Build M live default: non_sch_d_capital_gains lands at $74.6B vs its $10.2B SOI target (+635%); Schedule-D CGD route is empty

[MaxGhenis]

Summary

The live HF-latest release (populace-us-2024-buildm-sparse-rmloss100-98bf731-20260717T075101Z) ships non_sch_d_capital_gains at $74.6B against its own $10.2B SOI target — a +635% relative error recorded in the release's calibration_diagnostics.json — while schedule_d_capital_gain_distributions is empty. The #282 route split (Schedule-D-routed capital gain distributions carved out of long_term_capital_gains, direct route on non_sch_d_capital_gains) is not realized in the shipped artifact, and the build manifest records no CGD-split stage.

Observed (all reproducible from the published release)

From releases/populace-us-2024-buildm-sparse-rmloss100-98bf731-20260717T075101Z/calibration_diagnostics.json:

target (Ledger fact)	bound variable	target	final estimate	rel. error
irs_soi.ty2023.table_1_4.all.capital_gain_distributions_amount@2024 (Pub 1304 Table 1.4, concept irs_soi.capital_gain_distributions_1040)	non_sch_d_capital_gains (sum)	$10,155,465,319	$74,617,447,202	+6.3475
irs_soi.ty2023.table_1_4.all.capital_gain_distributions_returns@2024	non_sch_d_capital_gains (indicator_sum)	3,209,131	3,209,038	−0.0000
irs_soi.ty2023.table_1_4.all.net_capital_gains_amount@2024	capital_gains_gross (sum)	$1,270,864,366,489	$945,431,772,792	−0.2561

The returns-count target is hit exactly while the dollar target is 7.3× over — the inflation is per-record amounts, not weights. The TY2022 historic_table_2 state net_capital_gains_amount targets all land within ~0.2%, so the −25.6% national miss is specific to the TY2023 Table 1.4 national row.

Simulated at pe-us 1.775.8, year 2026, on the artifact:

• live latest: non_sch_d_capital_gains $85.5B across 3.26M tax units; schedule_d_capital_gain_distributions $0.0B; capital_gain_distributions $85.5B; long_term_capital_gains $1,116B; AGI $18,962B.
• pe-us pinned default (populace_us_2024.h5@populace-us-2024-c86a631-6e1bcd0271a5-20260619T002242Z, 75,112 hh, pre-Split the Schedule D route of capital gain distributions out of long-term gains (US) #282 lineage) for comparison: non_sch_d_capital_gains $13.69B across 4.01M units — consistent with the SOI direct-route concept (Table 1.4 TY2023 $10.2B; PUF E01100 uprated).

The overshoot ($74.6B − $10.2B ≈ $64.5B) is close to the Schedule-D-route CGD magnitude documented on #274 (SOCA TY2015: $68.05B routed via Schedule D line 13), and the national net-capital-gains line is short by a similar order — consistent with Schedule-D-route dollars sitting on the direct-route input rather than inside long_term_capital_gains. That attribution is a hypothesis; the two discriminating checks are below.

build_manifest.json for the release contains no stage touching non_sch_d, schedule_d, capital_gain_distribution*, or split_component* (the only "split" hit is an unrelated mortgage-interest note).

Why it matters

• Consumer impact: any analysis keyed to non_sch_d_capital_gains on the live default computes on ~7× the SOI base. Concretely: the GROWTH Act (H.R. 2089) groundwork that motivated pe-us #8828/#8839 measures ~$13.7B (2026) of direct-route distributions on the pe-us pinned default, but ~$85.5B on this artifact.
• Engine treatment: post pe-us #8839/#8850, non_sch_d_capital_gains adds to gross income and net_capital_gain — so ~$64B of income rides the wrong route line. Whether total AGI/CG is net-neutral depends on whether those dollars were also removed from long_term_capital_gains (the −25.6% national miss suggests partial offset, not confirmed).
• Gate question: a +635% relative error on a national SOI dollar target shipped inside a release whose headline is 90.2% within-10%. Should national Table 1.4 dollar rows be within-tolerance-blocking (same family as the Frozen eCPS artifacts ≥1.78.2 (incl. final 1.115.5) carry a misc-income point-mass defect that ~doubles tax aggregates — consumer advisory + release guard #457 consumer-advisory precedent and the Dense arm fails input-mass parity on miscellaneous_income (−58%); Build M dense needs a target-side remedy #393 dense-arm parity remedy)?

