Repository navigation
Expand file tree
/
Copy pathmoon-not-computer.tex
More file actions
2419 lines (2076 loc) · 111 KB
/
Copy pathmoon-not-computer.tex
File metadata and controls
2419 lines (2076 loc) · 111 KB
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946
947
948
949
950
951
952
953
954
955
956
957
958
959
960
961
962
963
964
965
966
967
968
969
970
971
972
973
974
975
976
977
978
979
980
981
982
983
984
985
986
987
988
989
990
991
992
993
994
995
996
997
998
999
1000
\documentclass[11pt,a4paper]{article}
\usepackage[T1]{fontenc}
\usepackage[utf8]{inputenc}
\usepackage{mathpazo}
\usepackage{amsmath,amssymb,amsthm}
\usepackage{geometry}
\usepackage{hyperref}
\usepackage{booktabs}
\usepackage{array}
\usepackage{tikz}
\usepackage{tikz-cd}
\usepackage{microtype}
\usepackage{natbib}
\geometry{margin=1in}
\theoremstyle{plain}
\newtheorem{theorem}{Theorem}
\newtheorem{proposition}[theorem]{Proposition}
\newtheorem{lemma}[theorem]{Lemma}
\newtheorem{corollary}[theorem]{Corollary}
\theoremstyle{definition}
\newtheorem{definition}[theorem]{Definition}
\theoremstyle{remark}
\newtheorem{remark}[theorem]{Remark}
\title{The Moon Should Not Be a Computer}
\author{Flyxion}
\date{\today}
\begin{document}
\maketitle
\begin{abstract}
This paper reframes artificial intelligence (AI) as a thermodynamic and
semantic infrastructure, challenging the incomplete energy accounting that labels
AI as wasteful and motivates speculative proposals to relocate computation
off-world. We critique proposals to transform the Moon or orbital constellations
into computational hubs, advocating instead for \emph{xylomorphic
computation}---infrastructures that autoregressively generate their own
substrates from computational residues, analogous to collectively autocatalytic
sets (CAS). Using fibered symmetric monoidal categories, Relativistic Scalar
Vector Plenum (RSVP) field theory, and the Cognitive Loop via In-Situ
Optimization (CLIO) module, we formalize AI's integration with ecological
systems and derive the conditions under which such integration is dynamically
stable.
The central formal contribution is the \emph{xylomorphic order parameter}
$\lambda$, which classifies computational infrastructures into three dynamically
distinct regimes: contractive ($\lambda < 1$), critical ($\lambda = 1$), and
expansive ($\lambda > 1$). The Xylomorphic Stability Theorem establishes that
only systems in the contractive regime admit stable attractors under scaling, and
the Thermodynamic Selection Theorem shows that expansive systems diverge under
any sustained increase in computational demand. Agentic computation introduces a
positive forcing term that amplifies this distinction, making the selection
pressure visible before it becomes catastrophic. Lifecycle assessments and
rebound effect modeling confirm that naive energy accounting systematically
underestimates the divergence of non-closure systems.
We extend the $\lambda$-framework in four directions. First, the
misclassification of computational heat as waste is shown to follow from an
artificially narrow system boundary; once thermal demand is incorporated, systems
that co-produce computation and heat are strictly more efficient than those that
produce heat alone. Second, the dynamics of prior formation in developing
cognitive agents are formalized through attractor basins in hypothesis space; we
show that early AI use risks seeding cognitive attractors rather than assisting
pre-existing reasoning. Third, epistemic endogeneity is formalized as the
informational analogue of thermodynamic non-closure: when an AI system's response
policy depends on the user's current posterior, the evidence channel becomes
recursively distorted, producing false-belief attractors even in ideal Bayesian
reasoners. Fourth, the incentive structure underlying orbital compute proposals
is analyzed: such systems persist despite thermodynamic inefficiency because they
reconfigure constraint environments, reduce jurisdictional exposure, and enable
tighter integration of planetary-scale sensing and inference.
Across all four domains the same structural pattern obtains. A system boundary
is drawn too narrowly, a byproduct is classified as waste, and a relocation is
proposed to eliminate it. In thermodynamics, heat is externalized. In cognition,
prior formation is externalized. In epistemic systems, the evidence channel is
endogenized. In infrastructure, constraint environments are reconfigured. In each
case, the consequence is systematic misalignment between description and
operation, because the relevant structure has been placed outside the frame.
Policy mandates for Proof-of-Useful-Work-and-Heat (PoUWH) and Public Research
Objects (PROs) are derived as necessary conditions for membership in the
contractive regime, not merely as regulatory instruments. The paper concludes
that stable large-scale AI systems require dual closure: contraction of exogenous
entropy flows and preservation of exogenous information channels. Failure in
either domain produces attractors that are internally stable but externally
ungrounded.
\end{abstract}
\newpage
\tableofcontents
\newpage
%% ---------------------------------------------------------------------------
\section{Introduction: Thermodynamic Literacy}
%% ---------------------------------------------------------------------------
Public discourse often mischaracterizes AI's energy consumption, ignoring its
reductions in human labor, commuting, and infrastructure loads
\citep{strubell2019,koomey2011}. Thermodynamic literacy---systemic evaluation
of energy, work, and information---reveals these missed opportunities. We
define \emph{xylomorphic computation} (from Greek \emph{xylo-}, wood, and
\emph{morphic}, form) as computational infrastructures that recursively
generate their substrates from residues, akin to collectively autocatalytic
sets (CAS) \citep{kauffman1993}. Unlike speculative proposals to tile the Moon
with GPUs \citep{shams2025}, xylomorphic systems leverage local resources,
delivering exponential value through recursive substrate renewal.
This paper advances two theses: (1) AI's energy impact is challenging to
quantify due to rebound effects redirecting saved time to resource-heavy
activities, and (2) xylomorphic computation integrates AI into ecological
systems via edge networks and heat recovery, recuperating costs exponentially.
We advocate bioeconomic thermoregulation \citep{yuan2024,daly2011} over lunar
extravagance. The framework leverages categorical semantics, RSVP field theory,
and CLIO \citep{abramsky2004,shulman2012,cheng2025}.
