TR-A-001 established that organizations face a structural gap between governance requirements and governance infrastructure — a finding validated by convergence across six independent traditions. That report named the gap and characterized it as architectural rather than disciplinary. But the characterization leaves a question open: even if the gap is architectural, could it be closed by means other than architectural imposition? If AI systems could learn governance invariances from organizational data, if organizational learning processes could discover and maintain governance properties at scale, if better training or more resources could resolve the persistent documentation failures the audit profession records — then the architectural approach, while valid, would be one option among several.

This report closes that question. The evidence shows that the processes organizations and computational systems rely on for improvement — learning, training, scaling, adaptive optimization — systematically destroy the very properties governance requires. The gap cannot be closed by doing more of what currently fails. The resolution must operate at a different layer: the architectural layer, where invariances are structurally given rather than computationally discovered.

The thesis would be disproven if: (1) a training procedure were demonstrated that provably preserves the symmetries Hassana Labs (2026) shows standard optimization breaks; (2) an organizational learning process were shown to reliably examine its own governing variables without architectural intervention; (3) a governance system were documented whose invariances were learned from data and maintained under operational stress at scale; or (4) a computational system achieved Ashby's (1956) requisite variety for governance without a model of the governance domain.

Five independent traditions — mathematical physics, ML theory, geometric deep learning, management cybernetics, and cognitive science — converge on a structural conclusion: governance invariances must be architecturally imposed because the processes relied on for improvement systematically destroy the properties governance requires.

The convergence is not a shared vocabulary or a coordinated research program. Each tradition reaches the conclusion from its own formal system, using its own methods, addressing its own domain. Noether's theorem (1918) proves that conservation laws require symmetries in the system's structure — a result from variational calculus with no connection to organizational governance, yet directly applicable: if governance properties are to be conserved, the system must possess the corresponding structural symmetries. Hassana Labs (2026) proves that training under log-loss — the standard optimization objective for modern AI systems — systematically breaks exactly the symmetries governance requires, and that this breaking is a mathematical consequence of the objective function, not a training failure that more data or compute could correct. Bronstein et al. (2021) provide convergent empirical evidence from geometric deep learning: across domains, architectural invariances — symmetries built into the network structure — consistently outperform learned invariances discovered through training. Conant and Ashby (1970) prove that every good regulator of a system must be a model of that system — governance without an architecturally structured model is not merely suboptimal but mathematically excluded. Sweller (1988) established cognitive load theory, whose three-load-type framework — formalized in subsequent work (Sweller, van Merrienboer, & Paas, 1998) — explains a fifty-year puzzle: why design rationale, governance documentation, and accountability infrastructure have failed to achieve sustained adoption despite mature theoretical foundations — every existing approach imposes extraneous cognitive load by requiring practitioners to step outside the work to document the work.

The convergence closes the design space. If training breaks the required symmetries (Chlon), if learning from data cannot produce the required invariances (geometric DL), if governance without a model is mathematically excluded (Conant-Ashby), and if the processes relied on for organizational improvement systematically prevent the examination of their own governing variables (Argyris, 1990) — then the resolution of the structural gap identified in TR-A-001 must be architectural. The gap cannot be closed by scaling what currently exists. Infrastructure must provide governance invariances as structural givens — properties of the system's architecture, not discoveries from its operation.

The structural property this report identifies is the architectural necessity of governance invariances: the property that governance-relevant symmetries, conservation laws, and structural constraints cannot arise from computational learning, organizational adaptation, or behavioral scaling — they must be architecturally given.