Discriminating checks (for the Build M lane)

1. Column masses on the Build M dense parent / base-m pool: does it carry schedule_d_capital_gain_distributions with ~$60–70B and non_sch_d_capital_gains at ~$10–14B (split ran, sparse export dropped the memo column — Sparse-57k export drops 94 no-formula input columns from the dense parent (tips, overtime, disability flags, ESI premiums, rent, QBI basis) #361 class, since it is a no-formula input), or non_sch_d at ~$75B with schedule_d absent (split never ran / donor stage mapped total CGD onto the direct-route variable)?
2. Whether the Split the Schedule D route of capital gain distributions out of long-term gains (US) #282 split_component_by_share stage is invoked anywhere in the Build M release path (the manifest records none).

Repro

from policyengine_us import Microsimulation  # 1.775.8
live = Microsimulation(dataset="hf://datasets/policyengine/populace-us/populace_us_2024.h5")
print(live.calculate("non_sch_d_capital_gains", 2026, map_to="tax_unit").sum() / 1e9)          # 85.5
print(live.calculate("schedule_d_capital_gain_distributions", 2026, map_to="tax_unit").sum())  # 0.0

Relates #274, #282, #361, #393, #457; pe-us PolicyEngine/policyengine-us#8839, PolicyEngine/policyengine-us#8850; Build M PR #455. Found re-verifying GROWTH Act groundwork numbers against current artifacts.

[MaxGhenis]

Root cause traced — it is not the #282 split stage; it is a QRF tail point-mass on non_sch_d_capital_gains in the puf_tax_detail stage (the #457 signature, now in the Populace artifact).

Evidence

Pipeline: non_sch_d_capital_gains is one of 64 outputs of the puf_tax_detail stage's fit_weighted_qrf op (predictors: employment/self-employment income, interest, both dividend legs, capital_gains, filing status — us/source_stages.json). The capital_gain_distributions stage only writes schedule_d_capital_gain_distributions (from long_term_capital_gains_before_response × the SOCA share) and never touches non_sch_d.

Distribution on the live artifact (person table, 2024 values, pe-us 1.775.8):

• 2,295 nonzero records → 3.21M weighted (the SOI returns-count target, hit exactly).
• Median nonzero value $2,979 — the body is SOI-consistent (3.21M × ~$3.2k ≈ the $10.2B Table 1.4 target).
• p99.9 = p100 = $594,484 — a repeated ceiling value (at least 4 records carry exactly $594,484, with weights 0–14,199).
• Top 100 records carry 89.0% of the $74.6B. The whole +$64B excess is a few hundred tail draws; the SOI dollar target had no blocking tolerance to catch it.

So the count target pinned the number of recipients while a handful of donor-tail broadcast values (the $594,484 point mass looks like a donor max/topcode) inflated dollars 7.3×. Same failure family as the frozen-eCPS misc-income defect in #457 and the historical eCPS income-tax blowup (unweighted/tail-oversampled QRF donors broadcast as point-mass landmines).

schedule_d_capital_gain_distributions = $0 is a second, independent gap: it is absent from release_input_coverage_manifest.json (9 keys; neither route variable listed), so nothing guards whether the split stage's output survives to the export — #361 class.

Suggested fixes

1. Gate (small, immediate): make national SOI Pub 1304 Table 1.4 dollar rows within-tolerance-blocking (a +635% rel. error on a national dollar target should fail certification), and add a tail-concentration check (top-k mass share) for QRF-imputed sparse dollar variables — the 89%-in-100-records signature is detectable at build time.
2. Export contract: add non_sch_d_capital_gains and schedule_d_capital_gain_distributions to the input coverage manifest so route presence is guarded (Sparse-57k export drops 94 no-formula input columns from the dense parent (tips, overtime, disability flags, ESI premiums, rent, QBI basis) #361 remedy scope).
3. Imputation (Build N Build N epic: CD target surface + base rebuild bundle (quartet, assets, QRF fix, feed regen, gate hardening) #449 "QRF fix" scope): sparse tax-detail variables need explicit zero-inflation handling and a donor-support guard (draws bounded to observed donor support with weight-aware tail treatment) — non_sch_d is nonzero on ~2% of returns; conditional QRF without a participation margin both over-spreads and tail-broadcasts.

Happy to take (1) and (2) as PRs from a fresh worktree unless the Build M/N lane prefers to fold them in — flagging here first to avoid colliding with in-flight base-m work.

[MaxGhenis]

Claiming fixes (1) and (2) from the comment above in a fresh worktree off origin/main (branch claude/musing-bardeen-994886): (1) national SOI Pub 1304 Table 1.4 dollar rows become within-tolerance-blocking in the release gate path, plus a tail-concentration gate (top-k weighted-mass share) for QRF-imputed sparse dollar variables that catches the 89%-in-100-records signature at build time; (2) non_sch_d_capital_gains and schedule_d_capital_gain_distributions added to the release input coverage manifest (#361 remedy class). Fix (3) — the QRF imputation change itself — remains Build N #449 scope. PR to follow, cross-linking #361 / #457 / #449.

