QBI base −31%/−40% vs targets and JCT; add per-component repeal-revenue benchmarks (amount residuals can't see yield gaps)

[MaxGhenis]

Summary

Two related findings from validating deduction/credit repeal scoring on the dense release populace-us-2024-f0af251-703bd81a565c-20260620 (bundle 4.18.7), against external scorekeeper estimates:

1. The QBI base runs well below external benchmarks. The release's own diagnostics show the QBI-linked target residual at −31% (relative_error on the JCT qualified_business_income_deduction revenue-loss target), and repealing §199A in isolation yields $45.9B (TY2026) vs JCT's $76.4B FY2026 individuals line (JCX-45-25) — a −40% gap, the largest among the deduction toggles after tips/overtime (which have no input base at all).

2. On-surface amount residuals cannot see repeal-revenue gaps — propose adding per-component repeal-revenue benchmarks to the validation surface. The itemized-deduction amount calibrates to +1.8% on this release, yet repealing each Schedule A line in isolation scores far below its own JCT line:

Component repealed alone	PE TY2026 ($B)	JCX-45-25 FY2026 ($B)	Diff
SALT (cap → $0)	23.9	59.5	−60%
Mortgage interest	23.0	53.0	−57%
Charitable	60.1	81.5	−26%
Medical	9.1	13.8	−34%
Casualty/theft	0.0	0.2	no base in data

The claimed bases are plausible in level (SALT $288.5B under the $40k cap; mortgage $215.7B among 17.7M itemizers), so the shortfall is yield per dollar of base — the marginal-rate composition of who holds the deductions, plus the un-aged TY2022/23 target vintage (aging tracked in PolicyEngine/chronicle#71). A benchmark family of "repeal component X → revenue vs the JCT line" (computed per release, like the reform-validation checks in calibration-diagnostics) would catch this class directly, where amount residuals structurally cannot.

Suggested actions

1. Investigate the −31% QBI target residual on the dense build (it was −81% on June builds — improved but far from closed).
2. Add per-component repeal-revenue benchmarks (SALT / mortgage / charitable / medical / QBI vs their JCX-45-25 lines; full-itemized vs TPC's interaction-adjusted estimate, since JCT lines are non-additive) to the release validation surface.
3. Casualty-loss inputs are absent entirely (immaterial at ~$0.2B/yr, but a completeness note alongside Sparse-57k certified release zeroes untargeted input bases (IRA/HSA/SE-pension/childcare); CDCC and ALD reforms silently score ~$0 on bundle 4.18.8 #278's input-coverage findings).

Related: #278 (sparse-release input coverage), #257 (degenerate-input gates), #272 (validation-config input checks).

Repro: policyengine[us]==4.18.7, managed_microsimulation, zero each component's cap/ceiling (or medical.floor→1.0, casualty.active→false) for 2026+, sum the income_tax delta.

[PavelMakarchuk]

Confirmed the same mechanism on the sparse 57k national-only release (populace-us-2024-sparse-l0-refit-57k-71a0887-national-only-20260701), via the dashboard's in-sample JCT rows — which are the calibration's own final_estimates passed through in_sample_estimates (no second estimator involved).

JCT cy2024 target	target	calib. initial	calib. final	rel err	SOI amount target err	PE yield vs JCT-implied yield
medical_expense_deduction	11.40B	4.48B	5.26B	−53.9%	−0.3%	6.6¢/$ vs 14.2¢/$
charitable_deduction	60.40B	43.55B	42.96B	−28.9%	+2.2%	19.8¢ vs 28.5¢
qualified_business_income_deduction	63.10B	45.44B	45.12B	−28.5%	−0.06%	21.7¢ vs 30.3¢
deductible_mortgage_interest	24.80B	27.14B	26.06B	+5.1%	−15.7%	14.8¢ vs 11.9¢
salt_deduction	21.70B	21.26B	21.71B	+0.03%	+6.6–8.0%	16.6¢ vs 17.7¢

relative_error ≈ (PE yield / implied yield) × (amount ratio) − 1 reproduces all five residuals.

Two additions to the issue's picture:

1. The initial estimates already carry the full gap (medical −60.7%, charitable −27.9%, QBI −28.0% pre-refit; charitable/QBI got marginally worse through calibration as the optimizer favored the SOI rows). Data-composition property, not a sparse-selection artifact — the sparse and dense magnitudes match.
2. The target surface is internally over-constrained at the data's current yield: JCT revenue-loss rows and SOI amount/returns rows over the same records are jointly infeasible. SALT shows the trade in the mirror: its yield gap was only ~7%, so the optimizer nailed the JCT row by overshooting the SOI SALT amount targets +6.6–8.0%. Mortgage is the inverse (yield too high → interest amount left 15.7% under SOI, JCT still +5.1%). Hitting medical's JCT row by reweighting would need ~2.17× weight on medical itemizers → medical SOI amount +117%; the optimizer correctly sacrifices the single JCT row.

Medical's extra severity has direct code evidence: fiscal_targets.py excludes the SOI 500k_to_1m/1m_plus medical bins because the base microdata has zero medical-deduction support there — the high-marginal-rate itemizers that supply JCT's yield don't exist in the data. Fixing yield means fixing the composition (imputation of who holds the deductions), not the weights — consistent with this issue's per-component repeal-revenue benchmark proposal.

(From the external-benchmark audit session, 2026-07-06; full per-target diagnostics on the calibration dashboard.)

[MaxGhenis]

Rerun of the same downstream validation on the certified Build O release (populace-us-2024-buildo-sparse-rmloss100-22bd902-20260722T232627Z, bundle 4.22.3), TY2026 static:

Per-component repeal revenue all moved toward the external lines (aged dollar targets in this lineage; each component repealed in isolation vs its JCX-45-25 FY2026 line):

Component repeal	Build O ($B)	dense f0af251 ($B)	JCT ($B)	Build O diff
SALT cap → 0	31.7	23.9	59.5	−47%
Mortgage cap → 0	26.6	23.0	53.0	−50%
Charitable ceiling → 0%	65.1	60.1	81.5	−20%
Medical floor → 100%	9.3	9.1	13.8	−33%
Full itemized (joint)	112.0	99.0	~130 (TPC)	−14%

The proposal in this issue stands: the release hits its itemized amount residual almost exactly (−1.8% on itemized_deductions_amount@2024) while yielding −47/−50% on the two largest component repeals — an on-surface amount residual cannot see the yield gap. QBI itself improved to −29.8%/−27.5% (2026/2027) from −31%/−40%.

New, related finding — top-tail burden concentration. AGI-ranked burden shares of the full itemized repeal on Build O: top 20% = 90.0% (TPC-via-CRS ~87% ✅), but top 1% = 24.1% vs TPC's ~43% (the dense release scored 39.4%). The sparse arm carries materially less itemized-repeal burden in the extreme tail even though aggregate bases calibrate. A claimed-amount or repeal-revenue benchmark by top AGI band (top 1% / 0.1%) on the validation surface would catch both this and the SALT/mortgage yield gap.

Reproduction lives downstream in scripts/scoring_validation/ (run_components_local.py, capture_agi.py, distributional_check.py, addendum §4).