at head.

Recommendation: APPROVE. The two substantive gaps are genuinely closed with hand-checkable arithmetic; the only residuals are the cosmetic unused arg and the honestly-annotated pre-window VI value, neither blocking.

🤖 Generated by /review-program re-verification (Claude Code)