[WesIngwersen]

Why

The domestic L·y = q check (test_compare_output_and_L_y, domestic case) uses y_nab. derive_y_nab back-computes y_nab as (I − A_dom) q from the same A_dom and q the test then compares, so L_dom · y_nab = q holds by construction for any tables. It confirms the Leontief inverse is computed correctly, not that the tables are consistent.

The change

• usa_matrices.derive_y_dom_from_use_final_demand() reads final demand from the Use table only:

  • every final-demand column, with exports included and the F05000 imports column negative;
  • plus the imports that industries use (Uimp row sums).

  The result is the domestic output taken by final demand, exports included. It never sees A or q. Negatives are kept, as in the published tables. The function is uncached and has no consumers other than the test.

• test_domestic_L_y_from_use_final_demand_equals_q checks L_dom · y = q under the canonical config at 1% per sector.

  • Because y is now independent, a sector fails when its Make output differs from its Use row (intermediate use plus final demand). That is the commodity identity Step 6 breaks the commodity output identity: Make vs producer-price Use, $189bn in 2017 to $1.46T in 2022 #1161 is about, plus anything A construction adds (scrap correction, negative clipping, waste disaggregation).
  • It is marked eeio_integration, so CI does not run it. I ran it locally.

What it shows

Sector output is commodity output q, and the gap is L_dom · y − q ($bn).

tables	sum of absolute gap	sectors off by more than 1%	median absolute gap	gap as a share of total output
canonical, MUT v0.5.0_a166975 (2024, before #1161)	1,604	285 of 405	3.35%	3.16%
2024 rebuilt in memory with #1162 (not stored)	78	13 of 405	0.05%	0.15%

On the stored tables, the largest gaps are retail and wholesale (454000 −$220bn, 424A00 +$89bn). This is the margin and tax mismatch #1162 fixes.

After #1162, the remaining failures are:

• mostly the waste-disaggregation sectors: 562111 −9.4%, 562HAZ +41%, 562910 +25%, 562OTH +15%, 562212 −13%;
• 331410 nonferrous smelting, +46%;
• S00402 (used goods), where q is zero.

Waste disaggregation splits each parent sector's output across its children without keeping each child's row identity. This is the same cause as the existing xfail on test_compare_Uset_y_dom_and_q_usa.

The test is xfail until the stored MUT is rebuilt with #1162. Rebuilding it is on hold. Even after the rebuild, the waste sectors will keep it failing, so whether to fix them or exempt them is a separate decision.

Checks run locally

• black --check, ruff check, mypy clean on both files.
• The new test was run with --runxfail under the canonical config. It failed as expected: 286 sectors, the 285 above plus S00402.

🤖 Generated with Claude Code

[WesIngwersen]

@jvendries @bl-young @MoLi7 Our simple Ly=q test was still circular where y=y_nab=(I − A_dom)·q . y_nab has to be derived independently for this to be a valid test.

[bl-young]

Looks good to approve. The old domestic L·y = q path was circular (y_nab = (I − A_dom) q), so it only checked the inverse. Reading y = Y @ 1 + Uimp @ 1 from the Use table is the right independent demand, and under the canonical nowcast config derive_Aq_scaled is a no-op so units match. Keeping it eeio_integration + xfail is fine until the stored MUT moves off a166975 and #1165 lands.

Two small nits, neither blocking:

1. Please leave a one-line comment on the domestic branch of test_compare_output_and_L_y that it only verifies L inversion given back-computed y_nab -- the independent check is the new test. Otherwise the circular path stays an easy place to hide table inconsistency again.

2. Tighten the new test's xfail reason: #1162 is already on main; what still fails on the pinned v0.5.0_a166975 is the pre-rebuild MUT, and after a #1162/#1170 rebuild the leftovers are mostly waste kids (#1165) plus 331410 / S00402. Citing #1161 as the live cause is a bit stale.

[WesIngwersen]

Both nits addressed in ad20d5a:

1. The domestic branch of test_compare_output_and_L_y now says it is circular (y_nab = (I - Adom) q), so it only checks that L inverts (I - Adom), and points to test_domestic_L_y_from_use_final_demand_equals_q as the independent check.
2. The xfail reason now says the pinned v0.5.0_a166975 is the pre-rebuild MUT ($1.6T, 286 of 405 sectors), and that on the Close the Step 6 commodity identity on the balanced Supply table (#1161) #1162 rebuild (feat(config): point the v1.0 configs at the MUT build v1.0.0_5514c3b #1170, v1.0.0_5514c3b) 15 sectors still fail: the waste children (562*, Waste disaggregation breaks each waste child's commodity identity (Make output vs Use row + final demand) #1165), 331410 and S00402.