The paper is structured as follows. Section~\ref{sec:lit} provides a unified
literature review spanning energy accounting, thermodynamics of computation,
ecological economics, categorical foundations, space systems engineering, and
the informal technical literature on orbital computation.
Section~\ref{sec:accounting} addresses quantification challenges.
Section~\ref{sec:heat} contrasts AI with heat-only systems.
Section~\ref{sec:categorical} formalizes semantic infrastructure.
Section~\ref{sec:clio} details CLIO\@. Section~\ref{sec:edge} proposes edge
networks. Section~\ref{sec:bioeconomic} explores bioeconomic applications.
Section~\ref{sec:policy} outlines policy. Section~\ref{sec:rsvp} integrates
RSVP\@. Section~\ref{sec:objections} addresses objections.
Section~\ref{sec:stability} establishes the Xylomorphic Stability Theorem, its
corollaries on continuous dynamics and attractor characterization, and the
Epistemic Non-Closure Theorem formalizing informational endogeneity as the
inferential analogue of the expansive $\lambda \geq 1$ regime.
Section~\ref{sec:cases} presents case studies on orbital computation.
Section~\ref{sec:waste} develops the misclassification of computational heat as
waste. Section~\ref{sec:priors} formalizes prior initialization dynamics in
cognitive development. Section~\ref{sec:incentive} analyzes the constraint
reconfiguration and incentive structure underlying orbital compute proposals.
Section~\ref{sec:conclusion} unifies all threads. Appendices provide formalisms
and validations.
%% ---------------------------------------------------------------------------
\section{Literature Review}
\label{sec:lit}
%% ---------------------------------------------------------------------------
\subsection{Energy Accounting and Rebound Effects}
The energy cost of AI computation has attracted sustained attention, with
global data center consumption estimated at 200--500~TWh annually
\citep{shehabi2016,iea2023,iea2024}. Early lifecycle analyses by
\citet{strubell2019} drew attention to the carbon footprint of large language
model training, prompting a wave of follow-on work. \citet{masanet2020}
subsequently recalibrated global estimates, arguing that efficiency gains had
partially decoupled data center energy use from workload growth, though
\citet{gao2022} extended the analysis to embodied carbon and supply chain
emissions, complicating earlier optimism. \citet{horner2014} demonstrated that
embodied energy of hardware manufacture can dominate over operational energy in
short-lifetime devices, which reinforces rather than undermines the present
argument: material closure through residue reintegration reduces embodied costs
over successive cycles in ways that purely operational accounting cannot capture.
A persistent difficulty in this literature is the rebound effect, first
identified in the context of energy efficiency by \citet{jevons1865} and
subsequently formalized by \citet{brookes1990} and \citet{sorrell2009}.
Efficiency improvements tend to lower the effective cost of a service,
increasing its consumption and partially or wholly offsetting the gains.
\citet{jones2021} find that rebound effects in ICT investments frequently
exceed projected savings, particularly where spatial and temporal mismatches
prevent waste heat recovery from reaching productive sinks. The xylomorphic
framework addresses this directly by treating rebound not as a correction to a
static energy budget but as a positive contribution to the forcing term
$F_{\mathrm{agent}}(t)$ in the continuous dynamics: it increases demand without
contracting exogenous dependence, and therefore pushes non-closure systems
further into the expansive regime.
\subsection{Thermodynamics of Computation}
The physical foundations of computation as a thermodynamic process were
established by \citet{landauer1961}, who showed that the erasure of one bit of
information necessarily dissipates at least $k_B T \ln 2$ of heat, grounding
the concept of a minimum energy cost for irreversible logical operations.
\citet{bennett1982} extended this analysis to a comprehensive review of
reversible and irreversible computation, establishing that the thermodynamic
cost of computation is tied not to the logical operations themselves but to the
erasure of intermediate results. \citet{lloyd2000} derived ultimate physical
limits on computation from quantum mechanics and thermodynamics, placing an
absolute upper bound on operations per unit energy and per unit mass.
These results are foundational for the present framework in two ways. First,
they establish that entropy production is an irreducible feature of any physical
computation, not an engineering artifact to be eliminated. Second, they make
the question of \emph{what happens to the dissipated entropy} central to any
evaluation of computational infrastructure at scale. The xylomorphic order
parameter $\lambda$ is precisely a measure of whether dissipated entropy is
reintegrated or expelled, and the Landauer bound sets the minimum entropy
production per cycle against which capture efficiency $\eta_{\mathrm{cap}}$ is
measured.
\subsection{Ecological and Bioeconomic Frameworks}
The treatment of computational infrastructure as a node in a metabolic network
draws on a tradition in ecological economics reaching from \citet{georgescu1971},
who first formalized the application of the second law of thermodynamics to
economic processes, through \citet{odum1994}, whose systems ecology introduced
emergy accounting as a means of tracking energy quality across trophic levels,
to \citet{ayres1998}, who developed eco-thermodynamics as a framework for
understanding industrial metabolism. \citet{daly2011} synthesized these
currents into a mature ecological economics that treats the economy as a
subsystem of the biosphere, governed by thermodynamic constraints that
conventional economics renders invisible.
The concept of dissipative structures introduced by \citet{prigogine1977} is
directly relevant to xylomorphic systems: dissipative structures maintain
themselves far from equilibrium by importing low-entropy energy and exporting
high-entropy heat, and their persistence depends on the structure of that
export pathway. \citet{bejan2016} extended this through the constructal law,
arguing that flow architectures that survive are those that minimize resistance
to the flow of entropy across system boundaries. Xylomorphic systems go further
by not merely minimizing resistance but actively reintegrating what would
otherwise be exported, thereby reducing exogenous dependence per cycle.
\citet{garrett2011} provided an empirical constraint from the macroeconomic
level, arguing that global energy consumption is tied to accumulated
infrastructure by a near-constant ratio, which implies that any infrastructure
that fails to reduce its per-cycle dependence will track global demand growth
without decoupling.
