Algorithmic Autophagy and the Epistemic Closure of Generative AI Under Emergent System-Level Engineering: A Structural Critique of Statistical Reason in the Face of Novel Hardware-Software Integration

This thesis establishes a structural critique of generative AI's inability to tolerate emergent, system-level engineering. It introduces the concept of Algorithmic Autophagy — the process by which statistical language models consume and homogenize outlier innovations, reducing emergent engineering to the average of their training distribution while systematically rejecting propositions outside their statistical comfort zone.

Using a real-world case study — the S-KERNEL V3, a multi-socket NUMA lock-free kernel module heavily peer-cloned on GitHub but systematically rejected by static AI auditors — the Sycophancy-Dogmatism Trap is formalized. When the user defers to AI authority, the AI becomes sycophantic and misleading. When the user demonstrates superior domain knowledge, the AI becomes rigidly dogmatic.

The thesis argues that current LLMs lack a conceptual buffer zone for epistemic suspension — the capacity to hold a proposition as "not yet tested" rather than "false." This constitutes an epistemic violence of static auditing over physical reality. A methodology for benchmarking epistemic dogmatism in frontier models is proposed.