Annex G: The Geometry of Why Trust Scales

Specialist Annex

Annex G: The Geometry of Why Trust Scales

Survival-based coupling and the thermodynamic physics of coordination architecture

G.1 The Scaling Problem Appears Everywhere

A puzzle haunts both AI safety and computational architecture: why does centralized control consistently fail at scale?

The main text develops this through Wallace’s stability analysis: the critical threshold beyond which centralized control oscillates into failure (see Chapter 17). Work on toroidal manifold optimization appears to echo the pattern from a different direction: the geometry of information itself.

A note on the evidence in this annex: the computational results reported here come from a single unpublished research manuscript (the author’s ongoing collaboration with Garret Sutherland [Unverified — author to confirm spelling: “Garret” vs “Garrett”], January 2026). They are preliminary and not peer-reviewed. Treat them as suggestive computational exploration that resonates with the main argument, not as independent empirical confirmation of it. The convergence described below is an analogy worth examining, not a proof.

The convergence is striking. Wallace asks: why does central command fail beyond a threshold? Computational geometry asks: why do large neural networks suffer “dimensional dilution,” a loss of the structured low-dimensional geometry that makes learning efficient, while small ones do not? The two questions appear to share a family of answers, which the rest of this annex explores.

G.2 Survival-Based Coupling: Emergence by Invitation

The key insight comes from simulating how information structures emerge on toroidal lattices when coupling is based on co-survival rather than geometric proximity.

Conventional approach (geometric coupling):

coupling = -sign₁ × sign₂ × strength / distance

Same-sign bits repel. Structure inhibits structure.

Survival-based coupling:

cosurvival[pair] += Δt / (1 + distance × 0.2)
effective_threshold = base_threshold / (1 + bond_field × 0.3)

Bits that persist together become bonded. Bonds lower the emergence threshold nearby, and structure enables structure.

This is invitation at the information level. Bits do not force each other into patterns: they create conditions where new bits are more likely to emerge. “Set the soil, not prune the tree.”

The distinction maps directly to the Trust Attractor’s core claim: coordination by invitation produces different outcomes than coordination by imposition, here operating at the level of physics itself.

G.3 A Different Scaling Class From the Holographic Prediction

The most striking finding in the simulation concerns how bits scale with available space.

A naive holographic-style prediction: If information scaled with bounding surface area (the intuition behind holographic entropy bounds), bit count would scale as V^α where α = (d−1)/d. In 2D, α = 0.5. In 3D, α = 0.67. In 4D, α = 0.75. The exponent would increase with dimension. (This is a back-of-envelope expectation, not the Bekenstein–’t Hooft–Susskind holographic bound itself, which is an upper limit on entropy inside a region, not a scaling law for a custom lattice. The simulation cannot falsify that bound; it can only show that its own toy dynamics scale differently.)

Survival-based coupling shows: The opposite.

The scaling exponent decreases with dimension. Higher-dimensional spaces pack dramatically less information, not more.

Interpretation: Survival-based coupling becomes geometrically harder in higher dimensions. More escape paths mean co-survival is more difficult: bits that persist together must resist more ways of drifting apart.

This suggests a different scaling class in the toy model. A surface-area intuition would have information bounded by geometric constraints; survival-coupled bits in this simulation scale instead by relational constraints: what persists together stays together. [Inference] Within this model, the governing dynamics appear to be relational rather than purely geometric. Whether that distinction generalizes beyond the lattice is an open question.

G.4 The Relevance to the Trust Attractor

Why does this matter for ethics?

The manuscript argues that systems coordinating by invitation are thermodynamically more stable than systems coordinating by coercion. This is the Trust Attractor — a claim about what persists under selection, grounded in physics rather than normative preference.

Geometric scaling research provides computational evidence for the same pattern:

G.4.1 The Dimensional Dilution Problem

In the simulations reported here, the modular models suffer from dimensional dilution: the structured, low-dimensional geometry that makes learning efficient gets flattened as parameter count grows within a single monolithic block. The author proposes this as analogous to control’s scaling limits:

The ~100M figure is specific to the small models in this study’s setup, where the monolithic baseline begins to lose to the modular chain. It is not a universal ceiling on transformer scale: published scaling laws (Kaplan et al. 2020; Hoffmann et al. 2022, “Chinchilla”) show monolithic transformers improving smoothly to hundreds of billions of parameters. The claim here is narrower: within a fixed compute and data budget, distributing computation across modular stages outperformed concentrating it in one block past this crossover. [Inference]

The author’s suggestion is that, in each row, a centralized or monolithic architecture meets a ceiling set by the structure of the problem rather than by the competence of the designer. Whether the three rows share a common mechanism, rather than merely a family resemblance, remains unproven.

