Money moves in steps the networks can absorb.

One jump ruins the read

Halving one channel and doubling another changes two variables at once, so next week cannot tell you which of the two moves was right. Networks also re-enter learning after a large change, so the first days of new data describe the disruption rather than the channel. Stepped movement keeps the before-and-after comparison usable.

The threshold is an account setting

Daily movement limits are configured per account and per campaign; the model never sets them at runtime. A step that would exceed the limit is truncated to it, or held for approval where the Constitution requires one. Protected campaigns sit outside reallocation altogether.

Every step is reversible by construction

Each step stores the prior state, the inputs behind the decision and the identity that approved it. Reverting restores the earlier split in one action, which is itself logged. Minimum evidence rules mean a step with too little supporting signal is never proposed in the first place.