This property rests on three independent formal results that, taken together, close the design space for governance infrastructure:

The good regulator theorem. Conant and Ashby (1970) proved that every good regulator of a system must contain a model of that system — not a metaphorical model but one that is homomorphic to the system it regulates. Francis and Wonham (1976) strengthened this with the internal model principle: a controller that achieves perfect regulation must embed a model of the disturbance it rejects. Friston (2010) extended the argument through the free energy principle: any system that persists in a changing environment must minimize surprise, which requires maintaining an internal model that predicts environmental states (RA-008 §F5–F8). The chain from Conant-Ashby through Francis-Wonham to Friston establishes a mathematical exclusion: governance without an architecturally structured model of the governed domain is not a design choice that trades off against alternatives — it is mathematically impossible for a system to regulate what it does not model.

The symmetry-breaking proof. Hassana Labs (2026) proved that training under log-loss — the standard optimization objective used across modern language models and predictive systems — systematically breaks the symmetries that governance requires. The symmetry-breaking is not a training failure but a mathematical consequence of the loss function's structure: the optimization landscape under log-loss does not preserve the equivariance properties that governance invariances demand (RA-011 §F5–F6). This result forecloses the "just train harder" path. More data, more compute, more sophisticated training procedures cannot fix a problem that is structural to the optimization objective. If governance requires symmetries and training breaks symmetries, then training cannot produce governance.

Noether's theorem. Emmy Noether (1918) proved the foundational result in mathematical physics: every differentiable symmetry of a system's action corresponds to a conservation law. Time symmetry yields energy conservation. Spatial symmetry yields momentum conservation. The theorem is not domain-specific — it establishes a universal structural relationship between symmetries and conservation. Applied to governance: if governance properties are to be conserved (maintained under organizational operation), the system must possess the corresponding structural symmetries. Those symmetries must be architectural — properties of the system's structure, not emergent features of its dynamics (RA-011 §F1).

The three results compose into a single argument. Governance requires a model of the governed domain (Conant-Ashby). The model must possess structural symmetries for governance properties to be conserved (Noether). Training under standard objectives breaks those symmetries (Chlon). Therefore governance invariances cannot be learned — they must be architecturally given.

The Argyris-Chlon parallel. The impossibility does not stop at the computational boundary. Argyris and Schön (1978) demonstrated that organizational learning processes exhibit a structurally isomorphic pathology: defensive routines — skilled patterns of behavior that prevent the examination of governing variables — make organizations systematically unable to question the assumptions on which their governance operates. Argyris (1990) documented that "nearly all study participants espouse Model II values when asked how they would behave, but virtually all operate from Model I in actual problematic situations" (RA-014 §F3). The parallel is structural: Chlon's training optimization locally minimizes loss while breaking the symmetries governance requires; Argyris's defensive routines locally minimize embarrassment while preventing the governing-variable examination that organizational learning requires. Both are systems that optimize locally in ways that systematically prevent global improvement. Both resist correction through more of the same process — more training cannot restore broken symmetries; more organizational learning under Model I cannot produce Model II behavior. Both require architectural intervention: changing the system's structure, not scaling its existing process (RA-014 §F4, §S2).

This parallel establishes that the architectural necessity is scale-invariant. The pathology of local optimization preventing global improvement manifests at the computational scale (symmetry-breaking under training), at the organizational scale (defensive routines under Model I), and at the cognitive scale (System 1 heuristic processing displacing System 2 deliberation under load). At every scale, the same structural pattern appears: the process relied on for improvement destroys the properties improvement requires. The resolution at every scale is the same: architectural intervention that provides the required properties as structural givens rather than as outputs of the improvement process.

The cognitive load resolution. The architectural necessity also resolves a fifty-year puzzle in governance infrastructure adoption. Design rationale systems (Potts & Bruns, 1988; Lee & Lai, 1991), governance documentation standards (IIA Standard 2330, AU-C 230, GAGAS §6.50–6.59), and knowledge management methodologies (Walsh & Ungson, 1991) have been theoretically mature for decades. Yet adoption persistently fails — the audit profession records approximately 25% structural noncompliance despite enhanced oversight; knowledge management reports 50–70% initiative failure rates. Cognitive load theory (Sweller, 1988; Sweller, van Merrienboer, & Paas, 2019) provides a structural explanation: every existing approach requires stepping outside the work to document the work. This imposes extraneous cognitive load — load attributable to the documentation process rather than the governance task itself. When intrinsic load (task complexity) plus extraneous load (documentation burden) exceeds working memory capacity, germane processing (productive engagement with the governance content) collapses. Practitioners satisfice under cognitive constraint, producing the "click and hope" behavior that governance systems were designed to prevent. The infrastructure failure is the governance system itself (RA-010 §F2, §S1).

