Substrates: The Dies and Molds of Knowledge Work

The fastest way to find hidden compounding in AI operations is to separate substrates from the projects that happen to create them. Prompt templates, eval sets, runbooks, classifiers, and playbooks become real assets when they have versions, owners, holdout metrics, and pinned consumers. The hard part is capital allocation: fund the substrate, maintain it, and stop pretending every downstream use starts from zero.