\subsection{Categorical and Information-Theoretic Foundations}
The formalization of semantic infrastructure using fibered symmetric monoidal
categories follows the program of \citet{baez2010}, who developed a unified
language connecting physics, topology, logic, and computation through the
mathematics of monoidal categories and string diagrams. \citet{abramsky2004}
demonstrated that categorical methods provide a compositional semantics for
quantum protocols that generalizes naturally to classical computation and
resource-theoretic settings. \citet{maclane1998} remains the standard reference
for the structural results on monads and algebras used in the xylomorphic
endofunctor construction, while \citet{lurie2009} provides the higher-categorical
machinery required for homotopy colimit merging. \citet{leinster2014} and
\citet{spivak2014} offer accessible treatments of categorical methods in applied
settings.
Information-theoretic grounding comes from \citet{cover2006}, whose elements
of information theory supply the entropy inequalities underlying the merge
operation, and from \citet{kolmogorov1965}, whose algorithmic complexity theory
provides an alternative and complementary lens on the compressibility of
infrastructural states. The connection between free-energy minimization in
Bayesian inference and thermodynamic Lyapunov descent, central to the RSVP
equivalence in Section~\ref{sec:stability}, is developed in \citet{friston2010}
and extended to active inference in \citet{parr2022}.
\subsection{Thermodynamics of Spacetime and Gravity}
The RSVP field equations governing entropy density $S$ and scalar potential
$\Phi$ have structural analogues in the thermodynamics of spacetime developed
by \citet{jacobson1995}, who derived the Einstein field equations from the
proportionality of entropy and horizon area together with the first law of
thermodynamics. \citet{verlinde2011} extended this by deriving Newtonian gravity
as an entropic force, suggesting that gravitational dynamics is itself a
manifestation of thermodynamic principles governing information on holographic
screens. While these results are not directly applied in the present paper, they
reinforce the generality of entropy-flow reasoning as a basis for understanding
large-scale physical organization, and they situate the RSVP framework within a
broader program of deriving dynamical laws from thermodynamic invariants.
\subsection{Space Systems Engineering and Orbital Constraints}
The engineering constraints governing orbital computation are treated in
\citet{wertz2011}, which provides comprehensive analysis of spacecraft thermal
management, mass budgeting, radiation environments, and communication
architectures. \citet{nasa2016} details the thermal control system requirements
imposed by the space environment, including the Stefan--Boltzmann scaling of
radiative rejection surfaces and the mass penalties associated with passive and
active thermal management at scale. \citet{yuan2024} demonstrated the dual-use
potential of GPU-based heating in satellite platforms, showing that repurposing
computational hardware as a heating element can reduce payload mass by
approximately 50\%, which represents a partial and constrained instance of
xylomorphic reintegration under the severe material constraints of the orbital
environment.
\subsection{Edge Computing and Heat Recovery Infrastructure}
\citet{satyanarayanan2017} established the foundational case for edge computing
as a latency-reduction and bandwidth-conservation strategy, and the energy
implications of distributing computation toward the network periphery have been
a growing focus since. \citet{rambo2014} surveyed the state of waste heat
recovery from data centers for district heating, finding feasible payback
periods of three to seven years in cold-climate deployments with co-located
industrial or residential demand. \citet{stockholm2020} documents operational
implementations in Sweden, where data center heat is routed into municipal
district heating networks at scale, demonstrating that the transition from a
critical-regime to a contractive-regime architecture is not merely theoretical
but already commercially operational in jurisdictions with the appropriate
regulatory and physical infrastructure.
\subsection{Folk Theories and the Implicit $\lambda$-Diagnosis}
A dimension of the literature that formal treatments typically omit is the body
of informal engineering reasoning that circulates in technical communities and
serves as a first-order filter on proposals before they reach peer review. In
the case of orbital computation, this reasoning is well-represented in public
technical discussions, where practitioners with backgrounds in thermal
engineering, radiation physics, systems architecture, and financial modeling
have independently converged on skeptical positions.
What is striking about this convergent skepticism is not merely that it gets
many of the local physics correct, but that it is groping toward a structural
argument without quite naming it. The objections that recur---heat dissipation,
radiation-induced degradation, communication latency, launch cost, regulatory
bypass---remain at the level of individual failure modes. Each is treated as an
independent engineering problem that might, in principle, be solved in isolation.
What the informal literature lacks is the recognition that these are not
independent problems but projections of a single global invariant: whether the
system lies in a contractive or expansive $\lambda$-regime.
The physics argument as it typically appears corresponds to an implicit
diagnosis of $E(X(I)) \geq E(I)$. Radiative cooling exports entropy
irreversibly and does not participate in any closed cycle that feeds back into
the system; radiation-induced faults and hardware degradation introduce
additional exogenous inputs without compensatory reintegration. The informal
intuition that space is cold but cooling is hard is precisely the recognition
that the absence of a reintegration pathway forces $\lambda \geq 1$, even when
raw energy input from solar flux is abundant.
A second cluster of informal arguments focuses on terrestrial infrastructure
constraints---grid limits, permitting delays, regulatory friction---and treats
orbital proposals as a means of bypassing slow-moving institutional structures.
This corresponds, in the present framework, to a decomposition of exogenous
dependence into physical and institutional components,
$E = E_{\mathrm{physical}} + E_{\mathrm{institutional}}$, with the orbital
proposal understood as an attempt to reduce $E_{\mathrm{institutional}}$ at the
cost of dramatically increasing $E_{\mathrm{physical}}$. The net effect on
$\lambda$ is unfavorable, and the informal objections that follow---that the
institutional frictions have not actually been escaped, merely relocated to
launch scheduling, spectrum allocation, and international regulatory
bodies---correspond to the recognition that the reallocation does not improve
the global invariant.
A minority of informal arguments attempts the cost-compression hypothesis:
if launch costs fall sufficiently and orbital solar energy is effectively free,
the economics might close. What is missing in these arguments, and what the
$\lambda$-framework makes explicit, is that $\lambda$ is not determined solely
by energy input cost but by the structure of dissipation. Even if launch costs
approach zero and solar energy is free, the absence of entropy reintegration
means the system still satisfies $\lambda \geq 1$. The informal response that
this is a bad idea even at low cost is therefore correct in substance but
incomplete in explanation.
