An O'Reilly Radar article argues the AI stack has a "preservation gap": the industry invests heavily in evaluating capability but rarely preserves what is needed to understand a past execution. The author distinguishes reproducibility (getting the same result again) from historical reconstructability (identifying the conditions of the original run) and introduces Orrery, a reference architecture that records runtime bindings as they occur — model endpoint, input transformation, retrieved context, tool contract, policy and runtime configuration — along with resolvable identities for the historical states they reference. The piece contends that if an external effect is committed while its usage record never reaches durable storage, the system offers only best-effort historical evidence, so preservation failure needs explicit semantics; it closes by mentioning AIGov Core treating reconstructability as a test.