Say what #720 changed in the unmarried-partner gap's reason

[MaxGhenis]

What this changes

Wording only, in the reviewed is_unmarried_partner_of_household_head US parity gap (ecps_parity_known_gaps.json) and the release manifest reason regenerated from it (tools/build_us_release_input_coverage_manifest.py).

Two statements described Build J, which predates microcosm #720, as if they were current:

• reason said the older-year relationship fallback "never identifies a partner", implying the pool still uses it. Since Pooled ASEC 2022/2023 vintages lack the NOW_* coverage recodes: reported Medicaid at interview thins to 24.6M vs ~58M survey under 65 on the certified artifact #720, tools/build_us_puf_support_base.py and tools/build_us_asec_pooled_source_base.py bind every pooled year to its pinned Census person file. asec_pool._prepare_year_input then restores Census A_EXPRRP before _with_relationship_recode, so no vintage those tools pool reaches the fallback. Build J's base_j.summary.json (2026-07-10) records derived:line_spouse_parent for 2022 and 2023, so the old sentence was true of that build. The new text keeps that history and says what current code does.
• semantic_non_substitutes.A_EXPRRP said "Only the 2024 locked input carries it". That is now scoped to Build J's three HDF inputs. The note adds that the pinned income-year 2025 input carries it, and that since Pooled ASEC 2022/2023 vintages lack the NOW_* coverage recodes: reported Medicaid at interview thins to 24.6M vs ~58M survey under 65 on the certified artifact #720 every pooled vintage carries the Census recode.

The reason now also states a fact it previously left unsaid. The #720 restore (ASEC_CENSUS_PERSON_COLUMNS) excludes PERRP, PECOHAB and A_FAMREL, although the pinned 2022-2024 Census person files carry all three. The A_EXPRRP note now gives code 13's make-up in those files: PERRP 44 and 47 (unmarried partners without relatives) plus PERRP 55 (housemates or roommates without relatives). PERRP 43 and 46 partners fall in code 12.

Not changed: classification (source_unavailability), issue (#38), gating, and every other evidence field. Whether this exclusion should stay as classified is a methodology call queued for Max (decision d1099, memo in the evidence folder). This PR does not prejudge it.

Relation to #1146

#1146 (spec-symbol-citations-us-uk, open) rewrites this entry's derived_A_EXPRRP and hermetic_build_contract, which on main still cite drifted asec_pool.py line ranges and say the pool reads the H5 tables "without refreshing omitted Census columns". This PR leaves those two fields to #1146. The two branches touch different lines and git merge-tree merges them cleanly at #1146 head 0a841e448. Once both land, the entry is consistent. The new text uses only commit-pinned line citations and dotted microcosm.* names that resolve, so #1146's citation guard accepts it (checked with its line_citations and DOTTED_MICROCOSM_SYMBOL on this branch).

Review

The independent review (Subfleet 20261008-175311-review-1157, Opus) approved head 0d8d0a56a. It recomputed every figure from SHA-verified files and found no factual errors. It raised seven nits.

Commit 7f50ba142 applies five of them:

• The reason names Build J, because the repo also uses "the hermetic build" for the current build.
• The closing sentence now has a clear referent.
• "Since Pooled ASEC 2022/2023 vintages lack the NOW_* coverage recodes: reported Medicaid at interview thins to 24.6M vs ~58M survey under 65 on the certified artifact #720" now reads "Since Pooled ASEC 2022/2023 vintages lack the NOW_* coverage recodes: reported Medicaid at interview thins to 24.6M vs ~58M survey under 65 on the certified artifact #720 (PR Read the ASEC at-interview coverage recodes the 2022/2023 inputs lack from the pinned Census archives (#720) #991)".
• The fallback's inputs now include household size.
• The PERRP 55 counts now carry their years.

It also adds the sixth nit's point. A raw-stage checkpoint built before #720 keeps the fallback codes. The US spec's asec_raw_stage pin is one such checkpoint: it has been pinned since 2026-08-17 (f85c248), and #720 landed on 2026-09-23.

Deferred: the PECOHAB sibling note still says "Only the 2024 locked input". That is true of Build J, but the pinned 2025 input also carries PECOHAB. Editing that line conflicts with #1146's rewrite of the adjacent derived_A_EXPRRP line, so it waits until #1146 merges, or until d1099 rewrites the entry.

Disk on the build machine was under 40 GB when the nits went in, so pytest was not re-run for 7f50ba142. The same assertions were checked directly instead:

• the reason prefix, classification, required_person_columns, hermetic_inputs and the rejection prefix;
• manifest reason == gap reason, through load_release_input_coverage_manifest;
• Cite US and UK spec sources by symbol or at a commit, never by a drifting line #1146's citation guard.

A delta re-review is running against 7f50ba142.

Evidence (computed 2026-10-08, members checked against their pins)

income year (member)	PERRP partners 43/44/46/47	A_EXPRRP == 13	PERRP inside code 13	PERRP 43/46 → code
2022 (pppub23.csv, 19b56537…)	3,889	4,312	44: 2,628 · 47: 184 · 55: 1,500	12
2023 (pppub24.csv, 21a2b9e0…)	3,893	4,387	44: 2,597 · 47: 189 · 55: 1,601	12
2024 (pppub25.csv, 06921fe8…)	3,751	4,257	44: 2,464 · 47: 179 · 55: 1,614	12

Counts are unweighted person records. Census labels come from api.census.gov/data/2024/cps/asec/mar/variables/{PERRP,A_EXPRRP}.json: PERRP 55 is "Housemate/Roommate without Relatives" and A_EXPRRP 12 is "Nonrel w/own rels in hhld". The 2025 H5 (4c5a3218…) carries A_EXPRRP (3,816 code-13 records), PECOHAB and A_FAMREL, but not PERRP. Its Census member pppub26.csv was not on disk and was not checked. The reproduction script and its output are in ~/reviews/microcosm-unmarried-partner-gap-2026-10-08/.

Checks

• test_us_relationship_inputs.py (4 passed, 0 skipped). The 2022-2024 locked H5s are mounted locally, so test_locked_artifacts_confirm_source_presence_and_absence ran rather than skipping. test_unmarried_partner_exclusion_is_source_unavailability asserts the manifest reason equals the gap reason.
• engine_free/shared/test_release_input_coverage.py, engine_contract/us/test_release_input_coverage.py, test_us_parity_reference.py, test_us_register_consistency.py, test_us_asec_sources.py, test_country_spec.py, microcosm-data test_contract.py: 579 passed in total with the relationship file, 0 skipped, 0 failed.
• Manifest regenerated with uv run python tools/build_us_release_input_coverage_manifest.py. Its diff is the one reason line.

Invariant (test-enforced, unchanged): the manifest's reviewed_exclusions reason equals the gap register's reason byte for byte.

🤖 Generated with Claude Code