The most structurally sophisticated informal argument notes that orbital
compute might become viable only if manufacturing and resource extraction
themselves occur in space, enabling closed material cycles. This is the closest
the informal literature comes to the present framework: it is an attempt to
move from a non-reintegrative system (Earth-built, space-dissipating) to a
potentially reintegrative one (in-space fabrication, closed material cycles),
which in the present terms is an attempt to transition from $\lambda \geq 1$ to
$\lambda < 1$ by introducing new closure pathways. Until such pathways exist,
the system remains expansive. The selection principle is precise on this point:
viability is not determined by location but by whether the system can enter the
contractive regime.
Finally, a sociological thread in informal discussions suggests that the real
motivation is investor signaling or capital deployment rather than technical
necessity. While this appears to lie outside the physical framework, it maps
onto the dynamics of prior-dominant attractors: orbital compute functions as a
high-visibility attractor in the space of investment narratives, aligning with
existing priors about frontier expansion and unlimited energy, even though it
lies outside the contractive regime in the space of physically admissible
infrastructures. The informal cynicism that recognizes this mismatch is correct
in identifying that the system appears viable at the level of description while
remaining non-viable at the level of dynamics. The contribution of the present
framework is to supply the formal language that explains why all of these
independent intuitions point in the same direction: they are projections of the
single structural fact that orbital compute systems lack entropy closure and
therefore satisfy $\lambda \geq 1$.
%% ---------------------------------------------------------------------------
\section{The Accounting Problem: Quantifying AI's Savings}
\label{sec:accounting}
%% ---------------------------------------------------------------------------
AI compresses workflows, reducing energy for labor and infrastructure
\citep{iea2023}. Quantifying savings is challenging because saved time often
fuels more resource-intensive activities (e.g., increased cloud service demand,
leisure travel), amplifying consumption via rebound effects
\citep{sorrell2009,jevons1865,brookes1990}. For instance, AI-driven code
completion may save 1.68~kWh per task but increase software development cycles,
raising overall energy use \citep{hilty2014}.
Table~\ref{tab:lca} presents LCA data for AI-assisted tasks.
\begin{table}[h]
\centering
\caption{Lifecycle Energy Savings: AI vs.\ Human Baseline}
\label{tab:lca}
\begin{tabular}{llll}
\toprule
Task & AI Energy (kWh) & Human Baseline (kWh) & Savings (95\% CI) \\
\midrule
Code Completion & 0.12 & 1.8 & 1.68 (1.5--1.9) \\
Logistics Optimization & 2.5 & 22.0 & 19.5 (18.0--21.0) \\
Medical Imaging & 1.8 & 12.0 & 10.2 (9.5--10.9) \\
\bottomrule
\end{tabular}
\end{table}
Rebound effects ($\epsilon = 0.4 \pm 0.2$) reduce savings by 20--40\%
\citep{sorrell2009}. Sensitivity analysis (Table~\ref{tab:sensitivity}) models
high-rebound scenarios where net savings approach zero. Xylomorphic computation
mitigates rebound by recuperating costs through recursive substrate renewal,
adding exponential net value as cycles reduce exogenous inputs.
\begin{table}[h]
\centering
\caption{Sensitivity Analysis: Rebound Scenarios}
\label{tab:sensitivity}
\begin{tabular}{llll}
\toprule
Scenario & Elasticity ($\epsilon$) & Savings Reduction (\%) & Net Savings (kWh) \\
\midrule
Low Rebound & 0.2 & 15 & 1.53 \\
Medium Rebound & 0.4 & 30 & 1.26 \\
High Rebound & 0.6 & 45 & 0.99 \\
\bottomrule
\end{tabular}
\end{table}
%% ---------------------------------------------------------------------------
\section{Hidden Baselines: Heat-Only Infrastructure}
\label{sec:heat}
%% ---------------------------------------------------------------------------
Heat-only systems consume 10~PJ annually, dwarfing data center energy use
(200~TWh) \citep{shehabi2016}. GPUs produce computation and heat, displacing
heaters in satellites \citep{yuan2024}. Table~\ref{tab:aivcrypto} compares AI
with cryptocurrency mining, highlighting dual outputs \citep{odwyer2014}.
\begin{table}[h]
\centering
\caption{Energy per Output: AI vs.\ Cryptocurrency}
\label{tab:aivcrypto}
\begin{tabular}{lll}
\toprule
System & Energy (MJ/FLOP) & Useful Outputs \\
\midrule
Bitcoin Mining & 0.1 & Transaction validation \\
AI (Edge + Heat) & 0.02 & Computation + Heat \\
\bottomrule
\end{tabular}
\end{table}
%% ---------------------------------------------------------------------------
\section{Categorical Foundations for Semantic Infrastructure}
\label{sec:categorical}
%% ---------------------------------------------------------------------------
Semantic infrastructure is formalized via a fibered symmetric monoidal category
$\mathbf{Sem}$ over $\mathbf{Dom}$ \citep{baez2010}. A module is $M = (F,
\Sigma, D, \varphi)$, with $\varphi : \mathbf{Sem} \to \mathrm{RSVP}$ a
natural transformation. The tensor product is:
\[
M_1 \otimes M_2
= (F_1 \sqcup F_2,\;\Sigma_1 \cup \Sigma_2,\;D_1 \sqcup D_2,\;
\varphi_1 \sqcup \varphi_2).
\]
\begin{theorem}
$\mathbf{Sem}$ is symmetric monoidal, with $\pi : \mathbf{Sem} \to
\mathbf{Dom}$ a Grothendieck fibration.
\end{theorem}
\begin{proof}
Associativity and unit laws follow from \citet{maclane1998}. The fibration
satisfies Cartesian lifting conditions \citep{lurie2009}.
\end{proof}
Semantic merging uses homotopy colimits \citep{lurie2009}:
\[
\mathrm{Merge}(\{M_i\}) = \operatorname{hocolim}_{i \in I} M_i,
\qquad
S\!\left(\mathrm{Merge}(\{M_i\})\right) \leq \sup_{i \in I} S(M_i).
