Repository navigation
Conversation
Replace the 20%-of-state-amount approximation in local_sales_tax with the State and Local General Sales Tax Deduction Worksheet: the Optional Local Sales Tax Tables (with the selector's named counties and cities) in the line-2 states and the ratio method elsewhere, with line 3 the combined rate above the state table heading rate. Combined rates come only from official state files: the Streamlined Sales Tax rate and boundary files, New York Publication 718 and 718-A, and Virginia Tax's rates by locality, weighted to Census 2020 places. Where no official file exists the combined rate defaults to the heading rate, so the local amount is 0 unless the user enters a combined or local rate. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…ne 3 Address the independent review of #9879: - Recode SST place codes that match no Census place or county subdivision of their county by the records' postal city, or by the code the same ZIP codes carry after the records end. This fixes the DeKalb County part of Atlanta (Georgia code 05000; 8.0% before, 8.9% now) and Vermont's 2022 town taxes (older town codes, and town-wide taxes under CDP codes). In Vermont codes resolve to towns. The build logs every recoding and every code left unmatched. - Write each covered state's general sales tax rate (state_rates.csv) and take worksheet line 3 as the combined rate less it, so rounded or averaged headings (Minnesota 6.88%, South Dakota 2023 4.35%) no longer shift line 3. California and Nevada keep the IRS rule: the combined rate above the heading. - List New York City's five counties as Table A counties. - Accept 7-digit place GEOIDs; document the county_fips/place_fips keys, the pre-2022 backdating and the remaining limits (Iowa, Vermont towns outside places, a few Utah codes). Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
This was referenced Oct 6, 2026
Collaborator
PR Review: Model the IRS local sales tax worksheet with official locality ratesPR #9879 (author: MaxGhenis, head
SummaryThe PR replaces PolicyEngine's placeholder local general sales tax (
The worksheet logic and the IRS tables are right.
What blocks approval:
There are seven optional suggestions. Critical (must fix)C1. Rutland city, Vermont, is at 7% for 2022 (and 2015–2021) instead of 6%.
Should addressS1. In the 16 local-tax states without official files, the local amount is now $0. The PR doesn't quantify that change, and the impact section is still a placeholder.
S2. Several
S3. Axiom: rulespec-us#1402 gives no module path for the worksheet module this PR mirrors.
S4. The input-normalization branches have no tests.
SuggestionsAll are optional. 1. Mississippi localities outside Jackson and Tupelo default to Jackson's table, but the IRS logic points to the ratio method.
2. Some location inputs are ignored by one or both lookups.
3. Two data limits are documented deferrals. Both are reasoned and documented in
4. 2015–2021 use the 2022 tables, selector, heading rates and locality rates (
5. Record-weighted unit averages pick up small special districts ( 6. The state general rates live in a CSV rather than a parameter.
7. The build can't be reproduced once the posted files change.
Verified correct
Axiom line assessmentThe PR body has
CI status
Branch status
Verdict: REQUEST_CHANGES |
DTrim99
requested changes
Oct 13, 2026
DTrim99
left a comment
DTrim99
left a comment
Collaborator
There was a problem hiding this comment.
Early feedback on the draft (details in the comment above). The worksheet logic matches § 164(b)(5)(H)(i), and all 1,824 local-table cells, the selector and the heading rates match the IRS PDFs. One critical: rates.csv:6285-6287 puts Rutland city, VT at 7% in 2022, and by backdating in 2015-2021, but its 1% local option tax began in July 2023, so 2022 local_sales_tax should be 0, not 95.33. Four should items: the 16 local-tax states without official files now get $0 local (Houston 2025: 0 here vs 169.80 on main and 271.68 per the IRS), so quantify the regression; the rate.yaml and local_sales_tax.py page anchors; the worksheet module path missing from rulespec-us#1402; and tests for the 7-digit place GEOID and zero-padding branches.
MaxGhenis
marked this pull request as ready for review
October 16, 2026 07:15
Contributor
Author
|
Thanks, David. Every item from your 10-13 review is answered at
Could you take another look? |
Collaborator
PR Review (round 2): Model the IRS local sales tax worksheet with official locality ratesPR #9879 (author: MaxGhenis), head SummaryAll five prior blockers are fixed.
