Two schools run production scheduling. One freezes the plan, this week is committed, disruptions wait at the fence, because stability lets materials stage, crews set, and suppliers sync. The other rolls, the plan re-computes as reality moves, because a frozen plan meeting a down machine produces fiction, and the floor should chase the best current plan, not last Monday's. Both are right, at different distances from execution.
What freezing actually buys
Commitment is infrastructure. A frozen horizon lets purchasing cut POs against certainty, staging pre-kit materials, crews plan setups, and attainment mean something (a plan that moves cannot be missed, which is convenient and useless, see attainment). Multi-plant operations and long-lead component makers, firearms manufacturing is the classic case, freeze weekly because allocation across plants needs a stable target.
What rolling actually buys
Truth. A rolling plan absorbs the 9 a.m. breakdown, the late trailer, the hot order, and always represents the best available sequence. High-mix job shops and short-cycle plants roll because their disruption rate makes any freeze longer than a day fictional by Wednesday.
The fence, and where to put it
The workable answer is both: frozen inside a fence, rolling outside. Inside (today, maybe tomorrow): committed, materials staged, changes require a human override with a reason. Outside: fluid, re-optimized freely. The fence's width should equal the horizon where re-planning causes more damage than the disruption: if moving tomorrow's sequence strands staged materials and breaks a validated allergen chain (allergen sequencing), tomorrow stays frozen; if it costs nothing, roll it.
Live data moves the fence
The freeze exists to protect against uncertainty, and uncertainty is mostly ignorance: not knowing true inventory, machine health, or real changeover costs makes late changes dangerous, so plants freeze wide. With live material truth, machine status from the PLCs, and measured changeover times, re-planning near execution becomes safe, and the fence narrows. The propose-and-approve pattern completes it: the system re-proposes freely outside the fence, proposes-with-warning inside it, and the planner owns every override (full architecture). Plants typically land at a one-day hard freeze with honest data, down from a week of protective fiction without it.