Where does the number come from?
The paper by Zhang and colleagues starts from a definition rather than a forecast archive. Predictability, as they set it out, is how long any information about the evolving state of the atmosphere can be retained if you knew its dynamics, its state at one moment and its future boundary conditions exactly, and did not know the quantum fluctuations of the energy arriving from the Sun. In a closed system that list would be complete and predictability would be unbounded. Sunlight is what breaks the closure: it injects variance of unknown character, and that variance behaves as noise.[1]
The rest is arithmetic on quantities that are actually measured: the total energy held in the atmosphere, and the radiative fluxes crossing its upper boundary. Noise destroys information at the smallest scales first, then propagates upward through the system, and predictability is exhausted once the atmosphere's total energy has been entirely replaced. That turnover buys 57 days beyond today's skillful range, plus a further 57 days of guidance sitting just below today's forecast quality. Add the 14 days forecasts already deliver and the total is 129, with a stated uncertainty of 7 days.[1]
Why the ceiling and a timetable are different things
The paper is careful about what it has bounded. The 129 days describe how long information could in principle survive; when and by how much that headroom becomes operational is left to future improvements in forecast technology. The open flank is one Wei Zhang points at himself: nobody has observed, derived or simulated how errors smaller than those in current forecasts actually grow. The estimate leans on an energy argument precisely because the error-growth route it replaces cannot be checked.[1]
That distinction matters for anyone buying sub-seasonal products. A ceiling at 129 days says nothing about how fast skill inside that range will improve. It says that effort spent past the ceiling has no physical return, and that the binding quantity is the atmosphere's own energy turnover rather than computing power or observation density. Water managers, grid operators and agricultural planners currently buy forecasts inside the 14-day range and hedge beyond it. The paper leaves that boundary where it is, and tells them the headroom above it is finite.[1]
The signal worth watching
Two of the paper's own numbers make a usable test. The first 57-day extension is defined as forecasts at today's quality; the second 57 days sit just below it. If the estimate is describing the physics correctly, operational skill should improve across one connected stretch and then fade into low-confidence guidance around the same place, rather than stopping abruptly. Nothing in the paper says when that stretch gets built.[1]
The measurable claim to hold the paper to is narrow. If, by the end of 2028, published operational scores for the leading global forecast systems still put useful deterministic skill at about 14 days, the ceiling will have explained nothing new about practice. The interesting case is the one where those scores start climbing into the first 57-day block, because then data rather than assumption is testing the energy argument. Either way the atmosphere is doing the spending, and the budget it draws on is the sunlight it takes in.[1]