[PavelMakarchuk]

Update the emulator's bundled NBER TAXSIM test executables for macOS and Linux to build cd2026102718 (October 27, 2026). This is the build that the TAXSIM-35 census audit now uses. Windows stays at the August build (see below).

Platform	On main	Updated	Source
macOS	cd2026081819	cd2026102718	https://taxsim.nber.org/out2psl/osx
Linux	cd2026081819	cd2026102718	https://taxsim.nber.org/out2psl/linux
Windows	cd2026081318	cd2026081318 (kept)	https://taxsim.nber.org/taxsimtest/taxsimtest.exe

Why this build

• The macOS and Linux builds give the same output on all 15,000 records of the CPS comparison sample. Each build also gives the same output on repeated runs: 8 runs on macOS for the IA and CO batches, and 3 runs on a hosted Linux runner.
• Earlier October builds did not: the October 8 Linux build stopped with SIGFPE on some batches, and the October 24 macOS build gave different IA 2021–2022 results from run to run (Does the current TAXSIM executable return two different 2021–2022 Iowa results for the same two-earner joint record? #1471).

Changes

• Windows kept at August. NBER's later Windows builds have not matched the macOS and Linux builds. For example, the October 8 Windows build gave an Indiana 2024 v39 of $63.20 where they gave $0.00. On October 28, taxsimtest/taxsimtest.exe was still the October 22 upload. So taxsimtest-windows.exe stays at cd2026081318, with a note in the manifest and the README.
• Binaries. manifest.json records the URLs, build stamps, sizes, last-modified dates and SHA-256 checksums. The README table is updated.
• Parser. TAXSIM prints debug lines between records: d1 <value> before Minnesota 2024–2025 records (October 8 build), and d1 <integer> <integer> before Colorado 2021 records (October 27 build). TaxsimRunner skips these exact shapes, and only these shapes. Without the second shape, every Colorado 2021 batch failed. Tests cover both shapes.
• AddMed. From cd2026100817, TAXSIM keeps the Additional Medicare Tax out of fiitax in every year, the correction reported on Does taxsimtest count the Additional Medicare Tax twice (in fiitax and in fica/tfica)? #1225. The new build does the same (2015, 2022 and 2024 checked). The 2013–2023 test now xfails only for builds before cd2026100817, which keeps the Windows build's behaviour.
• Known differences. tests/taxsim_known_regressions.py reads the platform's bundled build stamp. A test can then tolerate the exact values that the current build changes, and it fails if any other value changes. When TAXSIM fixes a listed difference, the test fails again, so that the entry is removed. Three differences apply to this build. In each, the emulator keeps the pinned values and only the binary moved:
  • Does the current TAXSIM executable count nonprop twice in AGI before 2024? #1482: nonprop is counted twice in federal AGI in every year before 2024 (1990–2023 checked), and in the state AGI that starts from it. This affects the pre-2024 records of the nonprop guard.
  • Does the current TAXSIM executable apply the 33% of New York's Empire State child credit a second time? #1461: New York's 2024 Empire State child credit is multiplied by 33% a second time (sctc 108.90 instead of 330), so dependent-count record 9 has siitax −791.50 instead of −1,012.60.
  • Does the current TAXSIM executable leave the child care credit out of Credit Limit Worksheet A for the 2024–2025 additional child tax credit? #1483: in 2024–2025, the child care credit is left out of Schedule 8812 Credit Limit Worksheet A when the additional child tax credit is computed, so student-status row 2 has actc 528.00 instead of 928.00 and fiitax −672.98 instead of −1,072.98.
  • The nonprop guard also keeps the CA 2021 siitax change first seen in cd2026100817 ($0.72 at CA taxable income $45,197). The cause is not established.
• Does the current executable start the federal EITC phase-out at zero income? #1416 fixed. The EITC phase-out values of cd2026100817 (Does the current executable start the federal EITC phase-out at zero income? #1416) match the pinned values again in this build. So the MD and MFS guard tests return to their main versions, and the Does the current executable start the federal EITC phase-out at zero income? #1416 lists are removed.

Validation

• Formats. The executables are Mach-O x86-64, ELF x86-64 and PE32+ x86-64. The macOS and Linux files embed cd2026102718.
• Probes. The VA 2021 opt(30) probe gives siitax 2068.05 and srebate 500. A single filer with $400,000 of wages gives fiitax = v28 in 2015, 2022 and 2024.
• Tests on macOS (Apple silicon with Rosetta 2), for every test file that runs the bundled binary: 361 passed, 4 xfailed, 8 skipped and 0 failed, in 14 files. The 4 xfails are Does the current TAXSIM executable apply the 33% of New York's Empire State child credit a second time? #1461, Does the current TAXSIM executable count nonprop twice in AGI before 2024? #1482 and Does the current TAXSIM executable leave the child care credit out of Credit Limit Worksheet A for the 2024–2025 additional child tax credit? #1483 above, plus the existing otheritem divergence. The 8 skips are the hosted end-to-end suite (6) and the Docker wrapper tests (2).
• Dashboard. The TAXSIM-only dashboard run on Split the dashboard refresh into a TAXSIM half and a model half #1444, with this branch merged, finished all five years on a hosted Linux runner (run 40686561393). Its first run, before the Colorado parser fix, failed on 1,476 Colorado 2021 records (run 40684445967).

Other open questions about this build are in the TAXSIM-35 audit issues (for example #1446, #1448, #1461, #1472, #1473, #1482 and #1483). Benchmarks and dashboard runs stay pinned to their binary hash. Comparisons made before and after this update can differ even with the same PolicyEngine code.

🤖 Generated with Claude Code

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated
policyengine-taxsim	Ready	Preview	Oct 28, 2026 12:28pm UTC