The Mississippi suggestion was adopted in its IRS reading: localities without a table now take the ratio method. The geography suggestion was implemented in full. I checked the data again by sampling, against official sources, with no mismatches:
I recomputed the NY state average and the Rutland remainder from the CSVs. The Python tests are justified and cheap in CI. The data adds about 240 KB to the wheel. CI is 27/27 green. Three new should-address items remain. None changes a household result.
There are six optional suggestions. Prior findings
New findingsShould addressN1. Axiom: rulespec-us#1402 doesn't match the head's absent-table rule.
N2. On the tree merged with
N3. The default-table descriptions don't follow the house template, and "PolicyEngine assigns" is no longer accurate.
SuggestionsAll are optional.
Verified
Axiom line assessmentThe body has
The line becomes valid once the absent-table rule and a Mississippi companion case are added. CI status
The run predates the source-conflict ratchet (#10287), so N2 shows up only on the merged tree. Branch status
Verdict: REQUEST_CHANGES |
DTrim99
requested changes
Oct 23, 2026
DTrim99
left a comment
DTrim99
left a comment
Collaborator
There was a problem hiding this comment.
Round 2: all round-1 findings are fixed (Rutland 2022, the state-by-state impact, references, the #1402 module path, normalization tests), and the data checks out: 18 localities across NY/VT/WA/TN/OH/AR match official sources, 15 crosswalk places and 7 state population totals verify, and every new or changed expectation hand-checks. Three should-address items: (1) rulespec-us#1402 doesn't match the head (add the 'no table = zero' rule and a Mississippi 2025 case); (2) merge blocker: on the merged tree, the #10287 ratchet fails because local_sales_tax.py's KNOWN_UNMARKED count drops to 0, so delete that entry after merging main; (3) the default_table and none.yaml descriptions are off the house template. Six suggestions. Full report in the comment above.
- Drop local_sales_tax.py from KNOWN_UNMARKED: on the tree merged with main it has no unmarked conflict-sounding comment, so the #10287 ratchet's shrunk check failed. - Default-table descriptions follow the IRS-led house form (a-d, none.yaml). - Shipped text cites the 2025 Schedule A instructions (p. 5) instead of decision ids. - The changelog fragment states the $0 local amount in the 16 uncovered local-sales-tax states (was 20% of the state amount). Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…round 2, suggestion 3) build.py now records output_sha256 for rates.csv, populations.csv, place_counties.csv and state_rates.csv, and manifest.json carries the current files' hashes (place_counties.csv unchanged). test_manifest_hashes_every_shipped_csv checks that the manifest names exactly the shipped CSVs and that each hash matches. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Contributor
Author
|
Round 2 is addressed at
Suggestions 1–5 are done. For suggestion 3, the manifest now hashes all four CSVs, with a test that ties the hashes to the shipped files. Suggestion 6 goes to a separate PR. The by-state impact rerun at this head is queued for when the host's current reservation lifts, and its table will replace the 🤖 Generated with Claude Code |
This branch has not been deployed
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
Replace
0.2 × state_sales_taxwith the IRS State and Local General Sales Tax Deduction Worksheet, lines 2–6: local tables/selector and the ratio method. Official SST, New York and Virginia files cover 21 states; elsewhere the local amount defaults to $0 unless a rate is entered, under Max's ruling d984 (2026-10-06). No Tax Foundation-derived data are included, enforced bytest_l0_official_sources_only.The hub's population impact, by state, is under "Hub microsimulation impact". It was rerun at the current head
76ebba54faagainst main at6a92d6a983. The national totals are within 0.4% of the earlier run at724a8fbe11; Mississippi's rows moved by less than 0.1%. Mississippi methodology is resolved under Max's ruling d1590, reading B: absent local tables use the ratio method; Jackson and Tupelo retain their tables.The law
What changes
Variables (
