Chesterton's-fence audit: wire NIPA wages/proprietors, net worth, LIHEAP; fence every exclusion

[MaxGhenis]

Chesterton's-fence audit of the target-parity contract

Follow-up to #404. Max flagged that macro_control_total was over-broad: us-data deliberately targeted NIPA totals to capture wages that never reach tax returns (filed-return / CPS wages cover ~$4T; NIPA wages are the ~$12.4T economy-wide universe including nonfilers). The rule is absolute: us-data-targeted ⇒ compiled, unless source-absent. A category label is no longer a sufficient reason to drop a target — every reviewed exclusion now carries a fence {origin, purpose, verdict_basis} recovered by git/PR archaeology on the archived policyengine-us-data repo, and the regenerator refuses a fenceless exclusion. Where archaeology is inconclusive, the fence is rebuilt (wired), not removed.

The authoritative us-data target surface is the eCPS loss matrix (utils/loss.py) + db/etl_national_targets.py, read from the GitHub archive.

Wired — reviewed_exclusion → compiled (following us-data's exact concept mapping)

family	engine counterpart	origin	purpose (verbatim)
bea_nipa.total_wages_salaries (red line)	employment_income_before_lsr national, $12,387,929,000,000 exact	PR #994	"add BEA/NIPA all-population macro targets for wages … nonfiler income"; local fit showed microdata wages $4.06T vs the $12.388T target
bea_nipa.proprietors_income	self_employment_income_before_lsr + sstb_… + farm_operations_income + partnership_s_corp_income national	PR #994	"Proprietors' income … all persons, including nonfilers"
federal_reserve_z1.households_nonprofits_balance_sheet	net_worth national	PR #282	no prose rationale in-thread → wire-by-default; linear direct-sum pins the post-reweight wealth aggregate the input-stage SCF imputation does not
hhs_acf_liheap.national_profile	indicator sum of spm_unit_energy_subsidy (recipient households)	PR #688	"we could not calibrate LIHEAP recipient counts … despite having CPS/SPM baseline LIHEAP receipt" — the SNAP-caseload pattern

bea_nipa.total_wages_salaries joins ssa_supplement.ssi_recipients in the gate's red-line set: silent loss of economy-wide wage targeting is the exact failure this audit exists to catch. Net-worth and proprietors carry a documented period-aging caveat (the feed's cy2023 fact is aged to 2024; a cy2024 fact would replace the estimate).

The classified fence table — us-data targets populace does not compile

family	classification	origin	verdict + basis
bea_nipa.personal_interest_income / personal_dividend_income	macro_control_total	PR #994 (added) → #1059 (removed)	us-data itself declined: "BEA personal interest/dividends include imputed interest, pension-plan dividends, and trust flows, so those macro totals should not directly calibrate tax/CPS interest and dividend variables" (loss.py:67-70). Calibrated from SOI/CBO (compiled).
bea_regional.state_wages_salaries	deferred	PR #1034	deferred: the feed carries raw place-of-work SAINC4 line-50 wages (DC = $96.8B, far above residence wages); us-data residence-adjusts (line 42 apportioned) + scales to the national total in bea_regional.py, not reproducible from the feed. National wage aggregate is compiled.
cms_nhe.medicaid_title_xix_expenditures	non_linear	PR #292	architecture-retires: PE-US PR #1138 made medicaid cost a person_weight-dependent SLCSP allocation → not a linear row; populace keeps it calibration_role=validation_only.
irs_soi.traditional_ira_contributions, roth_ira_contributions, form_w2_401k_elective_deferrals, form_w2_designated_roth_401k_contributions	input_side	PR #496 / #554 → #1125	architecture-retires: us-data #1125 and populace both govern retirement contributions as imputed pre-limit input columns (*_contributions_desired, POST_REFERENCE_ECPS_REQUIRED_INPUTS), with PE-US applying statutory limits — calibrating here would double-govern. The feed carries the amounts, so this is not a source gap.
irs_soi.form_w2_social_security_tips	not_modeled	PR #220 (fixes #215)	deferred: "'No Tax on Tips' requires a clean tip_income field … tipped workers skew lower-income" — but PE-US yields a structural zero for tip_income in the base microdata until the tip source stage is ported; the target is unsatisfiable, not silently dropped.
bls.consumer_expenditure, wic.national_summary, hud.housing_assistance	source_absent	loss.py BLS_CE_TOTALS / etl WIC / etl_housing_assistance	source-absent: no BLS/WIC/HUD fact in the pinned feed.

Every other BEA/KFF/census-projection/SNAP-persons/TANF-caseload/SOI-CD family carries a "not a us-data calibration target" fence — verified absent from loss.py and etl_national_targets.py by archaeology (the only BEA direct-sum targets were wages + proprietors, PR #994; the programs some NIPA series measure are calibrated from SSA/CMS/CBO admin sources, compiled).

Schema + mechanics

• TargetFence(origin, purpose, verdict_basis) — required on every reviewed exclusion; TargetFamily rejects a fenceless exclusion; the loader parses it and the anti-rot enforces it.
• The wirings reuse the existing dispatch/materializer machinery: _bea_reference_from_fact (source-specific dispatch, like ssa/census_pep), a DIRECT_LEDGER_TARGETS entry for net worth, an INDICATOR_LEDGER_TARGETS entry for LIHEAP (the snap_households pattern), and CBO-income-source aging roles for the cy2023 NIPA levels.
• Registry 5,689 → 5,693 specs; manifest 25 → 29 compiled, 55 → 51 reviewed exclusions; deterministic; Ruff clean.

Testing

Wiring tests per family (NIPA wages exact-value + proprietors expression + other-BEA-not-compiled + state-wages-deferred + net-worth + LIHEAP indicator), fence-schema tests (fence required, three-field validation, every shipped exclusion fenced, deferred state-wages fence names PR #1034), the red-line generalization, and the guarded regeneration test proving the committed artifacts reproduce from the pinned feed. Full contract + fiscal-targets + builder suites green.

Please do not merge — conductor reviews. Gates Build M alongside #404.

🤖 Generated with Claude Code