Regenerate TANF data for 2026 (policyengine-us 1.715.3 + Indiana fix) — complete 56/56

[hua7450]

Summary

Regenerates all 56 precomputed TANF data files from policyengine-us 1.715.3 (+ the Indiana TANF fix, PR #8543) at tax year 2026, replacing the stale 1.598.0 / 2025 data. Also adds a vectorized data generator that is ~600× faster than the existing one and validated bit-for-bit identical to it.

Of the 51 jurisdictions, 23 produced changed values; the other 28 recomputed to byte-identical data (no diff from the prior committed values).

New: vectorized generator (scripts/precompute_vec.py)

The old generator built 15,376 Simulation objects per state (one per grid cell) — that construction, not the math, was the bottleneck (~25–45 min/state). The new generator builds 16 per state:

• Household-structure dims (adults × children) stay a 16-iteration loop — they change entity counts, so they can't be axes.
• Income dims (31 earned × 31 unearned = 961 cells) become two PolicyEngine axes groups → one Simulation computes all 961 cells in a vectorized pass.

Measured: 4.3s vs ~47.6 min on Illinois (~664×). Full 33-state completion run took 74s.

Validated bit-for-bit identical to the cell-by-cell generator on 10 states covering every code path — plain states, person-level monthly special vars (IL), county selection (CA), and SPM-unit annual vars (CA/CO): 0 mismatches across 138,384+ cells. A per-structure fallback to the cell-by-cell path guarantees the run can never emit wrong/missing data. Documented in scripts/README.md.

Notable changed values (vs 1.598.0, $0-income household)

State	before	after (2026)	why
IN	2p 255 / 3p 320	2p 409 / 3p 513	🐛 the fix — was paying the standard of need, now the amount of assistance (IC 12-14-2-5). Exact benchmark match.
KY	3p 524	3p 341	reflects KY's Nov-2025 benefit cut
CT	3p 833	3p 897	2026 COLA (CT indexes annually)
NH	—	3p 1,366	2026 COLA
HI	3p 763	3p 611	old data was stale/high; now matches benchmark (~610)
KS / IL / AK	403 / 759 / 924	429 / 782 / 925	2026 updates

Benchmark verification (NC / CBPP "maximum benefit")

Compared all 51 jurisdictions, both household sizes (1 adult + 1 child, 1 adult + 2 children), against the external benchmark:

• 39/51 match exactly (≤ $4 of source rounding, both household sizes).
• 12 diverge — all explained, none are computation errors:
  • Zero-rent test household (AZ, NY, VT): no rent input → no shelter allowance; adding rent reproduces the benchmark.
  • Benchmark is an older year (CT, NH 2026 COLA; KY Nov-2025 cut) — PE is current.
  • Bundled allowance (MA): PE includes the clothing allowance (~$41.67/mo per person).
  • Region selection (PA, VA): benchmark uses a specific locality; PA Group 1 matches exactly.
  • Idaho flat $309 all sizes is correct; the benchmark's escalating $389 is the error.
  • Rounding (IL, AK, HI, NE, OH, SC): annual÷12 vs single-month, ≤ $5.

No genuine model concerns remain (the previously-flagged Indiana gap is fixed by #8543, which this data already carries).

Generator / metadata changes (scripts/precompute.py, metadata.json)

• YEAR 2025 → 2026
• OUTPUT_DIR fixed: ../frontend/public/data → ../public/data (repo moved Vite → Next; old path no longer exists)
• FPG dict → 2026 values (CONTIGUOUS 15,960/5,680 · AK 19,950/7,100 · HI 18,360/6,530)
• stdout line-buffered + a --workers flag
• metadata.json: year 2026, FPG 2026, version 1.715.3, CA/PA/VA county-gated

Provenance

Generated locally from policyengine-us 1.715.3 + the Indiana fix (PR #8543). Once #8543 is released, a clean re-run on the released version reproduces this output.

🤖 Generated with Claude Code