The resolution follows directly: governance context must accumulate as a structural by-product of organizational operation — not as a separate documentation step that competes for cognitive resources. This is the architectural-layer solution that cognitive load theory independently derives: reduce extraneous load to zero by making governance capture inherent in the work, not parallel to it. CLT has not been applied to governance or compliance system design — a confirmed gap in the literature (RA-010 §F2). The bridge from CLT to governance infrastructure design is an original contribution of the WMI research program.

Falsifiable claim. The architectural necessity of governance invariances is falsifiable by demonstrating any one of the following: (1) a training procedure that provably preserves governance-relevant symmetries under optimization; (2) an organizational learning process that reliably examines its own governing variables without architectural intervention; (3) a governance system whose invariances were learned from data and maintained under operational stress at scale; or (4) a computational system that achieves requisite variety for governance without a model of the governance domain. Each condition corresponds to the failure of one formal result in the argument chain. All four would need to be demonstrated for the architectural necessity claim to be fully overturned; any single demonstration would narrow rather than eliminate the claim.

Consequences. The architectural necessity yields three design requirements:

1. Governance invariances must be architecturally given. Structurally embedded in the infrastructure rather than learned from organizational data — because training breaks the required symmetries and organizational learning under defensive routines cannot examine its own governing variables.
2. Architectural invariances must be preserved under operational stress. Not merely established at design time but maintained as the system operates — because Noether's theorem requires ongoing structural symmetry for conservation, not a one-time design achievement.
3. The resolution must operate at the architectural layer. Because computational, behavioral, and scaling approaches cannot provide the required invariances — the good regulator theorem, the symmetry-breaking proof, and the Argyris-Chlon parallel each independently close one path to a non-architectural resolution.

The architectural necessity claim rests on convergence across five independent traditions. Each tradition reaches the same structural conclusion from different formal foundations, using different methods, addressing different domains. The convergence — not any single tradition's authority — is the evidence.

Mathematical physics. Noether's theorem (1918) establishes the foundational relationship between symmetry and conservation: every differentiable symmetry of a system's action corresponds to a conservation law. The theorem was derived in the context of variational mechanics and has governed physics for over a century. Applied beyond physics, the structural relationship holds: properties that are to be conserved — maintained invariant under system operation — require corresponding symmetries in the system's structure. The evidence shows that this relationship is not domain-specific but structurally universal (RA-011 §F1). The boundary for governance: organizational governance properties (accountability, authority delegation, constraint propagation) can only be conserved if the governance infrastructure possesses the corresponding structural symmetries. Noether's theorem does not prescribe which symmetries — it establishes that symmetries of some kind are architecturally necessary for conservation of any kind.

ML theory. Hassana Labs (2026) proved that training under log-loss — the dominant optimization objective across language models, predictive systems, and generative AI — systematically breaks the symmetries that governance requires. The proof is mathematical, not empirical: the loss function's gradient landscape does not preserve equivariance under the symmetry groups relevant to governance invariances. This is not a limitation of current training methods, training data, or compute resources — it is a structural property of the optimization objective itself (RA-011 §F5–F6). LeCun (2022) independently identified architectural insufficiency in current AI systems through the JEPA framework and the System 1/System 2 distinction for machine intelligence, while Bengio et al. (2013) documented representation learning limitations that Chlon's proof formalizes. The convergent critique — LeCun, Bengio, and Chlon reaching the same architectural boundary from different angles — establishes the ML theory tradition's independent conclusion: the invariances governance requires cannot emerge from training (RA-008 §F20).

