Rescene: band-limited stochastic forcing turns a frozen neural weather operator into a climate emulator

Machine-learning weather models now match or exceed ECMWF's high-resolution forecast in the medium range, but integrating them freely beyond their training horizon causes blow-up, drift, or loss of the seasonal cycle, and retraining for stability is expensive. We ask what can be recovered from a strictly frozen backbone. We present Rescene, a 0.4 M-parameter wrapper around a frozen 1.5°, 6-hourly vision-transformer operator, comprising a deterministic "slow clock" that blends the forecast toward a lead-aware day-of-year climatology and a generative head that adds a spectrally shaped stochastic perturbation at every step. The deterministic wrapper alone is stable for decades but collapses daily variability to 40% of ERA5. Adding the generative head restores 126% (Z500) and 130% (MSLP) of observed daily variability with pattern correlations of 0.89 and 0.92, recovers 82% of observed blocking frequency, keeps the ensemble calibrated (spread–skill 0.78–0.97 from day 7 to day 90), and integrates for 100 years with no detectable drift (+0.008 ± 0.014 K per century). Because the perturbation is band-limited to total wavenumber k ≤ 20, small scales are never forced, yet realistic k ≥ 20 power is sustained: a direct energy-budget decomposition shows the frozen operator supplies 28 times more energy than the perturbation at k ≥ 40. Large ensembles of century-scale climate variability thus become affordable without retraining the backbone.