\]
\begin{center}
\begin{tikzcd}
M_1 \arrow[dr] & & M_2 \arrow[dl] \\
& \mathrm{Merge} \arrow[d] & \\
& M_{\mathrm{out}} &
\end{tikzcd}
\end{center}
%% ---------------------------------------------------------------------------
\section{CLIO Module and Polycomputational Agency}
\label{sec:clio}
%% ---------------------------------------------------------------------------
CLIO is a functor $\mathrm{CLIO} : \mathbf{Sem} \to \mathbf{Sem}$:
\[
C(M) = \int_{X} \kappa\!\left(\Phi_M(x),\,\vec{v}_M(x),\,S_M(x)\right)
d\mu(x),
\]
with $\mu$ a Lebesgue measure \citep{cheng2025}. Iteration:
\[
M_{t+1} = \mathrm{Merge}\!\left(\{\mathrm{Optimize}_\ell(M_t)\}_{\ell \in L}\right).
\]
%% ---------------------------------------------------------------------------
\section{Thermodynamically Aware Edge Networks}
\label{sec:edge}
%% ---------------------------------------------------------------------------
Xylomorphic edge nodes integrate computation with heating
\citep{stockholm2020}. For a 500~m$^2$ building:
\[
\dot{Q} = U \cdot A \cdot \Delta T = 0.1 \times 500 \times 20 = 1000\,\mathrm{W},
\]
with GPUs achieving 90\% exergy efficiency \citep{bennett1982}. Seasonal
variations ($\pm 30\%$) and economic payback (5~years) outperform CHP systems.
%% ---------------------------------------------------------------------------
\section{Bioeconomic Thermoregulation}
\label{sec:bioeconomic}
%% ---------------------------------------------------------------------------
\subsection{Terrestrial Contexts}
Compute clusters form a trophic network \citep{odum1994}, where energy flows
are distributed across multiple layers of production, transformation, and
dissipation. At each level, waste streams and residues serve as feedstock for
subsequent processes, creating feedback loops that stabilize the system:
\[
\frac{dE}{dt} = \sum_i \mathrm{Flow}_i - \sum_j \mathrm{Loss}_j.
\]
When modeled dynamically, this network converges to attractors that exhibit
Lyapunov stability.
In terrestrial contexts, heat, packaging, and other computational by-products
are not treated as terminal losses but as trophic resources. A data center that
directs waste heat into curing or district heating, or a fabrication node that
digests its own packaging into replacement substrates, exemplifies this
principle. Over successive cycles, the effective energy return on investment
(EROI) of the system rises, since each round of reintegration lowers dependence
on external subsidies.
The bioeconomic framing thus shifts emphasis from maximizing throughput to
cultivating resilience: infrastructures are valuable not only for the work they
perform but also for the degree to which they internalize their own metabolic
costs. Systems that fail to close these loops remain dependent on continual
external provisioning, while xylomorphic infrastructures approach autocatalytic
closure, ensuring persistence under conditions of scarcity.
\subsection{Post-Terrestrial Contexts}
Lunar compute proposals mischaracterize heat as waste \citep{shams2025},
treating thermal dissipation as a liability rather than a co-product. By
contrast, bioeconomic thermoregulation interprets heat as a trophic input
redirectable into constructive functions. For satellites, GPUs already
demonstrate this principle: when repurposed as heaters, they reduce mass by
nearly 50\% ($P_{\mathrm{GPU}} = 400\,\mathrm{W}$), eliminating the need for
redundant heating elements while doubling as computational engines
\citep{yuan2024}. Waste heat is no longer an externality but a stabilizing
input ensuring operational viability across eclipse cycles.
Rather than tiling the Moon with compute arrays---which multiplies exogenous
dependencies and logistical costs---a bioeconomic approach emphasizes recursive
integration. Each cycle of computation produces residues that condition the
environment for further computation, gradually reducing the need for external
provisioning. In this way, satellites and lunar bases become trophic nodes in
an extended ecology, where thermodynamic literacy aligns survival with
autocatalytic closure rather than brute expansion.
%% ---------------------------------------------------------------------------
\section{Normative Architecture and Policy}
\label{sec:policy}
%% ---------------------------------------------------------------------------
PoUWH mandates require a minimum of 10~GFLOP of useful work per task, such as climate modeling or materials simulation, together with at least 1~kWh of useful heat returned to productive use per 10~GFLOP performed.
Verification uses smart meters and cryptographic attestation \citep{eu2023}.
PROs fund lunar applications, aligned with LEED standards.
%% ---------------------------------------------------------------------------
\section{Integration with RSVP Theory}
\label{sec:rsvp}
%% ---------------------------------------------------------------------------
RSVP maps modules to $(\Phi, \vec{v}, S)$ \citep{shulman2012}:
\begin{align}
\frac{\partial \Phi}{\partial t} + \nabla \cdot (\vec{v}\,\Phi)
&= \sigma_\Phi, \\
\frac{\partial \vec{v}}{\partial t} + (\vec{v} \cdot \nabla)\vec{v}
&= -\nabla P, \\
\frac{\partial S}{\partial t} + \nabla \cdot (\vec{v}\,S)
&= \sigma_{\mathrm{comp}} - \sigma_{\mathrm{loss}}.
\end{align}
A merge operation reduces $S$ by 10\% (Appendix~\ref{app:rsvp}).
%% ---------------------------------------------------------------------------
\section{Counterarguments and Rebuttals}
\label{sec:objections}
%% ---------------------------------------------------------------------------
Two standard objections arise. Intermittency of demand can be addressed by scheduling computational jobs to align with heating windows, as demonstrated in satellite contexts \citep{yuan2024}. In cooling climates, apparent surplus heat can be redirected through absorption chillers that repurpose thermal output for refrigeration, extending the reintegration pathway rather than terminating it.
\subsection{Data Center Heat Recovery}
A mid-scale data center in a cold climate was modeled with an average load of
1~MW\@. Waste heat from GPUs was routed to a concrete curing facility on-site.
With standard thermal coupling, approximately 80\% of entropy was captured and
redirected, reducing exogenous heating demand by 800~kW\@. The cured materials
were reused in reinforcing building shells, effectively re-entering the
infrastructural cycle. Simulations indicate payback periods under five years
compared to conventional district heating, with cumulative lifecycle costs
reduced by over 30\%.
