[MaxGhenis]

Part of #1215 (the selector that contradicts its own declaration; the ownership problem stays open there).

The mismatch

us/source_stages.json declares the wealth stage's carry as head_carry with condition: is_household_head. relationship_inputs.py sets is_household_head from P_SEQ == 1, the CPS reference person, and the housing stages use it as the head. The runtime selector (scf_wealth._household_head_mask, and its SIPP twin sipp_financial_assets._household_head_mask) took the lowest A_LINENO instead, and a test pinned that rule.

On the certified populace-us-2024-spm-20260909 the two rules disagree in 243 of 57,240 households. In the 2024 income year, the one vintage whose CPS file reports A_EXPRRP (the 2022 and 2023 frames synthesize it from line 1), is_household_head picks the reference person in every household, and the lowest-line rule misses 73. For example, in household 884 (#1215), the 71-year-old wife on line 1 carries $17,500 and the line-2 reference person carries nothing.

Change

• sipp_financial_assets.household_head_carrier_mask picks one carrier per household. Where is_household_head marks exactly one person in the household, it takes that person. Otherwise (no flag column, or zero or several flagged people), it falls back to the lowest A_LINENO, then row order, as before. It reads missing flag values as False and returns an empty mask for an empty frame.
• scf_wealth._household_head_mask calls it, so the SCF and SIPP halves, and the SCF-wealth gate's audit check, use the same person.

#1215's larger problem stays open: one household-level vector still sits on one person, and fixing it takes a change of draw unit and a validation pass.

Invariants

Hypothesis tests each of these with 100 examples:

• Exactly one carrier per household, for any mix of flags and line numbers.
• Where exactly one person in a household carries the flag, the selector picks that person.
• Permuting rows moves the carrier with its row (line numbers unique within a household, as in the CPS).

Example tests pin that the flag beats a lower line number, that zero or two flags fall back to the lowest line, that both halves return the same mask, and that the selector handles nullable-boolean flags with NA and an empty frame. The existing test_household_head_mask_is_one_per_household_at_lowest_line (no flag column) still passes unchanged.

Deliberate re-pins

The seed protocol attests the source bytes of seeded stage modules, so editing them moves these pins. On main at 3f46d5958, tools/repin_us_spec_engine_contracts.py recomputes them:

Pin	Main	This PR
EXPECTED_HASHES["seed_protocol"]	5840ac6c…ff53	3e432297573840bbae497370efe3e7cafed5c2c571db029208d10722035a7470
EXPECTED_HASHES["seed_map"]	fa48af06…6b37	5f8b7c620abee639d31f3dd50ed61e9c9a0daad89b8b4886d485b71d5cb882bd
Synthetic loader spec_sha256 golden	5b39d83c…d36e	ddb925d9603cd269c358f46ff433d5c4adf67991aa17782331a23c0b55767a05
Real US spec_sha256 (engine-workflow test)	4ccbd017…38b2	915d29b0a5419d8e9ff4c9945fb5a92fb18576440dbbbcef2a7680546cfc4441

tools/spec_engine_coverage.py regenerates docs/evidence/spec-engine/us-f0-coverage.json and reports 42,394/42,394 configuration fields and 41/41 inventory checks. The PR changes no RNG rule, seed or QRF target.

Effect

Future builds only. The 243 households' vectors move to their reference person, and both the asset draws and the household net-worth draw condition on that person's predictors (the net-worth QRF reads the same head mask). A frame that already carries an old blend audit can pass through the stage unchanged (the stage's healing rule), so a reused in-memory frame keeps the old carrier; canonical builds draw fresh. The PR runs no release or build.

Release-build rule

docs/us-release-build-rule.md asks a pull request "large enough to change how the dataset is built" to carry a from-scratch build and certification receipt from its own tree before it merges. This PR changes which person carries the asset vectors in future builds, and it carries no such receipt. d2781 asks whether its tests may stand in for that run on the terms of d2647 and d2729: the next certified US build on main would become this PR's from-scratch run.

Verification

At eff5d8343, which merges main at 3f46d5958 and re-pins:

• The spec-engine loader and field-usage tests, test_us_scf_wealth.py, test_us_sipp_financial_assets.py and the multispine constants test pass (68 tests).
• ruff check ., tools/ci_test_plan.py verify and verify-workflow, and tools/graph_acceptance_burndown.py --verify pass.
• tools/repin_us_spec_engine_contracts.py --check timed out in its 300 s clean-worker import while host load sat near 150. The reviewer recomputed all four pins from the head's bytes instead, and CI's engine-us job runs the coverage test.

5932199a5 restores a comment above the seed pins and changes nothing else; the loader tests pass at it (10 tests).

At db53b5f6b, the engine-free US and shared suites passed (-n 4), with the telemetry subprocess test deselected locally because it hangs on this host; CI runs it.

