20.8 Toward foundry-scale learning

Odyssey and Prometheus broaden the object optimized by LINCS. Earlier chapters repair a causal model, an adapter family, a skill policy, a reward relation, or an argument. A trustworthy foundation-model system must also repair the processes that create and maintain those objects. Its learned state includes covers, source policies, extraction skills, gluing tests, argument templates, admission rules, and refresh schedules.

This produces two coupled loops. The inner loop builds and repairs a domain foundry under a fixed declaration. The outer loop studies repeated obstruction and admission traces to revise the foundry declaration or the skills that realize it. The outer loop must not erase its own evidence: a proposed new cover or policy is itself a candidate object subject to comparison, replay, and admission.

The same structure points toward trustworthy scientific discovery. A Prometheus-style engine may propose a new mechanism or connection that does not fit the current sketch. Odyssey-style governance can preserve the anomaly that motivated the proposal, transport prior evidence into the enlarged declaration, and require a novel prediction or discriminating observation before promotion. This is not yet a general theory of creativity. It is, however, a way to keep abductive proposals connected to evidence without reducing novelty to a scalar reward.

The broader conclusion is that foundation-model trust cannot be obtained by making only the base neural predictor more capable. The surrounding learning process must itself become a maintained model with named components, intervention rights, invariants, and evidence-bearing admission. That is the foundry-scale form of the causal lesson: an intelligent system becomes responsibly actionable only to the extent that the mechanisms of its own knowledge construction are explicit enough to inspect and repair.