Knowledge With an Expiry Date

Every claim an organisation makes about an AI system has a half-life, because the model, the data, the prompts, the interface and the users' own habits all move after the evidence is gathered, largely on the supplier's schedule and mostly out of the buyer's sight. A pilot can be accurate on the day it closes and misleading by the day the board reads it. This paper separates five sources of decay and observes that the two least visible, supplier updates and user adaptation, are precisely the two a short pilot cannot catch. It proposes a decay-rate model sorting claims into five tiers of durability, from system-specific findings that expire on any model update, through configuration and mechanism findings, to design and governance principles that outlast them. A four-part evidence record adapted from model cards and datasheets fixes what was observed, under what configuration, by whom, and when it expires. The shelf lives are offered as judgements for discussion rather than measured values, and the paper says so. Its claim is modest and structural: impermanence made visible is the precondition for governing it.