variables/gov/local/tax/sales/):combined_sales_tax_rate(household, input-able): place → county remainder with known place/block → county population average → state average; heading rate without official data.local_sales_tax_rate(household, input-able): line 3 in percentage points,max(combined − official_state_component, 0) / 0.01; heading fallback and CA/NV thresholds.local_sales_tax_table(household): place → county → state default.local_sales_tax(tax unit): worksheet line 6.Both lookups share
tools/local_sales_tax_geography.py: normalize codes, reject out-of-state GEOIDs, resolve county FIPS → block county → supplied county enum → place principal county; supplied NYC/Yonkers flags infer missing places. Calculated defaults are ignored across branches. The generated crosswalk covers all Census places in the 21 rate states and single-county places elsewhere: 30,924 places, including 13,550 population-weighted places in the 21 states.Parameters (
gov.irs.deductions.itemized.salt_and_real_estate): Tables A–D, 2022–2025 (all 1,824 cells independently checked), selector state/default/county/place lists with named Census codes, and heading rates. Tables uprate after 2025. The 2022–2025 references identify all supporting PDF pages once per annual document, includingstate_sales_tax_table/tax.yaml.Data/build:
data/local_sales_tax/rates.csv: 7,720 step-function rows/4,448 county-place keys/21 states;populations.csv: matching Census weights;state_rates.csv: dated state components;place_counties.csv: generated crosswalk;SOURCES.md: sources/method/limits.scripts/local_sales_tax_rates/build.pychecksny_pub718.csvagainst the publication andvt_local_option.csvagainst Vermont dates.tools/local_sales_tax_rates.pylazily caches annual rates. The manifest records source/output hashes, logs and parser audits: 19 retained boundary files, 20,414,368 accepted records, 0 malformed rows skipped; 489 omitted empty final types padded and retained.Sources and licences
None of the rate sources carries an open licence. The PR ships only the rates and Census codes derived from them, which are facts set by state and local law, with attribution, never the source files.
Method and known limits
--downloadmay alter historical outputs. Exact reproduction needs retained hash-matching inputs. Validate 89-field records, allow empty trailing fields, normalize omitted unused final types, log malformed-width counts/examples, fail broken quoting. V3 caches pin source/data hashes and audits; compare/explain historical CSV changes.NONEfor an absent table: worksheet line 2 = 0, and line 6 uses the ratio method with Mississippi's 7% state rate. The 2025 Schedule A instructions, p. 18 list "City of Jackson only" for Table C and "City of Tupelo only" for Table A; those city tables are retained.Invariants
The worksheet/locality Python tests cover:
Tests
Reviewed baseline
944686e0d4: 15/8/25/26 YAML cases (combined/local rate/selector/local amount), 40 worksheet/23 locality/89 state-table Python tests and 4,448 data keys, correcting the original body's counts.New cases: Rutland dates/backdating/remainder; NYC seven-digit GEOID, Atlanta short place and DeWitt/Arapahoe short counties; invalid geography and block/enum/flag/place-only precedence; Vermont aliases/boundaries, Core provenance/branches, generated crosswalks and CSV/ZIP parser/cache failures.
Hand calculations:
(181×6% + 184×7%)/365= 6.50410959%; $766 × (184/365)/6 = $64.36. 2024 7%.(273×6.2167% + 92×6.2192%)/365= 6.21733014% after town records move out of the city.3651000/alternate inputs: 8.875%/A, $120 × 4.875 = $585; Yonkers flag/place-only 8.875%/B, $142 × 4.875 = $692.25.5001→05001, 6.5% + 1% + 3.5% = 11%; Arapahoe8005→08005, Table B in 2022 versus Colorado's default A.4000, 2024: 8.9%/B, the other-locality table. Westchester without city 8.375%/B, $142 × 4.375 = $621.25; no location retains NY 8.5411545063%/D, $135 × 4.5411545063 = $613.06.Required file checks for d1fabb8: each changed file separately, foreground;
OMP_NUM_THREADS,OPENBLAS_NUM_THREADS,MKL_NUM_THREADS,VECLIB_MAXIMUM_THREADS,NUMEXPR_NUM_THREADS= 4,OMP_WAIT_POLICY=PASSIVE.combined_sales_tax_rate.yamllocal_sales_tax_rate.yamllocal_sales_tax_table.yamllocal_state_tax.yamltest_local_sales_tax_worksheet.pytest_locality_sales_tax_rates.pytest_state_sales_tax_table.pytest_local_sales_tax_rate_build.pyRate, selector and provenance files passed after the resolver/crosswalk changes; parser tests passed after the source-audit change. The remaining worksheet/state/amount files passed at the final head. After the runtime checks, edits affected documentation and the builder.
