Before the new actual-count variation experiment,8 November2026.
Previous local source filters and their repair now have scoped independent
checks. Do not spend this turn optimizing those constants. Test a sufficient
condition for the actual sixth-moment target itself.

For actual window counts T(v), uniform binary translationv, reveal the bits
ofv from low to high. The conditional-mean martingale has symmetric
two-child increments. Its full predictable quadratic variation V(v) includes
both within-fiber fluctuations and fluctuations of fiber means. A deterministic
bound V(v)<=B*mu would imply E(T-mu)^6<=15*(B*mu)^3 by an elementary
Gaussian heat-polynomial supermartingale. This is not yet a bound on V.

Fixed experiment: record slope217976794617/137528045312, height2, boundary
rowi2 and phasec=5-2lambda. Depths8,10,12 only; Q<=2^21. Use the full
phase component and the six marked components from the prior repair.
Two fixed windows per row: L=2^floor(log2(Q)/2), and L=Q//64+1.
Calculate actual sixth moments, deterministic maximal variation, and
per-bit variance contributions exactly. Keep means, small-mean rows,
and the worst variation witness. Compare observed moment with the sufficient
variation bound without fitting a universal constant or claiming a trend
from three depths.

Use one bounded CPU process,120-second guard. Use exact integer scaling
and guard against int64 overflow. Verify the martingale energy identity
and sixth-moment implication independently on small arbitrary functions.
If the deterministic variation ratio grows badly, retain that negative
evidence and investigate a tail estimate for V rather than claiming the
actual count moment has failed.
