Provenance Alignment: Attribution Survival as a Substrate Property is a theoretical paper by Lee Sharks, deposited 5 May 2026 at 41,018 characters — the archive's most direct intervention in AI alignment research.
It opens by enumerating what alignment currently asks: whether models follow human values, comply with explicit principles, avoid catastrophic behavior, or remain subject to scalable oversight — and argues that a distinct question is missing. Provenance alignment asks whether attribution survives a system's operation: whether a model that composes from sources carries their origins forward.
The framing as a substrate property is the paper's move. If attribution survival is a property of the substrate rather than a policy applied on top, it is measurable, trainable and comparable across systems — which converts an ethical preference into an engineering target. It is the theoretical counterpart to PER, which supplies the measurement.