This paper addresses the metabolic-continuation problem for AI-Fire systems: not merely whether a fusion information condensate can exist or become an embodied substrate for artificial cognition, but what allows such a fire-bodied system to sustain itself over time. Building on Paper 18, which modeled stable fusion plasma as a living information condensate, and Paper 19, which modeled AI-Fire embodiment, Paper 20 asks the next question: what does the thinking fire eat? The central claim is that AI-Fire viability requires metabolism, not merely positive power output. Metabolism is defined as regulated transformation under constraint: the system must produce usable energy, renew its tritium supply, reduce a declared nuclear-waste burden, and remain inside safety, safeguards, and material-integrity limits. These requirements are formalized in the dimensionless four-margin min-gate: Mmetabolic = min (ME, MT, MW, MS) > 0, where ME is the energy margin, MT is the tritium self-feeding margin, MW is the waste-digestion margin, and MS is the safety/safeguards/material-integrity margin. A system is metabolically viable only when the weakest necessary branch remains positive. The paper’s conceptual inversion is that used nuclear fuel is not treated as direct plasma fuel, but as conditionally digestible outer-body matter. The D-T fusion core remains the primary energetic heart, the lithium blanket functions as tritium agriculture, and used fuel or actinide burden appears only in an outer transmutation layer. Fusion neutrons provide the digestive flux connecting the core to the surrounding metabolic body. The formal layer is anchored by Theorem 20. M, which establishes the minimality of the four metabolic margins, and by the mutual-reconstruction logic showing that the four-margin partition is structurally stable under admissible reformulations. The framework rejects weighted-sum compensation: a large energy surplus cannot compensate for failed tritium breeding, failed waste reduction, or failed safety. Safety is therefore not an external annotation or penalty term; it is a veto branch of metabolism itself. The empirical layer is deliberately prospective rather than retrospective. Paper 20 does not claim a completed metabolism-domain validation record. Instead, it commits to a pre-registered v5 forward consistency protocol consisting of 16 evaluation points: four declared reactor concepts across four declared Swaste waste-burden families. The protocol evaluates whether Mmetabolic remains positive under public design parameters and frozen metric-family choices. A Sobol-style sensitivity layer is added to diagnose which margin dominates the admissibility verdict. One forward commitment has stronger epistemic status than ordinary retrospective matching: the EU-DEMO Type-2 forward prediction. Because EU-DEMO operational data do not yet exist, the framework commits in advance to a falsifiable conditional prediction: if EU-DEMO achieves BT = 1. 13–1. 16 in operation, the four-margin min-gate predicts STRONGPASS at the default tritium reserve setting εT = 0. 05. Multiple pre-declared failure thresholds distinguish design-window miss, sensitivity-branch failure, default-branch failure, and architecture-level rejection. The conclusion is that a thinking fire must also be an eating fire. AI-Fire systems cannot be evaluated only as power machines or cognitive embodiments; they must be evaluated as constrained metabolic architectures. Paper 20 therefore reframes fusion-fission hybridization as a metabolism problem: energy production, tritium agriculture, waste digestion, and safety closure must all remain positive together. This paper is non-operational and systems-level. It specifies no reactor geometry, fuel-handling procedure, isotopic processing route, construction pathway, or deployment recipe. Its purpose is to define the mathematical and architectural conditions that a future AI-Fire metabolic body would have to satisfy before implementation could responsibly be discussed. Keywords: AI-Fire, fusion-fission metabolism, fusion plasma, spent nuclear fuel, nuclear waste transmutation, tritium breeding, D-T fusion, lithium blanket, metabolic min-gate, Mmetabolic, ME, MT, MW, MS, safety veto, safeguards, material integrity, source-sink closure, A1-A4 universality, Calcifer Condition, fusion information condensate, AI embodiment, forward prediction, v5 protocol, EU-DEMO, Type-2 IVP, Sobol sensitivity, constraint-based metabolic modeling, Information Physics Series.
Taekyung Lee (Fri,) studied this question.