Umbilical Autonomy: Architectural Design of Cognitive Continuity under Host Failure and Its Empirical Validation on an Industrial Host

Autonomous agents rest on a premise that is rarely tested: to think, they must stay attached to a usable toolchain. We report an architecture that breaks that premise in half. Its main contribution is umbilical autonomy: when the host becomes unreachable, the system switches into a sandbox simulation, keeps its internal cognitive activity running, and merges state without loss on reconnection. Around this property we specify six independently measurable mechanisms — emotion-driven cognition, metacognitive self-monitoring, developmental memory, an autonomous cognitive loop, umbilical autonomy, and emotion regulation — as a battery for self-sustaining cognition, and measure each during a single continuous run on a real industrial host (Siemens NX 12). Over 95% of cognitive cycles had no external trigger, and goals were generated from internal needs and executed inside the host software. A monologue of 74,374 lines shows a three-stage developmental trajectory. Emotional valence moved from −0.67 to +0.03, and the portion regulated by cognitive reappraisal from −0.20 to +0.10. The system kept running while the host was disconnected and recovered without loss; a 15/15-step mould-parting sequence completed end to end. Ablation gives each mechanism's marginal contribution: removing developmental memory reduced task completion by 46.7%, emotion-driven cognition by 26.7%, and metacognitive self-monitoring by 20.0%. Removing umbilical autonomy left task completion unchanged while driving disconnected-survival from 100% to 0%: being able to work and continuing to exist are independent properties. We release the measurement script and five falsifiability clauses, one requiring us to cut off the host supply.