Stratified Relational Workspace Theory

Download original file

This is the original TeX manuscript source. Use the paper’s PDF for typeset equations.

Read the typeset paper

1 % Front matter and introduction. Compile the repository root document instead.
2 % !TeX root = Stratified_Relational_Workspace_Theory.tex
3 \pagenumbering{gobble}
4 
5 \begin{titlepage}
6 \thispagestyle{empty}
7 \centering
8 \vspace*{1.2in}
9 {\Huge\bfseries\color{PartColor} Stratified Relational\\[0.18em] Workspace Theory\par}
10 \vspace{0.55in}
11 {\Large A Core--Transfer Framework for Distributed Cognition,\\ Collaboration, and Epistemic Systems\par}
12 \vspace{0.8in}
13 {\large Komorebi P.\par}
14 \vspace{0.65in}
15 {\large\bfseries Foundations, Dynamics, Agency, and Empirical Certification\par}
16 \vfill
17 {\color{RuleColor}\rule{0.70\textwidth}{0.8pt}\par}
18 \vspace{0.25in}
19 {\small August 2026\par}
20 \end{titlepage}
21 \clearpage
22 \pagenumbering{roman}
23 
24 \begin{abstract}
25 Collaborative reasoning increasingly spans people, computational models, documents, software,
26 instruments, and persistent records. Yet the existence of many interacting components does not,
27 by itself, establish a shared workspace, and successful group performance does not show that the
28 operative state is genuinely distributed rather than copied, centralized, or merely correlated.
29 Stratified Relational Workspace Theory develops a formal answer to this problem. A realization
30 is represented by heterogeneous local state spaces connected by transfer maps. Its exact shared
31 core is the space of global compatible sections, equivalently the fixed subspace of relational
32 holonomy when the transfers are invertible. A transfer Laplacian identifies approximate shared
33 directions and provides spectral stability, gauge covariance, and pathwise leakage bounds.
34 
35 The structural theory is joined to a dynamic and causal account. Canonical predictive states
36 determine whether a proposed workspace retains the conditional law relevant to a declared
37 future task. Feasible interventions, coalition risk, matched relational nulls, and reconstruction
38 tests distinguish causally operative distributed organization from archives, replicated memory,
39 and centralized controllers. Further results address persistence under changing realizations,
40 stable reconstruction after component loss, identity under replacement, local-to-global gluing,
41 hierarchical composition, learned transfer structure, boundary recovery, goal-conditioned action,
42 and certificate-directed repair. Because distributedness can be destroyed by merging components
43 or manufactured by splitting a centralized state into complementary shares, the theory also treats
44 decomposition as part of the claim. Faithful coarse-graining preserves exact structural cores, while
45 family-robust certification requires every gate to retain a positive margin across a declared set of
46 admissible decompositions. The same discipline is applied to the complete scientific specification.
47 Targets, losses, horizons, interventions, nulls, and thresholds are made explicit, and the theory develops
48 specification envelopes, transport radii, experimental-equivalence classes, partial-identification bounds,
49 and uniform finite-sample certification. A certificate is therefore distinguished from a favorable
50 analysis: it may be pointwise valid, specification-robust, sign-identified, or statistically unresolved.
51 The resulting rule is noncompensatory. Structural coherence, predictive alignment, causal efficacy,
52 relational irreducibility, non-collapse, reconstructive continuity, provenance, and self-maintenance are
53 separate claims, and failure of one cannot be hidden by success on another.
54 
55 Exact finite examples show how the same framework applies to human collaboration, human--AI
56 research loops, and multi-agent control. A review of adjacent work locates the individual
57 mathematical ingredients in established traditions including distributed cognition, local systems
58 and sheaves, spectral perturbation, sufficient statistics, structural causal models, frame reconstruction,
59 system identification, and feedback control. The candidate contribution is therefore
60 not a new name for any one of these constructions. It is their integration into a single falsifiable
61 account of persistent relational workspaces and the matched near misses that such an account
62 must reject. The framework does not infer phenomenal consciousness, personhood, or moral
63 status from coherence, agency, or self-maintenance.
64 \end{abstract}
65 
66 \tableofcontents
67 \clearpage
68 \pagenumbering{arabic}
69 
70 \section*{Introduction}
71 \addcontentsline{toc}{section}{Introduction}
72 
73 Scientific and technical work rarely resides in a single mind or device. A proof may be divided
74 among a conjecture, a critic's counterexample, a notebook, a computer algebra system, and a
75 verification script. A human--AI research process may distribute objectives, formal manipulations,
76 execution, and persistent memory across different components. Multi-agent systems divide sensing,
77 planning, communication, and repair in still other ways. In each case, it is natural to speak of a
78 shared workspace. The phrase is useful, but without a formal criterion it can conceal several
79 fundamentally different organizations.
80 
81 A common archive is not necessarily a workspace. Nor is a group whose apparent coordination is
82 produced by one complete controller surrounded by decorative channels. Two systems can contain
83 the same local information while differing in how that information is related, transported, and used.
84 Likewise, a system can perform well without preserving its operative organization through interruption,
85 replacement, or component loss. The central problem is therefore not simply whether information is
86 shared. It is whether there is a persistent relational state whose structure and causal capacities cannot
87 be reduced to any single declared component.
88 
89 The philosophical background is supplied by extended and distributed cognition, transactive memory,
90 collective intelligence, and theories of group agency \cite{Hutchins1995,ClarkChalmers1998,Wegner1987,ListPettit2011}.
91 Those traditions motivate the possibility that cognitive organization may span people and artifacts.
92 The theory asks a narrower mathematical question: what would have to be measured and proved
93 before such an organization should be called a relational workspace? The answer must distinguish
94 physical realization from abstract identity, local representation from globally compatible content,