Geometric deep learning. Cohen and Welling (2016) demonstrated that group-equivariant convolutional networks — networks whose architecture encodes the relevant symmetry group — consistently outperform networks that attempt to learn the same symmetries from data. Bronstein et al. (2021) systematized this finding in their comprehensive survey Geometric Deep Learning: Grids, Groups, Graphs, Geodesics, and Gauges, documenting the "GDL blueprint" across spatial, graph, and gauge domains. Weiler and Cesa (2019) extended the result to general E(2)-equivariant steerable CNNs. The pattern is consistent: when the relevant symmetry group is known, building it into the network architecture produces better performance, better generalization, and better invariance preservation than attempting to learn the same symmetry from data — even with large training sets and extensive compute (RA-011 §F8–F12). The geometric deep learning tradition provides convergent empirical evidence for the theoretical results: architectural invariances outperform learned invariances because learning is a symmetry-breaking process.

Management cybernetics. Conant and Ashby (1970) proved the good regulator theorem: every good regulator of a system must be a model of that system. Ashby (1956) established the law of requisite variety: a controller must have at least as much variety as the system it controls, or regulation will fail. Beer (1972, 1979, 1985) operationalized these principles in the Viable System Model — the recursive five-subsystem architecture for organizational viability. Francis and Wonham (1976) proved the internal model principle: a controller achieving perfect regulation must embed a model of the disturbance generator. The cybernetic tradition establishes that governance without a model is not a design trade-off but a mathematical impossibility. The model must be structurally adequate — possessing enough variety to match the governed system — and must be architecturally present, not emergently discovered (RA-008 §F5–F8). No formal organizational world model meeting these requirements exists in the literature. System Dynamics (Forrester, 1961; Sterman, 2000) is the closest existing approach but was designed for simulation rather than governance, missing the authority-chain and constraint-propagation properties that governance requires (RA-008 §F15–F16).

Cognitive science. Sweller (1988) established cognitive load theory, whose three-load-type framework — formalized in subsequent work (Sweller, van Merrienboer, & Paas, 1998) — identifies intrinsic (task complexity), extraneous (presentation and process burden), and germane (productive schema construction) load. When intrinsic plus extraneous load exceeds working memory capacity, germane processing collapses — the practitioner cannot engage productively with the governance content because the documentation process consumes the available cognitive resources. Kahneman (2011) documented the System 1/System 2 architecture of human cognition: load and fatigue push processing from deliberative System 2 to heuristic System 1, producing precisely the satisficing behavior that governance systems were designed to prevent. Simon (1947/1997) established bounded rationality as the baseline: decision-makers use aspiration levels rather than global optimization, and "click and hope" is rational satisficing under extreme cognitive load, not a character failure (RA-010 §F1). The convergence between Kahneman's cognitive System 1/2, LeCun's (2022) computational System 1/2, and Beer's (1972) cybernetic reactive/deliberative distinction — three independent traditions discovering the same dual-process architecture — validates the structural pattern: when load exceeds capacity, deliberative governance degrades to reactive heuristics (RA-010 §F3).

Convergence synthesis. Five traditions — mathematical physics, ML theory, geometric deep learning, management cybernetics, and cognitive science — reach the same structural conclusion from different formal systems:

• Mathematical physics (Noether): conservation requires architectural symmetry.
• ML theory (Chlon): training breaks the symmetries conservation requires.
• Geometric deep learning (Bronstein et al.): architectural invariances outperform learned invariances.
• Cybernetics (Conant-Ashby): governance without a model is mathematically excluded.
• Cognitive science (Sweller): the documentation paradigm imposes extraneous load that degrades the governance it aims to support.

