Certified populace-us drifts behind current eCPS on 10 parity layers; pin the parity reference + restore the gate runner

Summary

The certified US default (populace-us-2024-5da5a95-20260611, the dataset shipped in policyengine.py bundle 4.16.2) passes its build-time parity gate but drifts behind the current enhanced CPS on 10 populated layers. This was reproduced, not inferred — by pulling the build's own check_parity.py out of git history (it was removed from HEAD, see below) and running it against the sha-verified certified artifact with two references:

Same candidate both runs (populace_us_2024.h5, sha256 f32c2e5e…, == the certified_data_artifact in policyengine.py us.json). So this is not a gate escape and not a build bug — the build's parity_gaps=0 reproduces exactly against the eCPS it was gated on. The reference moved; the parity gate ran against an unpinned, since-changed eCPS, and nothing re-checks for drift.

exported_nonzero passes (308 stored columns, all non-zero) — populace's own columns are clean. The gaps are purely missing relative to the newer eCPS.

The 12 gaps vs current eCPS

Pure inputs (no formula) — add to the impute/donor set (PUF-derived):

• amt_foreign_tax_credit (eCPS populates 9.3% of records)
• general_business_credit (5.4%)
• excess_withheld_payroll_tax (4.6%)
• prior_year_minimum_tax_credit (4.0%)
• early_withdrawal_penalty (3.3%)
• other_credits (2.8%)

These feed net federal tax, so the certified default computes different (lower-magnitude) tax than current eCPS for the affected 3–9% of records.

Formula-owned but yields 0 — populate the upstream inputs the formula needs:

• self_employed_health_insurance_ald (7.4%) — registered formula exists, but its upstream self-employed health-premium inputs are absent in populace, so it computes 0.

Reported / comparison aggregates (not formula-consumed) — decide carry-or-document:

• spm_unit_net_income_reported (99.4%), spm_unit_total_income_reported (97.7%), ssi_reported (1.5%)

Vacuous (eCPS populates 0% too — no action):

• recapture_of_investment_credit, unreported_payroll_tax

Root-cause / process gaps (the more important fixes)

1. The parity reference is unpinned. The gate ran against a working-copy eCPS, not a recorded revision, so "parity 0" is not reproducible and silently rots when eCPS changes.
2. The certified release manifest records no gate result — no gaps count, no reference identity, no skipped-layer count. The verdict isn't in the artifact.
3. The gate runner was deleted from HEAD. packages/populace-data/build/us/check_parity.py (and the rest of the build/us snapshot) were removed in fda3838 "Remove retired build and benchmark surfaces". The gate library (populace.build.parity_gate) remains, but nothing invokes it against a reference — so the release contract is only reconstructable from git history.

Proposed fixes

• Re-home a reference-pinned simulation-level parity runner into populace-build (reuse parity_gate), recording the reference eCPS repo+revision+sha256 and the full result (gaps, skipped, layer counts) in the release manifest. (PR incoming.)
• CI drift-check: re-run parity against the latest published eCPS on every build so reference drift fails loudly instead of going invisible.
• Close the 6 pure-input credit/penalty gaps via the PUF donor/impute stage; populate the upstream for self_employed_health_insurance_ald; decide explicitly whether to carry the 3 reported aggregates or document them as a named exemption.
• Restore the build/gate surface to the repo (or move it into the package) so the release contract is auditable, not archived in a deleted commit.

Repro: check_parity.py from commit 33ed83d, run with POP_CANDIDATE/POP_REFERENCE env overrides.

Start with the historical packages/populace-data/build/us/check_parity.py at commit 33ed83d and the remaining populace.build.parity_gate library, then inspect policyengine.py's us.json certified artifact entry. Reproduce the parity check with the recorded candidate and references. Done should include an auditable pinned reference, recorded parity results, and a restored or relocated gate runner with drift checking addressed.