95 and statistical association from causal operation.
96 
97 The static theory begins with an interaction graph whose vertices carry heterogeneous local state spaces
98 and whose edges carry transfer maps. Each local state is decomposed into a transferable core and a
99 substrate-specific residual. Exact shared content is represented by global compatible sections. When
100 core transfers are invertible, the same object appears as the fixed subspace of the holonomy representation.
101 Approximate shared content is extracted as a low-energy spectral subspace of a transfer Laplacian.
102 This formulation supplies explicit obstruction, stability, gauge, and leakage calculations rather than
103 relying on an informal appeal to shared context.
104 
105 Structure alone is insufficient. A proposed workspace must also retain the predictive state relevant to
106 a declared future, respond under feasible intervention, outperform matched relational nulls, and resist
107 reduction to an individual node or a local-storage surrogate. Persistence introduces further requirements:
108 the workspace must remain trackable as its realization changes, preserve nontrivial content, support
109 reconstruction after loss, and maintain an identity profile under admissible replacement. Local workspaces
110 must glue without hidden incompatibility, and hierarchical organization must not conceal centralization.
111 Formation claims are evaluated separately from operational claims so that a scaffolded but functional
112 workspace is not dismissed, while an endogenously learned but causally irrelevant pattern is not promoted.
113 A separate decomposition analysis asks which claims survive faithful coarse-graining and plausible
114 alternative interface descriptions. This is necessary because arbitrary merging can conceal relationality,
115 while arbitrary splitting can manufacture it. The analysis is then extended from decomposition to the full
116 scientific specification. A workspace claim depends on its declared target, loss, horizon, intervention
117 family, null panel, and thresholds. The theory therefore distinguishes pointwise validity from robustness
118 across defensible specifications and distinguishes both from identifiability under the experiments that are
119 actually available. This prevents a favorable target or threshold from being mistaken for a general result
120 and prevents an observationally indistinguishable hidden controller from being ruled out by assertion.
121 
122 The theory extends to domain-relative agency only after these lower-level conditions are in place. Goals
123 must persist without repeated external rewriting, counterfactual values must guide feasible actions, damage
124 must be detected through the workspace, and repairs must improve a matched no-repair trajectory while
125 preserving identity. These criteria define an engineering notion of workspace agency. They neither assert
126 nor deny phenomenal consciousness. Consciousness, personhood, and moral status remain separate questions
127 that require independent evidence.
128 
129 Most of the mathematical tools used here are established. Inverse limits, holonomy, connection Laplacians,
130 perturbation theory, sufficient statistics, causal models, frame reconstruction, system identification, and
131 feedback control all have substantial prior literatures. The contribution sought here is integrative: a
132 noncompensatory framework that places these tools in one common architecture and specifies the near-miss
133 systems that each layer must reject. Exact examples provide calibration, while the motivating human--AI
134 collaboration is treated conservatively as a candidate case rather than as a certified instance.
135 
136 The paper proceeds from exact structure to approximate structure, dynamic operation, relational
137 irreducibility, persistence, reconstruction, identity, gluing, formation, boundary discovery, agency,
138 decomposition robustness, specification robustness, identifiability, and empirical certification. It then
139 gives worked cases, situates the framework relative to adjacent work, and closes with its philosophical
140 consequences, limitations, and open problems.
141 
142 \section*{Notation and Conventions}
143 \addcontentsline{toc}{section}{Notation and Conventions}
144 
145 A concrete realization is denoted by \(\mathfrak R\). Its interaction graph is \(G=(V,E,s,t)\), with local
146 state spaces \(X_i\) at vertices and transfer maps \(T_e:X_{s(e)}\to X_{t(e)}\) on directed edges. When a
147 core--residual decomposition is used, \(X_i=C_i\oplus R_i\), and \(H_e\) denotes the induced core transfer.
148 The exact structural workspace is the compatible-section space \(\Gamma(H)\); the transfer discrepancy is
149 \(D_H\), the transfer Laplacian is \(L_H=D_H^*D_H\), and \(\Gamma_\varepsilon(H)\) denotes a selected
150 low-energy workspace subspace. Dynamic workspace states are written \(W_t\), while \(Z_t\) denotes the
151 canonical predictive state associated with a declared future variable \(Y_t^+\).
152 
153 Unless stated otherwise, the structural results use finite-dimensional real or complex Hilbert spaces and
154 finite graphs. Probabilistic statements are made on standard Borel spaces when regular conditional laws are
155 required. All approximate conclusions are indexed by explicit tolerances, spectral gaps, intervention
156 families, loss functions, or null classes. A claim of relational irreducibility is always relative to a declared
157 decomposition of the realization into components. The symbol \(\mathcal D\) denotes one such
158 decomposition, and \(\mathfrak D\) denotes an admissible family used for robustness analysis. The complete
159 scientific specification is
160 \[
161 \varsigma=(\mathcal D,Y^+,\ell,\tau,\mathcal I,\mathcal N,\Theta),
162 \]
163 and \(\mathfrak S\) denotes an admissible family of specifications. A gate margin is written
164 \(m_q(W;\varsigma)\), with positive sign indicating passage under the declared direction and threshold.
165 
166 The terms \emph{workspace}, \emph{agency}, \emph{consciousness}, \emph{personhood}, and \emph{moral status}
167 are not interchangeable. A structural workspace may fail every operational test. An operational workspace
168 may lack persistence. A persistent relational workspace may lack goals or repair. A workspace agent, as
169 defined here, is a domain-relative functional organization and is not thereby a conscious subject. Definitions,
170 propositions, theorems, conjectures, and empirical tests are identified by their ordinary mathematical form;
171 all proved statements remain conditional on the assumptions given with them.

Back to source controls

Back to Stratified Relational Workspace Theory