Split the child-support gate band: expense floor 0.003

packages/populace-build/src/populace/build/us_runtime/child_support.py

@@ -118,9 +118,23 @@
         "weight",
     }
 )
+#: Per-column plausibility bands on the pooled (ASEC + PUF-clone) person share.
+#: The halves are NOT symmetric and must not share a floor: CPS asks custodial
+#: parents about receipt but relies on payer self-reports for expense, so the
+#: expense share runs structurally lower. Measured at the first full-scale run
+#: through this gate (Build M base, 2026-07-13, seed 0): raw ASEC person share
+#: 0.66% expense / 0.98% received; the QRF draws a covariate-consistent 0.27%
+#: on the PUF tax-detail clone channel; the ~50/50 pool blends to 0.466%
+#: expense — non-degenerate, deterministic, and reproduced bit-identically with
+#: the leaf-storage guard disabled. A shared 0.005 floor mislabeled that as
+#:  degenerate. The retired us-data pipeline's sequential imputation shipped
+#: 1.99% (extended CPS) and 2.96% (enhanced CPS) against the same 0.62-0.66%
+#: survey marginal — prevalence drift, not a reference; whether prevalence
+#: should instead be seeded to OCSE payer counts is populace#417. Degeneracy
+#: per channel is still enforced by _CHANNEL_NONZERO_SHARE_BAND below.
 _NONZERO_SHARE_BANDS: dict[str, tuple[float, float]] = {
     "child_support_received": (0.005, 0.15),
-    "child_support_expense": (0.005, 0.15),
+    "child_support_expense": (0.003, 0.15),
 }
 _CHANNEL_NONZERO_SHARE_BAND = (0.002, 0.15)
 

packages/populace-build/tests/test_us_child_support.py

@@ -481,3 +481,24 @@ def situation(received: float, expense: float) -> dict[str, object]:
     assert active.calculate("snap_child_support_deduction", "2024-01")[
         0
     ] == pytest.approx(200.0)
+
+
+def test_nonzero_share_bands_are_asymmetric_by_half() -> None:
+    """The expense floor sits below the received floor, deliberately.
+
+    CPS asks custodial parents about receipt but relies on payer self-report
+    for expense, so the expense share runs structurally lower: at the first
+    full-scale gate run (Build M base) the pooled expense share was 0.466% —
+    faithful ASEC channel (0.660%) blended with a covariate-consistent clone
+    channel (0.272%) — while received sat at 0.854%. A shared 0.005 floor
+    mislabeled that expense signal as degenerate (see populace#417). This pins
+    the split so a refactor back to one shared band fails loudly.
+    """
+    from populace.build.us_runtime.child_support import _NONZERO_SHARE_BANDS
+
+    assert _NONZERO_SHARE_BANDS["child_support_received"] == (0.005, 0.15)
+    assert _NONZERO_SHARE_BANDS["child_support_expense"] == (0.003, 0.15)
+    assert (
+        _NONZERO_SHARE_BANDS["child_support_expense"][0]
+        < _NONZERO_SHARE_BANDS["child_support_received"][0]
+    )