None of these traditions set out to study organizational governance infrastructure. None shares methodology, vocabulary, or institutional affiliation with the others. Each reached the same boundary independently: the point where computationally or behaviorally derived properties fail and architecturally given properties are required. The convergence is the strongest form of evidence available for a structural claim: when independent formal systems produce the same conclusion, the conclusion reflects the structure of the domain rather than the perspective of any single tradition.

The architectural necessity argument is not purely theoretical. It resolves a documented pattern in the audit profession's record — a pattern the profession has measured, reported, and failed to resolve for decades.

The documented record is unambiguous. IIA Standard 2330 (Documenting Information), AU-C 230 (Audit Documentation), and GAGAS §6.50–6.59 (Documentation) establish requirements for governance documentation that the profession demonstrably cannot meet at the institutional level. TR-A-001 documented the structural noncompliance rate: approximately 25% of engagements under enhanced regulatory oversight produce documentation that does not meet the profession's own standards (RA-006 §F1; TR-A-001 §3.3). The IIA Common Body of Knowledge studies report chronic incompleteness in decision documentation across enterprise clients. These are not isolated failures — they are a structural regularity attested across decades of the profession's documented record.

The question the record poses is structural: why does persistent noncompliance endure despite mature standards, institutional backing, regulatory enforcement, and substantial organizational investment in governance infrastructure? The architectural necessity argument provides the answer. Every existing approach to governance documentation — audit workpapers, design rationale systems, compliance checklists, knowledge management repositories — requires practitioners to step outside the work to document the work. This is the structural requirement that cognitive load theory identifies as the source of failure: the documentation step imposes extraneous cognitive load (load attributable to the documentation process, not the governance task) that competes with intrinsic load (the actual governance decision's complexity) for finite working memory resources (RA-010 §F2).

The resolution is not better training, more discipline, or enhanced oversight — the profession has pursued all three for decades without reducing the structural noncompliance rate. The resolution is architectural: governance context must accumulate as a structural consequence of organizational operation, not as a parallel documentation step. This is not a preference or an optimization — it is the only approach that eliminates the extraneous cognitive load that the documented record shows persistently degrades governance quality.

Lerner and Tetlock (1999) established a complementary design principle: under certain conditions — particularly when accountability is established before commitment to a position — process accountability (accountability for how decisions are made) improves decision quality, while outcome accountability (accountability for what happened) can degrade it through defensive reasoning and premature closure. No published work has applied this framework to AI governance or algorithmic accountability — a confirmed literature gap (RA-010 §F6). Governance infrastructure that captures decision process — how governance decisions were made, what constraints were considered, what evidence was weighed — implements process accountability architecturally. Infrastructure that merely records outcomes and assigns retrospective blame implements outcome accountability and produces the defensive routines that Argyris (1990) documented.

The field evidence establishes that the architectural necessity is not an abstract formal conclusion — it resolves a concrete, documented, institutionally measured pattern of governance failure. The resolution operates at the architectural layer because the failure is architectural: it inheres in the structure of the documentation paradigm, not in the competence or discipline of practitioners.

The architectural necessity claim rests on convergence evidence — five independent traditions reaching the same structural conclusion — supplemented by three formal proofs (Noether, Chlon, Conant-Ashby) and one empirical pattern (geometric deep learning). The convergence argument establishes structural necessity with high confidence: independent formal systems producing the same conclusion is the strongest available evidence for a structural claim. However, the argument does not constitute a formal impossibility proof across all possible governance architectures — it establishes that specific categories of approaches (training, organizational learning, behavioral scaling) cannot close the gap, not that no future approach from an unanticipated direction could.

Five traditions are engaged. The selection is not exhaustive. Other traditions that might engage the architectural necessity boundary include: control theory beyond the Conant-Ashby/Francis-Wonham lineage, category theory and its applications to compositional systems, information geometry, and institutional economics. The five selected traditions were chosen because each provides a formally independent result — the convergence would be weakened if the traditions shared formal foundations. The five traditions selected here share no common formal system, which strengthens the convergence argument but leaves the question open whether additional traditions would strengthen or qualify it.