[MaxGhenis]

Fixes (1) and (2) are up as PR #464 (gate + tail-concentration + export contract; fix (3) untouched, per the claim above). Verified against this release's own diagnostics: the 0.25 Table 1.4 dollar blanket fails exactly the two defect rows (+634.8% CGD, −25.6% net capital gains) and passes the other nine dollar rows; the top-100-share gate reads 89% on the shipped column vs the 75% threshold. One additional observation from the same pull of the artifact's diagnostics, for the Build N #449 gate-adjudication pass: two Table 1.4 returns-count rows are also wildly missed (estate_trust_net_loss_returns +495.9%, other_income_net_loss_returns +328.1%) — same family, outside #464's declared dollar-row scope.

[MaxGhenis]

Fix-3 discriminating checks answered against the certified Build M base (base_populace_us_2024_puf_support.h5, design weights):

1. The split never ran. schedule_d_capital_gain_distributions does not exist anywhere in the base file (verified by walking every HDF node). non_sch_d_capital_gains carries the full distribution total on the person table: $30.27B weighted across 7,452 rows (3.07M weighted carriers). So this is not the Sparse-57k export drops 94 no-formula input columns from the dense parent (tips, overtime, disability flags, ESI premiums, rent, QBI basis) #361 sparse-export-drop class — the pool never had the route leg.
2. Declared but never invoked. The capital_gain_distributions stage is fully declared (source_stages.json, kind split_component_by_share, output schedule_d_capital_gain_distributions, SOCA TY2015 + Pub 1304 provenance) and its executor is registered (source_runtime.py → split_us_component_by_share_from_manifest), but neither tools/build_us_asec_pooled_source_base.py nor tools/build_us_puf_support_base.py references the stage — which is exactly why shipped build manifests record no split stage.

Fix 3 is therefore a wiring fix: invoke the declared stage in the base path (following the neighboring stages' invocation/gate/manifest-recording pattern), then re-run the base through the staged pipeline with the frozen Build M selection. Implementation is in flight on branch cgd-split-462; conservation (non_sch_d + schedule_d == pre-split total per tax unit), manifest recording, and refuse-to-rerun behavior get pinned by tests. The #464 route-leg coverage requirement stays red until the re-run ships, by design.

One note for the re-run: at design weights the direct-route residual after the split should land in the SOI Table 1.4 concept class ($10–14B against the $10.2B target), which also relieves the −25.6% national net_capital_gains_amount miss from the wrong-route side.

[MaxGhenis]

Correction to my previous comment, after the implementation lane's contract reading (branch cgd-split-462, BLOCKED.md) — two claims were wrong:

1. "Fix 3 is therefore a wiring fix" — no. The declared capital_gain_distributions stage is a memo carve-out: schedule_d_capital_gain_distributions = long_term_capital_gains_before_response × 0.09852561497474391 (SOCA TY2015 share) where LTCG > 0 and no direct-route amount exists. It never touches non_sch_d_capital_gains — its packaged share is the Schedule-D CGD residual over LTCG-excluding-direct, not a repartition share for the $30.27B direct-route total. Wiring it satisfies the Blocking Table 1.4 dollar-row fit + QRF tail-concentration gates; export-guard both CGD route legs (#462) #464 route-leg coverage requirement and gives reforms the component; it moves zero dollars.
2. "Per-tax-unit conservation" — not the stage's contract, and its own unit test pins the memo semantics (test_us_capital_gain_distributions.py). I had inferred a split-and-conserve design; the executor's actual transform is additive-memo-only.

The substantive fix 3 remains what the root-cause comment above defined: zero-inflation + donor-support guards on the puf_tax_detail QRF for sparse dollar outputs — the +$64B lives in a few hundred tail draws (p99.9 = p100 = $594,484 point mass, top-100 records = 89% of mass), while the $30.27B design-weight body is a separate identification question the calibration + #464 tail gates own.

So the work is two independent pieces, both feeding the Build N base re-run: (3a) wire the declared memo stage into build_us_puf_support_base.py (insertion point identified: after qrf_finalization, before qbi_reconciliation; checkpointed stage recording is automatic) per the executor's existing contract; (3b) the QRF participation-margin + donor-support bounding. 3a is in flight; 3b design next.

[MaxGhenis]

3a is merged (#480): the declared memo stage now runs in the PUF base pipeline (outer stage between qrf_finalization and qbi_reconciliation, identity-asserted, stage-recorded), so schedule_d_capital_gain_distributions materializes on the next base build and the #464 route-leg coverage requirement can go green. 3b (QRF zero-inflation/donor-support guard for the non_sch_d amount tail) is the remaining piece; mechanism diagnostic in flight.