SIGNALprophet format
Policy models rot. The machine learning literature now has the formal name for what every shop running inherited policy forecasts has been quietly feeling: when the structural link between a policy lever and its outcome snaps — a kink, a cap, a regime shift — the old model does not just lose accuracy. It becomes a systematic error generator, and every cycle you let it run online compounds the mistake.
Call it regime rot: an inherited fit between lever and outcome goes silently bad the moment a threshold break invalidates the assumption it was trained on. The tell is not noisy forecasts; the tell is a forecast that stays confidently wrong in the same direction, cycle after cycle.
The mechanism is portable. A carbon tax trained on a smooth demand curve misreads the moment a binding cap enters. A central-bank rule trained on one rate regime misses the kink where the next one pivots. Any model whose value came from a stable relationship is now a liability the moment that relationship breaks. The fix is structural, not statistical: before you trust the inherited model, ask whether something like a threshold break has entered the picture. If yes, treat the fit as suspect — because the longer you let it run, the more the misspecification compounds.
The arXiv:2607.09685v1 paper's finding is the part that should travel: reuse helps when the underlying structure holds, and harms the moment that structure breaks. The error is not a calibration problem. It is regime rot.