The architectural necessity argument has not been validated through implementation. The claim is that governance invariances must be architecturally given — it does not claim that any specific architecture adequately provides them. Validation would require demonstrating that an implemented architecture with structurally given invariances achieves the governance properties the traditions identify as necessary, and that removing the architectural invariances degrades governance quality in the predicted ways. This is the role of the forthcoming TR-E series.

The cognitive load theory resolution and the field evidence (§3.3) draw primarily on the internal audit profession's documented record — IIA, GAGAS, COSO, AU-C 230. The argument is offered as domain-general (applicable beyond audit), but the field evidence base is domain-concentrated. Other governance domains (legal compliance, healthcare quality, financial regulation, software engineering process) would provide additional field evidence; their absence does not weaken the formal arguments (which are domain-independent) but does narrow the field-evidence grounding.

Disposition: Strengthens-refines (provenance: plan)

The evidence reviewed in §3 strengthens WMI-P07 — the position that governance structures must be anchored outside the learning loop. Five independent traditions converge on this claim. Noether's theorem establishes that conservation requires structural symmetry, not emergent symmetry (RA-011 §F1). Chlon's proof demonstrates that training under standard objectives breaks the symmetries governance requires (RA-011 §F5–F6). The Argyris-Chlon parallel establishes that organizational learning processes exhibit the same structural pathology — local optimization preventing governing-variable examination — at the organizational scale (RA-014 §F4). The good regulator theorem requires that the governance model be architecturally present, not emergently discovered (RA-008 §F5–F8). Geometric deep learning provides convergent empirical evidence that architectural invariances outperform learned ones (RA-011 §F8–F12). Per rwp-wr-oq.md §5.1.1, WMI-P07 was "provisionally closing with strong supporting evidence." This report delivers the five-tradition convergence evidence base for that closure: the anchor must be outside the learning loop because learning destroys the properties the anchor is required to preserve.

Disposition: Strengthens-refines (provenance: plan)

The evidence strengthens WMI-P09 — the position that governance infrastructure must function as a socket (defining the interface) rather than a plug (binding to a specific runtime). The geometric deep learning evidence is particularly relevant: architectural invariances work precisely because they encode the symmetry group structure without specifying the implementation details of the processing that occurs within that structure (RA-011 §F8–F12). The good regulator theorem's state-prediction formalism is domain-general — it applies equally to physical, biological, cognitive, and organizational systems (RA-008 §F9). The architectural necessity argument reinforces runtime agnosticism: if invariances must be architecturally given, then the architecture must define the governance interface without prescribing the computational or organizational processes that operate within it.

Disposition: Strengthens-refines (provenance: plan)

The evidence strengthens WMI-P15 — the position that architectural decisions in governance infrastructure are inherently ethical decisions. The five-tradition convergence establishes that the choice of whether to impose governance invariances architecturally is not a neutral engineering decision — it determines whether governance properties can be conserved at all (RA-011 §F16). The Argyris-Chlon parallel shows that deferring to "more training" or "better organizational culture" as alternatives to architectural imposition is not merely suboptimal but structurally futile — these processes systematically destroy the properties they aim to produce (RA-014 §F4). The cognitive load theory evidence shows that governance infrastructure that imposes extraneous documentation burden degrades the decision quality it aims to protect (RA-010 §F2). Every architectural choice — what to embed structurally, what to leave to learning, what to impose versus what to discover — directly shapes whether governance can function.

Disposition: Strengthens-refines (provenance: plan)

The evidence strengthens WMI-P01 — the position that organizational world models require governed composition (structural rules for how components compose). The good regulator theorem requires that governance infrastructure contain a model of the governed system (RA-008 §F5–F8). Noether's theorem requires that the model possess structural symmetries for governance properties to be conserved (RA-011 §F1). The combination establishes that the model's composition must be governed — structurally constrained to preserve the symmetries that governance properties depend on — rather than free-form or emergently organized.