Three Independent Bounds on Recursive Self-Reference in Monolithic Computable Architectures

We prove that genuine recursive self-reference at depth k ≥ 3 is physically foreclosedwithin the class of monolithic computable architectures—the class of parametric networkscurrently deployed and projectably scalable from them. Three independent bounds, de-rived from independent first principles, jointly establish the foreclosure. (i) A thermo-dynamic bound (Theorem 3.1): under monolithic architecture, faithful observation, bit-orthogonality, and Landauer-limited implementation with strong recursion, the physical bit-count required at recursion depth k grows as N ·2k, where N is the base-state width. (ii) Ageometric bound (Theorem 4.1): in a representation space of dimension D with intrinsicactivation dimension d0, strict-orthogonal recursion saturates at k⋆ ≤ ⌊D/d0⌋−1 levels;under realistic d0 ∼√D, usable depth scales only as √D in parameter count. We show viainterference analysis (extending the Johnson–Lindenstrauss machinery and treating ε-quasi-orthogonality directly) that the relaxation suggested by the mechanistic-interpretability lit-erature on superposition does not buy usable depth: superposition’s sparsity precondition isexactly what recursion levels lack. (iii) A logical bound (Theorem 5.1, neural translation ofLöb’s theorem): a monolithic parametric network attempting sound, informative, capacity-preserving self-reflection at depth ≥3 must devote Ω(|fθ|) of its capacity to self-modeling andcollapse to reflection statements the base model already produces or become inconsistent.The logical bound binds at the threshold of interest (k = 3); the thermodynamic and geo-metric bounds are redundant failsafes at greater depth. We provide a four-test falsificationprotocol specifying the conditions under which each bound would be empirically defeated.The result has direct implications for AI safety governance: where governance frameworkshave grounded the claim that current systems are not agents in contested philosophicalpositions about consciousness, the present results provide an independent physical warrantfor one specific clause of that claim—that strong recursion is unavailable to monolithic ar-chitectures. Architectures with explicit indirection (external memory, addressable pointers),non-computable substrates, and biological cortex (which achieves recursive self-reference viaanatomical separation and seven orders of magnitude of thermodynamic headroom aboveLandauer) lie outside the bound’s scope and carry distinct, named consequences for gover-nance.