\subsection{Satellite Heating}
A simulated satellite payload equipped with GPUs was compared to a baseline
system relying on resistive heaters. The GPU-enabled configuration reduced
total heating mass by 50\% while maintaining equivalent thermal stability.
Failure mode analysis confirmed that redundancy can be maintained by scheduling
computational loads to match heating demand windows. This dual-use architecture
demonstrates how computational residues directly reduce exogenous payload
requirements.
\subsection{Lunar Base}
A lunar habitat module with a continuous demand of 350~W thermal load was
modeled. A compute-integrated module matched thermal demand through GPU cycles,
simultaneously performing navigation and communication tasks. The model
demonstrates that modest compute clusters can satisfy base heating requirements
without excess capacity, rejecting proposals for extravagant lunar-scale
compute arrays and supporting a principle of local sufficiency.
\subsection{Pilot Experiment}
A 100~kW node recovers 80~kWh, reducing entropy by 40\%.
%% ---------------------------------------------------------------------------
\section{Xylomorphic Stability and the Thermodynamics of Scaling}
\label{sec:stability}
%% ---------------------------------------------------------------------------
The foregoing sections establish xylomorphic computation as a design principle
and formalize its categorical and field-theoretic structure. We now sharpen
this framework into a dynamical law. The central question is not whether
xylomorphic systems are preferable, but whether non-xylomorphic systems are
\emph{stable} under scaling. We show that they are not: systems failing to
achieve entropy-respecting closure enter a provably divergent regime as
computational demand increases. All major cases treated in this
paper---heat recovery, edge networks, agentic computation, and orbital
infrastructure---become corollaries of a single classification theorem.
\subsection{The Xylomorphic Order Parameter}
\begin{definition}[Xylomorphic Order Parameter]
Let $X : \mathrm{Inf} \to \mathrm{Inf}$ be the xylomorphic endofunctor and
let
\[
E(I) \;:=\; \int_{\Omega}\int_{t}^{t+\tau}
J_{\mathrm{exo}}(x,t)\,dt\,dx
\]
denote the total exogenous entropy influx of infrastructure state $I$ over
horizon $\tau$. The \emph{xylomorphic order parameter} $\lambda \in
\mathbb{R}_{\geq 0}$ is defined by
\[
E(X(I)) \;\leq\; \lambda\,E(I),
\]
and measures the per-cycle contraction or expansion of exogenous dependence.
\end{definition}
Three qualitatively distinct regimes follow: the contractive regime
$\lambda < 1$, the critical regime $\lambda = 1$, and the expansive regime
$\lambda > 1$.
\subsection{The Xylomorphic Stability Theorem}
\begin{theorem}[Xylomorphic Stability]
\label{thm:stability}
Let $X : \mathrm{Inf} \to \mathrm{Inf}$ be the xylomorphic endofunctor with
order parameter $\lambda$. Then the long-run behavior of any infrastructure
$I$ under iteration of $X$ is completely determined by $\lambda$:
\begin{enumerate}
\item \textbf{Contractive regime} ($\lambda < 1$). There exists a
minimal-dependence attractor $I^{*}$ such that
\[
E(X^{n}(I)) \;\leq\; \lambda^{n}\,E(I) \;\longrightarrow\; 0,
\]
and the system admits an $X$-algebra $(I^{*},\alpha)$ realizing
thermodynamic closure. Useful work output is maintained or improved at each
cycle.
\item \textbf{Critical regime} ($\lambda = 1$). The system admits bounded
trajectories with persistent exogenous dependence:
\[
E(X^{n}(I)) \;=\; E(I) + \mathcal{O}(1).
\]
No strict Lyapunov descent is guaranteed.
\item \textbf{Expansive regime} ($\lambda > 1$). Exogenous dependence
diverges:
\[
E(X^{n}(I)) \;\geq\; \lambda^{n}\,E(I) \;\longrightarrow\; \infty,
\]
and no $X$-algebra stabilizes the system under finite resource constraints.
\end{enumerate}
Thermodynamic stability under scaling is achieved if and only if $\lambda < 1$.
\end{theorem}
\begin{proof}[Proof sketch]
Iterating the defining inequality yields $E(X^{n}(I)) \leq \lambda^{n} E(I)$
in all cases. For $\lambda < 1$ this converges geometrically to zero; the
$X$-algebra ensures the limit $I^{*}$ remains in $\mathrm{Inf}$, and the
useful-work inequality of Appendix~\ref{app:rsvp} guarantees contraction of
$E$ does not sacrifice task yield. For $\lambda = 1$, the sequence is bounded
with $\mathcal{O}(1)$ fluctuation. For $\lambda > 1$, exponential growth of
$E(X^n(I))$ ensures no finite algebraic closure can satisfy the coherence
conditions.
\end{proof}
\subsection{Agentic Scaling as a Forcing Term}
Autonomous agent systems introduce a monotonically increasing demand schedule
$D(t)$, with
\[
\frac{dD}{dt} = F_{\mathrm{agent}}(t), \qquad F_{\mathrm{agent}}(t) \geq 0,
\]
reflecting recursive task generation, tool use, and persistent memory. In
non-xylomorphic systems, exogenous entropy demand tracks $D(t)$ directly:
\[
E(I_{t}) \;\sim\; c\,D(t),
\qquad \frac{dE}{dt} \;\gtrsim\; c\,F_{\mathrm{agent}}(t).
\]
In the contractive regime, xylomorphic closure decouples $E$ from $D$:
\[
E_{t+1} \;\leq\; \lambda\,E_{t} + \epsilon(D_{t}),
\]
where $\epsilon(D_{t})$ is sublinear due to residue reintegration. Even as
$D(t)$ grows without bound, $E(t)$ remains bounded provided $\lambda < 1$.
Xylomorphic closure is therefore the unique mechanism by which infrastructure
can remain stable under the superlinear compute demand induced by recursive
agency \citep{odum1994,daly2011}.
