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

[MaxGhenis]

Summary

The frozen Enhanced CPS artifacts on the Hugging Face model repo policyengine/policyengine-us-data carry a miscellaneous-income point-mass defect in every revision from 1.78.2 through the final 1.115.5 (the revision main resolves to). Any consumer running tax analysis against the frozen file's latest revisions gets roughly 2x-inflated income tax aggregates. Since the repo is deprecated and will not be re-cut, this needs a consumer advisory on the artifact itself plus a populace-side guard so the defect class can't recur in successor releases.

Evidence (measured on enhanced_cps_2024.h5)

Quantity	rev 1.73.0 (Aug 2025)	rev ≥1.78.2 (incl. 1.115.5)
miscellaneous_income, unweighted column sum	$92B	$2.07T
Largest single record	—	$795M
Weighted AGI (pe-us 1.772.0, 2026)	$15.7T (SOI-consistent, within 3%)	$23.9T
Weighted federal income tax (2026)	$2.43T (within 5% of SOI)	$4.4T

The defect is the known QRF imputation landmine: quantile-regression-forest draws placing extreme point masses on miscellaneous_income for a handful of records. It entered at 1.78.2 and persists through the final frozen revision.

Discovered while pinning data for the IARIW OBBBA incidence paper (PolicyEngine/obbba-paper — private; cleaning manifest and revision comparison in data/clean/manifest.json and the methods section there). The paper pinned 1.73.0 and validated it against IRS SOI aggregates.

Asks

1. Consumer advisory on the frozen artifact: a warning in the HF model card / README of policyengine/policyengine-us-data stating that revisions ≥1.78.2 (including main/1.115.5) should not be used for tax analysis, and pointing to 1.73.0 as the last SOI-consistent revision of that lineage (or to certified populace successors).
2. Benchmark guard in populace release gates: a max-single-record / tail-mass check on imputed income columns (misc income being the proven offender) so a $795M single-record draw can never pass certification. Related work: Dense arm fails input-mass parity on miscellaneous_income (−58%); Build M dense needs a target-side remedy #393 (dense-arm misc-income input-mass parity), Build I: selection-side misc-income support fix for a certifiable 57k default successor #359 (selection-side misc-income support fix) — this issue is the artifact-level consumer warning + guard, complementary to those pipeline fixes.

Reproduction

from huggingface_hub import hf_hub_download
import h5py
for rev in ["1.73.0", "1.115.5"]:
    p = hf_hub_download("policyengine/policyengine-us-data",
                        "enhanced_cps_2024.h5", repo_type="model", revision=rev)
    with h5py.File(p) as f:
        col = f["miscellaneous_income"]["2024"][:]
        print(rev, f"sum=${col.sum()/1e9:,.0f}B", f"max=${col.max()/1e6:,.0f}M")