[MaxGhenis]

Part of #1446 (step 2).

us_housing_inputs_signal_gate now fails when anyone outside a renter household carries positive pre_subsidy_rent. Until now the gate counted these people (positive_rent_nonrenter) and passed anyway. impute_us_pre_subsidy_rent now gives heads of owner and no-cash-rent households zero rent, so the CPS rent draw passes the stricter gate.

Why the gate fails this rent

#1444 fixed rent placement: impute_us_pre_subsidy_rent drew rent in household order and wrote it to heads in person order, so heads received other households' rent. with_us_housing_inputs returns a frame unchanged when the gate passes, and the gate passed misplaced rent, so a base built before the fix kept it.

The Route A support bases from before and after #1444 hold the same 907,863 people and differ only in person.pre_subsidy_rent. Each build writes the gate's counts to base_populace_us_2024_puf_support.summary.json:

Support base	Gate	positive_rent_nonrenter	positive_rent_nonhead	Weighted rent
c0513db, before #1444	passed	66,623	0	$648.26bn
9434c50, after #1444	passed	0	0	$672.02bn

With this change the first base fails the gate and the second passes.

Why the draw zeroes rent outside renter households

The ACS rent model draws for every CPS head, with tenure among its predictors. The shared QRF's zero-or-positive classifier gives each owner and no-cash-rent head a small chance of positive rent. On the Route A pre-clone frame (168,852 heads, 53,055 of them in renter households), the draw on main put rent on these heads outside renter households:

Seed	Heads	Weighted rent	Largest draw
0	0	$0	none
1	1	$4.18m	$5,520
2	1	$3.07m	$3,960
3	4	$49.99m	$20,400
4	2	$19.78m	$12,000
5	2	$55.48m	$24,000
6	3	$63.70m	$22,800
7	4	$103.36m	$63,600

Under the stricter gate, seeds 1 to 7 would each stop a support-base rebuild. The draw still runs for every head and then sets rent to zero outside renter households, so renters keep their draws. For seeds 0 to 7 on that frame, everyone in a renter household keeps main's draw, everyone else gets zero, and with_us_housing_inputs passes the gate. Route A's support base at 9434c50 used seed 0, and at seed 0 main's draw put rent on no one outside a renter household.

Where the gate runs

• tools/build_us_puf_support_base.py checks it before support cloning. When it fails, the builder draws rent again from the ACS donor that --acs-h5 names, and without that flag the build stops. If the gate fails again after PUF-support imputation, the build stops.
• tools/build_us_fiscal_refresh_release.py checks the base frame and raises RuntimeError with no redraw, so a release from a base that predates Place each ACS rent draw on its own household's head #1444 stops at the housing gate. The Route A release dry run reads the 9434c50 base through --base-h5, and that base passes.
• The multispine pool calls with_us_housing_inputs when it prepares source inputs, so it draws such rent again.

Limits

• If rent moves from one renter household's head to another's, the gate passes it, because both households rent.
• ACS rows get pre_subsidy_rent from transfer_acs_inputs, which fits its own QRF on the donor and draws with the same sign classifier, and nothing afterwards sets owners' rent to zero. Published ACS rows put rent on 30% of ACS owners (State child SPM rates: next steps from the November trace #1446). A fiscal refresh that takes its base from --pool-manifest checks ACS rows, so any such rent stops it at this gate until State child SPM rates: next steps from the November trace #1446 step 3 zeroes owners' rent after the transfer. That arm already fails the gate's tenure check on main when ACS households keep OWNED_OUTRIGHT, which acs_inputs.py assigns to ACS tenure code 2.

Invariants

packages/microcosm-build/tests/engine_free/us/test_us_housing_inputs.py checks these properties:

• For any household sizes and person order, the gate's count equals the number of draws that person-order placement moves off renter heads. The gate reports the renter failure line if and only if that count exceeds zero, and with heads listed in household order, placement moves nothing (Hypothesis, 40 examples).
• With a stub model that draws positive rent for every household, each renter head holds its own household's draw under any person order, and everyone else holds zero (Hypothesis, 40 examples). Under the same stub, with_us_housing_inputs passes the gate (Hypothesis, 40 examples).
• For any rent vector over {0, 1, 12,000}, the gate's count matches a brute-force count, and the gate reports the renter failure line if and only if it exceeds zero (Hypothesis, 60 examples).
• With rent on one person in an owner household, the gate fails with positive rent outside renter households: 1. With rent written in person order on a reversed person table, it fails with 4, and with_us_housing_inputs draws that rent again onto households 1 to 4.
• With the stub model, owner and no-cash-rent heads get zero rent, renter heads keep draws of 1,000 to 4,000, and the gate passes.

Removing the gate's new failure line fails four of these tests, and removing the zeroing line fails three others.

Locally test_us_housing_inputs.py, test_spec_engine_loader.py and test_spec_engine_coverage_tool.py pass (40 tests), and ruff check ., both repinner --check modes and tools/spec_engine_coverage.py --check exit 0. test_us_multispine_pool_tool.py needs policyengine_us, so the engine-us CI job runs it.

Re-pin

The change moves attested kernel source, so tools/repin_us_spec_engine_contracts.py rewrote seed_map, seed_protocol, two spec_sha256 pins and docs/evidence/spec-engine/us-f0-coverage.json. The bundle generator needs the US engine and did not run locally; the engine-us CI job runs the bundle tests. One repinner pass leaves covered_item_count at 40 in test_spec_engine_coverage_tool.py, and a second pass restores 41 (#1530).

axiom: n/a: build-pipeline gate, no policy rule changes

🤖 Generated with Claude Code

[MaxGhenis]

Merged at head bad73e4 with a merge commit (3f7cfc1). GPT-6.1 Sol reviewed that head through Subfleet (job 20261105-225819-pr1529-rereview) and approved it; an earlier Opus review at fdc785a requested the zeroing this head adds. Every check passed on the latest run, and no one requested changes. Sol's sandbox could not run pytest, so the local test and mutation counts in the body come from this session's runs. #1446 step 3 stays open for rent on ACS owners, and #1530 tracks the repinner's second pass.

🤖 Generated with Claude Code