Controller Scale Is Not Enough for Long-Running AGI: A Workflow Theory with Reusable Certified Libraries

This record contains a theoretical manuscript on long-running AI systems viewed as workflows rather than isolated predictors. The paper studies systems that select tasks, decompose them into typed subproblems, invoke tools, preserve traces, schedule audits, maintain certified records, and operate under drift and partial observability. Its central question is whether scaling the controller alone is sufficient for long-running AGI when replay, validation, and maintenance resources remain fixed. The main negative result is a controller-scale insufficiency theorem: under explicit replay ceilings for reusable audited modules and recursive validation throughput ceilings, controller-only scaling cannot guarantee robust certified coverage of a changing weighted target environment. The positive side is constructive. On bounded-interface typed task families, the manuscript develops a workflow-level theory based on proof-carrying certified entry, audited reusable libraries, delayed monitored target-membership calibration with logged propensities, novelty control on a finite reachable frontier, and record-level maintenance envelopes. These components yield predictable gain, novelty, and loss laws, which are combined through a stopped supermartingale argument to obtain a finite-horizon entrance theorem for certified AGI. The paper also gives a comparison theorem against budget-matched flat direct-control workflows without reusable audited semantics. Even when such systems are granted the same tools, generic memory, retrieval-compression, unaudited caches, latent summaries, and memoization, they face a hard ceiling on certifiable structured boundary mass. A repository-maintenance example is included as a concrete parametric instantiation. Overall, the manuscript argues that workflow organization, reusable certification, replay, validation, and maintenance are mathematically central to long-running AI capability, not only controller scale.