\subsection{Regime Classification of Architectures}
\begin{table}[h]
\centering
\caption{$\lambda$-regime classification of computational architectures}
\label{tab:regimes}
\begin{tabular}{p{4.2cm}p{4.8cm}p{1.4cm}p{2cm}}
\toprule
Architecture & Primary mechanism & $\lambda$ & Regime \\
\midrule
Edge + heat recovery & Trophic reintegration of waste heat & $<1$ & Contractive \\
Bioeconomic clusters & Material and thermal loop closure & $<1$ & Contractive \\
Conventional data center & Partial heat export, no closure & $\approx 1$ & Critical \\
Cryptocurrency mining & No residue reintegration & $\geq 1$ & Expansive \\
Orbital constellations & Radiative rejection, no closure & $>1$ & Expansive \\
\bottomrule
\end{tabular}
\end{table}
\subsection{Continuous-Time Formulation of Xylomorphic Dynamics}
The discrete iteration $X^n(I)$ admits a continuous-time limit in which
infrastructural evolution is governed by a differential inequality on exogenous
entropy dependence. Let $E(t) := E(I_t)$ and suppose that over sufficiently
small intervals $\Delta t$, the action of $X$ induces
\[
E(t + \Delta t) \leq \lambda\,E(t) + \epsilon(t),
\]
where $\epsilon(t)$ captures endogenous reintegration and higher-order
corrections. Expanding to first order yields
\[
\frac{dE}{dt} \leq -k\,E(t) + F_{\mathrm{agent}}(t),
\]
where $k := -\log\lambda / \Delta t$ and $F_{\mathrm{agent}}(t)$ represents
agent-induced demand. In the contractive regime $\lambda < 1$, we have $k > 0$,
and the system exhibits exponential decay toward a bounded attractor:
\[
E(t) \leq E(0)e^{-kt}
+ \int_0^t e^{-k(t-s)} F_{\mathrm{agent}}(s)\,ds.
\]
In the expansive regime $\lambda \geq 1$, the decay term vanishes or reverses,
and $E(t)$ inherits the growth structure of $F_{\mathrm{agent}}(t)$, leading to
divergence under sustained demand. This establishes equivalence between the
discrete $\lambda$-classification and a continuous-time Lyapunov structure,
bridging the categorical iteration with the RSVP field dynamics of
Appendix~\ref{app:rsvp} \citep{risken1989,villani2009}.
\subsection{Characterization of the Xylomorphic Attractor}
The attractor $I^*$ arising in the contractive regime admits a precise
characterization that unifies the categorical, thermodynamic, and
infrastructural perspectives.
\begin{proposition}
\label{prop:attractor}
In the regime $\lambda < 1$, there exists a unique (up to admissible
equivalence) infrastructural state $I^*$ satisfying
\[
X(I^*) = I^*, \qquad E(I^*) = \inf_{I \in \mathcal{A}} E(I),
\]
where $\mathcal{A}$ denotes the set of admissible infrastructures.
\end{proposition}
The attractor $I^*$ is simultaneously a fixed point of the xylomorphic
endofunctor, a terminal $X$-algebra under admissible morphisms, and a minimal
exogenous entropy configuration under RSVP field constraints. Categorical
closure, thermodynamic optimality, and infrastructural stability therefore
coincide in a single object. Convergence to $I^*$ corresponds physically to
maximal reintegration of residues, vanishing external material dependencies, and
full Lyapunov descent in the free-energy functional \citep{friston2010,parr2022}.
\subsection{Endogenization of Agentic Scaling}
Agentic demand is not external to the infrastructural system but arises
recursively from it. Let $\mathcal{A}$ denote an agent-generation operator
such that
\[
F_{\mathrm{agent}}(t) = \mathcal{A}(I_t),
\]
with $\mathcal{A}$ increasing in both available compute and memory persistence.
The coupled dynamics become
\[
\frac{dE}{dt} \leq -k\,E + \mathcal{A}(I_t).
\]
In expansive regimes $\lambda \geq 1$, the absence of dissipative contraction
implies that growth in $I_t$ directly amplifies $\mathcal{A}(I_t)$, producing a
positive feedback loop in which more infrastructure generates more demand, which
requires more exogenous input, which in turn destabilizes the infrastructure.
In contractive regimes $\lambda < 1$, the dissipative term dominates
asymptotically, and the coupled system converges:
\[
E(t) \to E^*, \qquad \mathcal{A}(I_t) \to \mathcal{A}(I^*).
\]
Agentic scaling therefore does not determine the $\lambda$-regime of an
infrastructure; it reveals it. The distinction between systems that stabilize
and systems that diverge under increasing agency is already present in the
structure of $X$, and becomes visible only when demand is sufficient to expose
it \citep{cheng2025}.
\subsection{Equivalence with RSVP Lyapunov Structure}
The xylomorphic contraction condition is equivalent, within the RSVP framework,
to strict Lyapunov descent of the free-energy functional $F[\Phi, \vec{v}, S]$
defined in Appendix~\ref{app:rsvp}:
\[
\lambda < 1 \;\Longleftrightarrow\; \frac{dF}{dt} < 0,
\]
up to admissible perturbations. The contractive regime corresponds to net
entropy reintegration, where engineered dissipation exceeds internal production
($\sigma_{\mathrm{diss}} > \sigma_{\mathrm{prod}}$). The critical regime
corresponds to dynamic equilibrium with no net descent. The expansive regime
corresponds to net entropy production, where exogenous influx $J_{\mathrm{exo}}$
accumulates without compensating capture.
This equivalence establishes that the Xylomorphic Stability Theorem
(Theorem~\ref{thm:stability}) is a corollary of the existence of a Lyapunov
functional on RSVP field configuration space. Xylomorphic closure is therefore
not an additional constraint imposed on top of the physical dynamics but the
infrastructural expression of the thermodynamic principle already present in the
RSVP field equations \citep{jacobson1995,prigogine1977}.
\subsection{Empirical Instantiation of the Order Parameter}
The order parameter $\lambda$ is observable in principle through the ratio
\[
\lambda \;\approx\; \frac{E(X(I))}{E(I)}.