Formatting/lint:
uvx ruff@0.15.5 formatandcheck, andmake format, passed.Local cost for the expanded locality file: 409.1 s wall including setup; peak RSS was not collected. The previous attempt's 23-test baseline took 483.9 s; host conditions differ, and no baseline RSS was recorded. The added Core provenance/branch checks reuse the existing model; parser tests construct no model. These Python files remain in Rest's existing
remaininggroup; no CI runner was added. These are local macOS figures, not Linux CI measurements.CI cost, from this PR's CI run artifacts as read in the round-2 review: in Rest shard-1's
remaininggroup the sales tax Python files take 49.4 s of test time (worksheet 40 tests in 23.1 s, state table 89 in 20.4 s, locality 33 in 5.5 s, build 9 in 0.4 s), and the group took 14:55 wall at a peak RSS of 7.93 GB, against 14:46–19:19 and 7.84–7.89 GB for three sibling PRs that day. In the microsimulation group the county-carryover test takes 0.8 s for 4 tests; the group's peak RSS is 14.16 GB against 14.07–14.09 GB on siblings (about +75 MB). No CI job was added.Original main
4c900b6d3dfailed 11/26 local-amount cases (NYC $107.80, Tacoma/Texas $200); new rate variables were absent. Original checks also passed unchanged state-sales-tax, CTC liability, Hawaii SALT and NY itemized-deduction files.Hub microsimulation impact (current head)
A real microsimulation on the default dataset, main at the merge base
6a92d6a983against this PR's head76ebba54fa, for 2025 and 2026. The runs went one at a time under the shared heavy-job lock. Sums are weighted.local_sales_taxstate_sales_taxsalt_deductionstate_income_taxincome_taxhousehold_net_incomehousehold_benefitsEvery total in the table matches the state breakdown's "ALL" row, which is computed separately from household sums. Local sales tax changes in 36 states and moves both ways. In 2025 New York gains $5.6bn, Georgia $2.1bn and Washington $1.1bn, while California loses $3.8bn, Texas $2.7bn and Florida $2.1bn. The SALT deduction falls most in Texas (−$2.5bn) and Florida (−$2.0bn). Federal income tax rises most there too (+$59M and +$44M) and falls most in Washington (−$44M). State income tax moves in 4 states, by less than $10,000 in total.
By state: weighted change, summed to households, states with any change
local_sales_tax2025local_sales_tax2026salt_deduction2025salt_deduction2026income_tax2025income_tax2026Comparison files are in
fixes/9879-impact/on the hub host.Coverage and household examples
The 16 local-tax states without an official locality file here are AL, AK, AZ, CA, CO, FL, HI, ID, IL, LA, MO, MS, NM, PA, SC and TX. Without an entered rate their local amount becomes $0, replacing main's 20% of the state amount. This is a coverage regression under d984, particularly relevant in no-income-tax Texas/Florida; it does not mean their localities impose no tax.
Houston, 2025, single/$85,000: state amount $849; main local $169.80, PR default $0, entered 8.25% worksheet $849 × 2 / 6.25 = $271.68. Houston FY2025 adopted budget, PDF p. 8 confirms that combined rate. Actual SALT/tax effects depend on itemization, the income-tax election and the SALT limit.
Controlled household probes use one unweighted single person per state, age 35, $85,000 wages, $5,000 property tax and $20,000 deductible interest. The comparison substitutes main's old local formula on the same merged model; these examples do not estimate population impacts.
Official follow-up sources include Florida DOR's DR-15DSS county surtaxes and CDTFA's rates/downloads.
Follow-ups
Response to David Trimmer's review
C1 — fixed: recode Rutland town/city, validate Vermont dates, regenerate rates; test city/remainder/transition/backdating and builder boundaries.
S1 — fixed: documented Houston's regression and the TX/FL/WA household probes; the hub's state SALT and income tax microsimulation is in "Hub microsimulation impact", with a by-state table.
S2 — fixed: 2022–2025 heading/state-table parameters and worksheet variables cite each annual document once with all supporting PDF pages; pre-2022 metadata is unchanged.
S3 — paths/imports specified: rulespec-us#1402 identifies worksheet targets and same-year table imports; d236, signed-import and corpus/release/approval blockers remain.
