ACS-multispine staging: SSI probe finds spine-inconsistent incidence (2.6% vs 1.9% baseline rate) and weight-composition skew (ASEC 1.5 persons/hh, ESS 0.3%)

[MaxGhenis]

Context

SSI resource-limit probe of the buildl ACS-multispine staging artifact (_buildl-runtime/acs_2024_1yr_calibrated.h5, local/unpublished: 3,583,746 persons / 1,606,726 households, ASEC+ACS spines, 742-target option_b_soi_totals calibration, simulation_ready=true). Method: exact 40k-household chunking through pe-us 1.764.6 Microsimulation (41 chunks), applying the lane's own engine-pass contract — reviewed_null_fills register + held_back_columns projection. TY2026, SSI-only metrics. Same harness as the #356/#374 records.

Headline (consistent with the certified-artifact family)

scenario	multispine staging	Build J certified (4.20.4)	dense f0af251	anchors
baseline SSI	8.31M / $71.8B	6.71M / $57.1B	8.05M / $63.6B	SSA ≈7.4M / $61B
$10k/$20k	+$5.96B / +845k	+$5.69B / +797k	+$1.60B / +499k	OACT ≈$1.0B / CBPP 224k
remove test	+$16.62B / +2.41M	+$18.19B / +2.45M	+$16.05B / +3.23M	ECPS +$15.75B

The $10k/$20k delta sits in the SCF-only family, as expected — the ACS spine's asset columns come from the same SCF transfer — so #374's blend/refit applies to this lineage too. Full removal stays in the robust +$16–18B band. Baseline SSI overshoots SSA (this surface carries no SSI target; the incumbent surfaces sat closer).

QA findings (pre-launch signals)

1. Spine-inconsistent SSI incidence. Baseline SSI rate: 2.59% on the ACS spine (6.08M of 234.7M person-weight) vs 1.85% on the ASEC spine (2.23M of 120.3M). Ten-twenty delta per baseline recipient: ACS $821 vs ASEC $433 — the same reform scores ~2× as intensely on ACS-spine records. Points at transfer-stage distribution mismatch (assets/disability/unit structure) between the spines; relates to the Canonical multispine operator ordering: assemble spines first, then clone/impute/derive/seed/simulate/calibrate #395 operator-ordering question.
2. Calibrated-weight composition skew. Implied persons per weighted household: ASEC spine 1.50 vs ACS 2.93 (each spine holds 80.08M household-weight; person-weight splits 120.3M vs 234.7M; total 355.1M vs census ≈346M). With effective_sample_size = 4,853.9 (ess_fraction 0.003), the solve is concentrating weight on few, small ASEC households — the runbook's decision-5 l2-sensitivity trigger looks warranted before this surface ships.
3. Consumer footnote on reviewed_null_fills.json: enum-kind entries log fill_value as a repr ("'CITIZEN'"); a consumer that applies the log verbatim writes quoted strings that the engine's enum encoder rejects. Logging raw values (or documenting the repr convention) would prevent that.

Repro scripts: ssi-asset-limits .local-results/populace-rerun-scripts/pass_h_*.py; outputs in populace-buildl-acs-multispine-* incl. by_spine.csv. Relates #395, #374, #368/#369.

[MaxGhenis]

Per-target residual audit of the same staging artifact (recomputed from checkpoints/national/{weights_latest.npz, target_frame_lean.h5, targets.json}; matches the diagnostics' 87.87% within-10%). 90 of 742 targets miss by >10%, in four clusters:

family	n off >10% / n	median |err|	worst
state population	4 / 51	0.7%	CA +23.0%, NY +17.6%, NJ +14.2%, PA +13.7%
CD population	11 / 436	0.7%	CA-17 +347%, CA-05 +326%, CA-36 +245%, NY-24 +234%, PA-08 +210%
SNAP caseloads	28 / 51 — all over	10.8%	NJ +162%, MN +156%, ID +129%
SOI PTC (state amount+returns)	30 / 102 (26 over)	0.6%	NY amount +374% / returns +363%; TN +116–153%; IN, MS >+100%
SNAP benefits	6 / 51	0.2%	NJ +110.7%
Medicaid enrollment	11 / 51	0.3%	NJ +105.9%

Patterns worth noting before this surface ships:

1. The four states that miss population (CA/NY/NJ/PA) are the same states whose CDs blow out 2–4.5× and NJ misses across every admin family at ~2×. Weight mass is systematically parked in the wrong (large) states, consistent with the ESS-0.3% concentration finding above — the solve appears trapped by the ratio-5 cap from a mass-misallocated warm start, and the per-target loss weighting lets dollar targets win over population/caseload counts.
2. SNAP caseloads are over everywhere they miss (28/28) while SNAP dollars sit at 0.2% median — the solve reconciles benefit totals against an over-enrolled caseload by shrinking per-household benefits; take-up on the ACS spine isn't constrained (the Recalibrate SNAP take-up per state to FNS household caseloads #376-style state take-up recalibration doesn't reach this frame).
3. State PTC is systematically over (26/30 misses, NY 4.7×) — ACA/marketplace enrollment on the ACS spine looks over-assigned (take-up flags are among the reviewed nulls → engine defaults).

Full 90-row miss table: multispine_target_misses.csv (repro in the ssi-asset-limits rerun scripts). Given the SSI-probe findings above and these residuals sharing one plausible root (weight allocation), a re-solve with l2>0 (the runbook's own decision-5 sensitivity) + population-target re-weighting seems like the cheapest first experiment.