[PavelMakarchuk]

The September Linux TAXSIM build aborts the 2024 and 2025 dashboard refresh in mdtax22_ (#1214). Add an explicit, opt-in fallback to the previous hash-pinned Linux binary for Georgia and Maryland in those two years. All other state/year combinations keep the new binary; all TAXSIM outputs for a fallback household come from a single binary.

Each regenerated TAXSIM row carries its binary SHA256; reused years retain their original per-year provenance. Checkpoint identity, summary provenance and draft release notes disclose the fallback. Existing memory/disk limits and full-population checks remain active; core dumps are disabled on hosted refresh workers. A dependency-free diagnostic checks the previous binary on all Maryland inputs and isolates the failing original Linux batch before the full refresh runs.

Validation: 12 lightweight refresh tests pass locally, including state/year scope, hash rejection, memory-limit termination and summary/source checks. Hosted reproduction and full regeneration are in progress. No heavy model runs were performed on the laptop.

This is a dependent draft against codex/refresh-new-taxsim-september, which combines the Alabama input correction (#1204) and binary update (#1206). Do not merge the underlying binary upgrade until #1214 is resolved or this temporary exception is explicitly accepted. This workaround does not resolve the upstream TAXSIM arithmetic bug.

The 2024 smoke test also exposed an uninitialized old variable in TAXSIM’s Georgia low-income credit routine. Georgia now uses the pinned previous binary too. Only 2024–2025 are regenerated; 2021–2023 artifacts are verified and reused unchanged from run 35893683416. Every remaining TAXSIM batch is preflighted before PolicyEngine calculations. Current full recovery run: https://github.com/PolicyEngine/policyengine-taxsim/actions/runs/35953332152 .

Completed refresh: run 35953332152 succeeded for all five years (2021–2023 reused). State absolute match rates: 72.1%, 76.5%, 83.8%, 85.1%, 85.9%. Federal rates unchanged at displayed precision. The 2025 state decline versus 87.0% baseline is dominated by Louisiana (91.5% to 26.9%); this is a discrepancy to investigate, not proof either model is legally wrong. Generated dashboard data is committed; the release remains draft and no live deployment has been performed. Hosted fallback verification and dashboard build: run 36000366107.

[vercel]

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

Project	Deployment	Actions	Updated
policyengine-taxsim	Ready	Preview	Sep 27, 2026 8:12pm UTC

[PavelMakarchuk]

Closing. The Maryland crash in #1214 is closed, and #1444 replaces these fallbacks: it splits the refresh into a TAXSIM half and a model half, and it uses a batch that TAXSIM stops on only when two reruns give identical outputs.