Immigration composition gate anchors on a Pew 2022 unauthorized stock that is not the runtime SSN-card residual (c-27 lane F2)

[MaxGhenis]

What is wrong

The US stacked-pool immigration composition gate compares the runtime SSN-card NONE residual (the pool's proxy for unauthorized status) against Pew's July 2022 unauthorized-immigrant estimate of 11.0 million, and fails candidate-26 on 18,919,043 / 11,000,000 = 1.72x. The sealed c-27 root-cause lane (run 20260903-073827-c27-root-cause, sol) read the archived Pew page and the runtime construct and found them not comparable [integrity-failed]:

• the runtime residual is a proxy after documented/EAD classifications; Pew's is a demographic residual with a lawful-population subtraction and an undercount adjustment;
• the pool includes 45,510 sampled ACS group-quarters people; the archived Pew page states no equivalent group-quarters universe;
• Pew counts over 3 million temporarily protected people (DACA, TPS, pending asylum, T/U — overlapping components); the runtime excludes DACA from NONE and has no matching categories;
• the pool construct is 2024; the anchor is July 2022, with no sourced rebase.

So the 1.72x ratio is not a valid acceptance criterion in either direction: it can neither fail nor pass a candidate. The lane proposes no replacement anchor, adjustment, or number — only same construct or comparable after stated adjustments would permit one.

What to do

1. Re-anchor the gate on a construct the runtime actually measures, or state the adjustments (universe, categories, reference year) that make the Pew stock comparable and receipt them in the gate's evidence.
2. Until then, the gate's verdict must be recorded as not comparable, not fail; a red verdict on a non-comparable anchor must not gate a release, and a green one must not certify.
3. Regression: the gate refuses to emit a numeric verdict when the anchor's construct receipt does not match the numerator's construct.

Evidence

Sealed deliverables: _buildo-runtime/out/candidate-27/root-cause/deliverables/ (IMMIGRATION-READING.md, immigration-reconciliation.json, FINAL-REPORT.md §R1, sha256 2f71c43cb68e2a942250b5158195eca7c17e0e1780fd264fec2490811dfa1c60; archived Pew page sha256 1215853ce1a97b944a76176aa19833b496ed7eda601cba6f8f8e059841f53ac1). Related: #857, candidate-26 exclusion.

[MaxGhenis]

Mechanics note, so this is not read as "the gate is wrong" (read from us_runtime/immigration.py and tools/build_us_fiscal_refresh_release.py at main 09abf2a):

• Pew enters the calculation twice. The 8.3M undocumented-worker control (and the 408k student control) is forced: the label stage spills excess residual workers/students to EAD until the remaining counts match. The 11.0M population figure is never forced; the total undocumented population is emergent.
• The composition gate compares the emergent NONE total to the 11.0M anchor inside the band (0.5, 1.6), described in code as a release-blocking backstop against collapse or explosion, not a calibration objective. It is binding: the release tool raises Release gates failed and stops (build_us_fiscal_refresh_release.py:8980–8996). That is what halted c-26 at 18,919,043 / 11,000,000 = 1.72x.
• The module docstring records the emergent total on the ASEC-only 2024 build at ≈13M ("inside the range of published 2023–24 estimates"). c-26's stacked pool reached 18.9M, so the ACS half roughly doubled the residual. Against a 2024-era anchor of 13–14M, c-26 would sit near 1.4x and pass the band.

So the two things to decide are separable: (1) whether the ACS-side growth from ≈13M to 18.9M is real (needs the replay that #863 blocks); (2) what construct and year the anchor should carry, per this issue. Rebasing the anchor alone must not be used to pass c-26.