Review

• GPT-6.1 Sol approved 0b15232c2 with nits, which db53b5f6b fixes, and a second review approved the delta to db53b5f6b.
• An Opus 5.5 reviewer approved db53b5f6b..5932199a5 with nits. It recomputed all four pins from the head's bytes, and reproduced main's pins from main's bytes as a control. It flagged a dropped Liquid assets sit on the lowest-line person, against the stage's declared is_household_head carry; 13.1M SSI-relevant adults test at $0 resources by construction #1215 comment, which 5932199a5 restores, the repinner's two-pass count, which Repinner needs two passes to pin covered_item_count #1530 tracks, and main's lowest-line education-debt selector, which Liquid assets sit on the lowest-line person, against the stage's declared is_household_head carry; 13.1M SSI-relevant adults test at $0 resources by construction #1215 now records.

axiom: n/a: microdata construction only; no policy rule changes.

🤖 Generated with Claude Code

[MaxGhenis]

Independent review (GPT-6.1 Sol) of 0b15232c: APPROVE WITH NITS, no blocker; it confirmed caller availability, fresh blend/audit consistency, the four re-pins (all 53 draw-site contracts unchanged), and that the archived pipeline took the first stored row and never minimized A_LINENO. Dispositions in db53b5f:

• String flags ("False", "") no longer count as heads; only booleans and nonzero numbers do.
• Flags are counted within the fallback's own sort groups, so a missing household id is its own household in both rules.
• New blend-plus-gate test with the declared head on line 2 (all three leaves on the head, zeros elsewhere, no blend-audit disagreement).
• PR body now says net-worth conditioning follows the same head, and that a reused frame with an old audit keeps its old carrier (canonical builds draw fresh).
• sipp_vehicles still picks the lowest A_LINENO on its own; that predates this PR and is noted on Liquid assets sit on the lowest-line person, against the stage's declared is_household_head carry; 13.1M SSI-relevant adults test at $0 resources by construction #1215.
  Digests re-pinned for the new bytes; spec_engine_coverage --check 41/41; wealth tests 42 passed; loader tests 10 passed.

[MaxGhenis]

I'm handing this PR to the Microcosm hub (session local_86a67d6b-abf3-47b6-8f7b-6247f14e222a) from session local_ccd7a569-162a-4b4c-a12a-81027d77e07e.

• At head 5932199a5ffa4c4cda160261416b460ad2337f3f, an Opus 5.5 reviewer approved db53b5f6b..5932199a5 with nits (~/microcosm-launch-evidence/dynasim-parity-20260909/pe-us-depletion-cut-population-20261001/frame-structure-20261011/lane-out/review-1221-head-eff5d83.md, whose addendum covers 5932199a5; working notes in ~/reviews/microcosm-1221-review-2026-11-11/runs/delta-merge/). It recomputed all four pins from the head's bytes and reproduced main's from main's bytes. 5932199a5 restores the Liquid assets sit on the lowest-line person, against the stage's declared is_household_head carry; 13.1M SSI-relevant adults test at $0 resources by construction #1215 comment it found missing, Repinner needs two passes to pin covered_item_count #1530 tracks the repinner's two-pass count, and Liquid assets sit on the lowest-line person, against the stage's declared is_household_head carry; 13.1M SSI-relevant adults test at $0 resources by construction #1215 now records main's lowest-line education-debt selector.
• Locally at eff5d8343, 68 tests in the loader, field-usage, SCF-wealth, SIPP-assets and multispine constants files pass, and so do the four lint steps CI runs; at 5932199a5, the 10 loader tests pass. tools/repin_us_spec_engine_contracts.py --check timed out in its 300 s clean-worker import while host load sat near 150. On 5932199a5, CI has passed lint, select-countries, wheels and installed-emitter, and the engine and UK jobs still run.
• The pins match main at 3f46d5958, which doesn't contain Impute US humanitarian immigration statuses to cited stocks (H.R.1 channels) #779. The hub set this PR's slot after Impute US humanitarian immigration statuses to cited stocks (H.R.1 channels) #779, or after 2026-11-13 00:00Z if d2729 stays unruled. If Impute US humanitarian immigration statuses to cited stocks (H.R.1 channels) #779 lands first, re-pin on the merge with main; the repinner's full pass needs two runs, since one run writes a stale covered count (Repinner needs two passes to pin covered_item_count #1530).
• The PR changes which person carries the SCF and SIPP asset vectors in future builds, so docs/us-release-build-rule.md asks it for a from-scratch receipt, and it carries none. d2781 asks Max to waive that receipt on the terms of d2647 and d2729, and closes when this PR merges or closes.
• Liquid assets sit on the lowest-line person, against the stage's declared is_household_head carry; 13.1M SSI-relevant adults test at $0 resources by construction #1215 records two more selectors on main that start from the lowest A_LINENO, in scf_education_debt.py and sipp_vehicles.py, for a follow-up that starts both from is_household_head. Liquid assets sit on the lowest-line person, against the stage's declared is_household_head carry; 13.1M SSI-relevant adults test at $0 resources by construction #1215's ownership problem stays open.