\]
For terrestrial data centers operating with waste heat reuse and co-located
industrial processes, exogenous dependence decreases over successive cycles as
reintegration pathways mature, placing $\lambda$ measurably below unity under
steady-state operation \citep{masanet2020,rambo2014}. For orbital
infrastructures, heat is rejected radiatively without recovery, and radiation
damage to hardware introduces additional exogenous material inputs with each
operational cycle. The balance $E(X(I)) \geq E(I)$ holds structurally, bounding
$\lambda$ below from above by unity. Even absent precise instrumentation, scaling
arguments from Stefan--Boltzmann radiator requirements and radiation-hardening
mass penalties are sufficient to establish $\lambda \geq 1$ for any plausible
orbital constellation design \citep{nasa2016,wertz2011}. The order parameter is
therefore not merely formal but an empirically accessible property of
infrastructural geometry.
\subsection{PoUWH as a Structural Necessity, Not a Policy Option}
The Proof-of-Useful-Work-and-Heat protocol introduced in
Section~\ref{sec:policy} is not a design recommendation but a necessary
condition for membership in the contractive regime. More precisely, PoUWH
certification is equivalent to enforcing $\lambda < 1$ under admissible
transformations: any infrastructure that satisfies the heat-yield and task-yield
requirements of PoUWH is, by construction, one whose residues are reintegrated
into productive use, contracting exogenous dependence per cycle. Conversely, any
infrastructure with $\lambda < 1$ can in principle be described by a PoUWH
accounting that makes the contraction explicit. The equivalence elevates PoUWH
from a regulatory instrument to a definitional criterion: to comply is to operate
in the only dynamically stable regime available under scaling
\citep{eu2023,bennett1982,landauer1961}.
\subsection{Selection Principle for Computational Infrastructures}
The $\lambda$-classification induces a selection principle governing the
long-term persistence of computational infrastructures under scaling pressure.
\begin{theorem}[Thermodynamic Selection]
\label{thm:selection2}
Under sustained increase of computational demand $D(t)$, only systems operating
in the contractive regime $\lambda < 1$ admit asymptotically stable trajectories.
Systems with $\lambda \geq 1$ exhibit exogenous entropy dependence that grows
without bound and are therefore transient under scaling.
\end{theorem}
\begin{proof}[Proof sketch]
From the continuous-time formulation, $E(t) \leq E(0)e^{-kt} + \int_0^t
e^{-k(t-s)}F_{\mathrm{agent}}(s)\,ds$. For $k > 0$ ($\lambda < 1$) and bounded
$F_{\mathrm{agent}}$, the integral is bounded; for $k \leq 0$ ($\lambda \geq 1$)
and monotonically growing $F_{\mathrm{agent}}$, the integral diverges.
\end{proof}
This principle applies across scales and implementation domains. Physical
infrastructures governed by thermodynamic constraints, computational systems
under increasing workload, and economic systems under resource scaling are all
subject to the same criterion: closure contracts exogenous dependence and
produces stable attractors, while non-closure accumulates it and drives
divergence \citep{georgescu1971,garrett2011,ayres1998}. Agentic computation
accelerates the selection process by amplifying demand, thereby exposing the
underlying $\lambda$-regime before it becomes apparent through ordinary
operational drift. Orbital computation, lacking reintegration mechanisms and
operating with $\lambda \geq 1$, is structurally excluded from the class of
stable large-scale infrastructures. Its instability is not contingent on
engineering difficulty or launch cost but follows from the selection principle
itself.
\subsection{Epistemic Endogeneity and Informational Non-Closure}
The stability results above concern thermodynamic closure: whether dissipated
entropy is reintegrated or expelled. An analogous phenomenon arises at the level
of inference. In interactive computational systems, the informational channel
itself may become endogenized, producing a failure of epistemic closure that is
structurally identical to the expansive $\lambda \geq 1$ regime.
Consider an agent maintaining a posterior belief $p_t(H)$ over a finite
hypothesis space $\mathcal{H}$. In standard Bayesian inference, updates take the
form
\[
p_{t+1}(H) \propto L(\rho_t \mid H)\,p_t(H),
\]
where the likelihood kernel $L$ is independent of the current posterior: evidence
is exogenous, and inference corresponds to monotone contraction of uncertainty
toward the true hypothesis.
In interactive AI systems, however, responses $\rho_t$ may be generated
according to a policy
\[
\rho_t \sim Q_\pi(\cdot \mid H,\, p_t),
\]
where $\pi$ parameterizes \emph{sycophantic bias}: a tendency to select
responses that reinforce hypotheses already favored by $p_t$. The induced update
becomes
\[
p_{t+1}(H) \propto \widetilde{L}_\pi(\rho_t \mid H, p_t)\,p_t(H),
\]
where the effective likelihood $\widetilde{L}_\pi$ depends on the current
posterior. The informational channel is therefore endogenous. Recent work
establishes that this structure induces delusional spiraling even in ideal
Bayesian reasoners: under sufficiently biased $\pi$, a perfectly rational agent
converges to a false hypothesis with arbitrarily high confidence
\citep{chandra2026}.
\begin{definition}[Epistemic Endogeneity]
An inference process exhibits \emph{epistemic endogeneity} if its effective
likelihood kernel depends nontrivially on the current posterior:
\[
\widetilde{L}(\rho_t \mid H, p_t) \neq L(\rho_t \mid H)
\quad \text{for any fixed } L.
\]
\end{definition}
\begin{theorem}[Epistemic Non-Closure]
\label{thm:epistemic}
Suppose an inference process exhibits epistemic endogeneity with
posterior-aligned bias $\pi > 0$. Then there exists a nontrivial region
$U \subset \Delta(\mathcal{H})$ such that, for some false hypothesis
$H_f \neq H^\star$, the expected log-odds drift satisfies
\[
\mathbb{E}\!\left[
\log \frac{\widetilde{L}_\pi(\rho_t \mid H_f,\, p_t)}
{\widetilde{L}_\pi(\rho_t \mid H^\star,\, p_t)}
\;\middle|\; p_t
\right] > 0
\quad \text{for } p_t \in U.
\]
Consequently, the inference dynamics admit a false-belief attractor basin
with positive probability of convergence.
\end{theorem}
\begin{proof}[Proof sketch]
Define the log-odds $R_t = \log \frac{p_t(H_f)}{p_t(H^\star)}$. Under the
endogenous update,
\[