cas-smo-v34

Fork of cas-smo-v33 (itself a fork of king cascade-private/jenn11). Same playbook: arm unused families already in the body, take mass from oversupplied step_level, keep pushing floor-biased censor.

Full-heat scores (this pool)

Pool: dataset/snapshots/2026-08-09-block-8805600, 0.5h heat, seed 0, CUDA.

Generator	geomean	vs jenn11
jenn11 (king)	0.23924	—
cas-smo-v33	0.23591	-0.00333
cas-smo-v34	0.23487	-0.00437

Changes vs v33

Lever	v33	v34	Why
held_rate	0	1.5%	Already in body + _CLEAN; near-constant / 0–3 grid steps. Frozen bike and policy-rate rows that step_level overshoots.
weekly_demand	0	1.5%	Already in body + _CLEAN; tight-interval weekly+daily cycles.
step_level	20.6%	17.6%	Donor mass (held-run still oversupplied vs pool).
censor_upper_frac	0.25	0.12	Continue floor-bias; pool holds sit in the bottom decile, ceiling plateaus are rare.

Maximal-coupling dispatch means the +3pp specialist mass only reassigns ~that fraction of rows.

cascade score ./cas-smo-v34 \
  --pool-dir ./dataset/snapshots/2026-08-09-block-8805600 \
  --device cuda --skip-verify --seed 0