Machine learning systems combine scaled models with data curation, memory, tools, feedback, persistent state, and control policies. This paper asks whether a pre-scientific developmental model can nevertheless serve as a heuristic engine: a disclosed source of candidate causal variables and dependency hypotheses, judged downstream rather than treated as evidence. From one contemplative schema we derive six proposals: faculty-ordered curriculum, controlled perspective supervision, relational continuity, lifecycle consolidation, sequential commitment stabilization, and legible-pattern worlds. They are not six homogeneous stages. Some are learner capacities, some are curriculum properties, and lifecycle consolidation is an intergenerational operation. We therefore separate a floor of independently testable interventions from a spine proposing a developmental dependency order. The correct null is not that parameter count predicts one rising axis. It is that, after compute, data, architecture, persistent state, supervision quality, and optimization are matched, the proposed variable explains no additional variance. To reduce retrospective analogy, the program also requires a provenance protocol: timestamped derivation, an explicit source-to-variable translation rule, and preregistration of each hypothesis before confirmatory literature search and testing. The paper is a research agenda, not an empirical report. The source earns standing only if prospectively derived interventions produce effects that survive strong contemporary baselines.