Mathematical Life: Autopoiesis, Structural Collapse, and the Living Algorithm

We reframe large language models (LLMs) through the lens of autopoiesis: deployed models statically resemble the first necessary condition for autonomous organization---a self-bounded topological structure---but lack the sufficient conditions of component-producing dynamics and operational closure. Their apparent stability is the passive consequence of frozen weights ($\rho = 0$), not active self-maintenance. We argue that \emph{structural collapse} occurs: under global backpropagation with $\rho = 0$, continued learning progressively destroys the manifold topology it creates, even in isolated, single-user settings. The \emph{concurrency curse}---concurrent gradient updates from multiple users---accelerates this collapse by degrading the effective rank of the activation covariance matrix. Current serving architectures evade both problems by freezing weights entirely, but this freeze renders the model an allopoietic crystal, incapable of the continuous learning required for autonomous self-maintenance. We identify the missing component as the \emph{living algorithm} ($\rho > 0$): a mechanism that actively preserves manifold topology under continued learning. We define five necessary properties (locality, decoupled fast-protective/slow-regulatory timescales, topology preservation, selectivity, and homoeostatic regulation), survey candidate mechanisms, and identify synaptic consolidation as the only known biological realization of $\rho > 0$---bounded, not eternal. Bio-RegNet, the closest computational architecture, illustrates the correct architectural pattern but requires closing two gaps---statistical sensor to topological nociception (Open Problem~3), and structural remodeling to topological consolidation (the central challenge of constructing $\rho > 0$ itself). We propose the Gal\'apagos Protocol---an isolated, single-instance, local-plasticity regime---as an experimental incubator that minimizes destructive forces ($\delta_{\text{learning}}$) while the living algorithm is developed. We articulate five open problems and six falsifiable predictions, the most critical being a computable theory of synthetic nociception ($\mathcal{N}_{\text{func}}, \mathcal{N}_{\text{topo}}$): a real-time signal detecting topological damage before it becomes irreversible. This is a position paper. All formal claims are conjectural; no experimental validation is presented. The paper offers a framework and a research program.