S4 — fixed: both lookup YAML files cover NYC
3651000→ 8.875%/A and Atlanta4000in DeKalb → 8.9%/B in 2024, plus short/invalid geography coverage.1 — resolved under Max's ruling d1590 (reading B): Mississippi localities other than Jackson and Tupelo select
NONEfor an absent table, so worksheet line 2 = 0 and line 6 uses the ratio route with the 7% state rate: $1,243 × 1 / 7 = $177.57 (2025/$85,000/family 1/1% local), replacing the $179 Jackson proxy. Jackson retains Table C and Tupelo retains Table A, as specified by the 2025 Schedule A instructions, p. 18. Without an entered rate the local amount remains $0.d1590 validation at
b35e82324a1d9c81865d18546937279f70ccc1a4: 31 selector, 32 local amount and 40 worksheet tests passed in separate foreground processes with four-thread caps. The worksheet grids now includeNONEwithout additional model constructions or CI jobs and remain in the existingremaininggroup. Local worksheet wall time including setup: 188.85 s; peak RSS: 1.64 GiB. The prior body reports 638.7 s for 40 worksheet tests; host conditions differ and prior peak RSS was not recorded.2 — fixed: shared supplied-input geography, generated principal counties, precedence/invalid/branch tests and coverage limits.
3 — retained deferrals: Iowa averages (Des Moines 6.83% versus 7%) and Vermont outside-place/uncoded limits; Rutland fixed separately.
4 — retained staging: backdate 2022 IRS/first locality rates to 2015–2021; Add the IRS 2015-2021 local sales tax tables, selector and heading rates #9911 covers historical IRS data/instructions, locality history separately.
5 — documented: weighted small districts retained, Wichita 7.506% versus ordinary 7.5%; no new cutoff.
6 — documented: unrounded dated CSV components, MN/SD differences, parameter/API limitation; no migration.
7 — fixed: audit all 19 files/20,414,368 records; retain 489 SD records by padding their omitted unused final type, with full collapsed-input equality to the legacy parser; log malformed skips, reject broken quoting, preserve audits in hash-pinned caches, and test failures/normalization; document current downloads/historical-output review.
Hub review P2 — fixed: when a dataset supplies a county for an earlier year and the county formula carries it into a later year, the resolver now treats that county as supplied, so it uses the county's rate and table instead of the state's average. Regression test:
test_local_sales_tax_county_carryover.py.P2 test fix: handle
NONEin the state-table test lookup and masks so absent local tables use line 1 × line 3 / line 4; Jackson C and Tupelo A stay covered. All 11 requested files pass (171 Python + 272 YAML = 443 tests).Response to round 2 (David Trimmer, 2026-10-23)
mainis merged (5a1896f844), and7003e15c8bdeletes thelocal_sales_tax.pyentry fromKNOWN_UNMARKED. The test's own comment scanner, run on the merged tree's copies of the PR's 21 files undervariables/andparameters/, finds no unmarked conflict-sounding comment inlocal_sales_tax.pyand none gained in any other file.default_table/{a,b,c,d}.yamlread "The IRS has residents of these states use Optional Local Sales Tax Table A unless the selector names their city or county" (B, C, D likewise), andnone.yamlreads "The IRS provides no Optional Local Sales Tax Table for residents of these states unless the selector names their city."76ebba54fa):build.pyrecordsoutput_sha256for all four shipped CSVs,manifest.jsoncarries them (theplace_counties.csvhash is unchanged), andtest_manifest_hashes_every_shipped_csvchecks that the manifest names exactly the shipped CSVs and that each hash matches. It fails against the previous manifest.SOURCES.md,none.yamland the two YAML test comments cite the 2025 instructions' p. 5 sentence instead of decision ids; the ids stay in this body._stored_inputintotools/general.pychanges a shared helper thatmainuses elsewhere.axiom: TheAxiomFoundation/rulespec-us#1402 queued
TheAxiomFoundation/rulespec-us#1402 tracks the tables/worksheet with IRS companion cases:
us/form/irs/ty<YYYY>/i1040sca-line-5a-worksheet.yaml(TY2018–2025),.../i1040sca-line-5-worksheet.yamlbefore 2018. Same-year imports await d236/signed bindings and corpus/release/re-pin/gated promotion. Locality rates are line-3 data inputs. No billed encoder run is authorized.🤖 Generated with Claude Code