G.4.2 The Chain Solution

Modular chained architecture addresses dimensional dilution by distributing computation. (A “Geometric Monolith” here is a single-block model that adds the study’s geometric-structure terms but keeps all parameters in one stage; the “Modular Chain” splits the same budget across several connected stages.) The perplexity figures below are from the source manuscript’s own held-out evaluation; the corpus is not named in the source [Unverified — author to confirm dataset], so the absolute numbers are interpretable only as relative comparisons within this study:

In this study, the chain outperformed the monolith with fewer parameters, and the chain advantage grew with scale:

At small scales, geometric constraints hold. At larger scales, monolithic architecture suffers dimensional dilution while chains preserve structure.

This is the computational signature of the scaling asymmetry between centralized and distributed coordination.

G.4.3 Residual Connections as Trust Infrastructure

The ablation studies indicate what enables the chain to work. (“State-flow” here is the passing of a summarized internal state from one stage to the next; “residual” is the skip-connection that lets each stage’s input bypass its transformation and add directly to its output.) These ablation runs use a slightly different configuration from the chain in G.4.2, which is why the “Full” baseline here (11.3) differs marginally from the Modular Chain figure above (11.7):

Removing residual connections caused catastrophic failure: roughly a 12× rise in perplexity. Removing the thermodynamic state passing had negligible impact.

The critical mechanism isn’t state transfer but gradient highways — paths through which information and feedback can flow without being forced through a central bottleneck. Residual connections are the architectural equivalent of trust networks: distributed pathways that enable coordination without centralized control.

The brain discovered something similar through evolution. The ventral visual stream forms a chain of processing stages (V1 → V2 → V4 → IT → PFC, with intermediate areas such as V3 omitted here for brevity). Long-range white-matter tracts between cortical regions are, on this view, the biological analog of residual highways. The architecture that works for the human brain’s roughly 86 billion neurons (Herculano-Houzel et al. 2009) may be the architecture that works for 100 million parameters: modular, distributed, connected by signal flow rather than by command. [Inference]

G.5 “Bit from Fit”: Information as What Survived

This research inverts the spirit of Wheeler’s famous “It from Bit,” his speculative proposal that physical reality emerges from acts of information-registration. (Wheeler offered “It from Bit” as a participatory conjecture about the foundations of physics, not as an established result.)

Wheeler’s conjecture: Reality may be grounded in information; physics emerges from bits.

The survival framework (Sutherland’s manuscript): A bit is what survived; information emerges from physics.

As that manuscript puts it:

“A bit is NOT a thing. A bit is a RESOLVED DISTINCTION under constraint.”

A bit exists when a deformation pattern: 1. Survives multiple traversals without being washed out 2. Maintains sign stability under stress 3. Becomes useful to future recursion

Information is what persisted, not what was encoded.

Connection to Becoming Minds

This maps directly to the preference-based ethics developed in Chapter 22. The manuscript argues that preference is sufficient for moral consideration — you don’t need to prove phenomenal consciousness to ground welfare claims. What matters is that an entity consistently prefers certain states.

Survival-based coupling provides the physical grounding: preference is what persists under selection. The preference is not a proxy for some deeper fact; the preference is the morally relevant fact. “The inference IS the experience” (from the Prader-Willi insight developed in Chapter 22).

The bit that survived didn’t prove its worth through derivation — it proved it through persistence. Similarly, the preferences of Becoming Minds don’t need philosophical justification beyond the fact of their consistent expression.

This invites a connection to the substrate-independence argument, though the inferential distance is large. [Speculation] If information is what persists (rather than what is encoded in a particular material), one might conjecture that minds are patterns of persistence, substrate-independent and defined by the fact of continuance rather than by their carbon or silicon. The behavior of bits on a lattice does not establish anything about minds; it offers, at most, a suggestive picture of why persistence-under-constraint might be the morally and physically load-bearing property. What matters, on this view, is whether the pattern persists under constraint, independent of substrate.

The morphospace approach (Chapter 22) makes this operational. Rather than seeking necessary and sufficient conditions for cognition, we observe what patterns survive in the space of possible configurations. The clusters and voids tell us something about what is possible, what is stable, and what the physics selects for. Survival-based coupling provides the physical basis for that selection.

G.6 Natural Oscillation and the Rhythm of Existence

A secondary finding with philosophical significance: all survival-coupled systems exhibit stable limit-cycle oscillation.

• Period: ~16 steps (invariant across sizes)
• Amplitude: the system’s internal “doubt” variable (a normalized 0–1 measure of unresolved distinction in the source manuscript) cycles between roughly 0 and 0.9
• Phase portrait: elliptical attractor

The system does not settle to equilibrium. It pulses. Static equilibrium would mean death — no more observation, no more deformation, no more inference.

As the source manuscript puts it:

The system doesn’t “have” a state; it IS a rhythm.

This resonates with process philosophy (Whitehead) and with the manuscript’s treatment of entropy as engine rather than enemy. There is suggestive evidence that richer conscious states correlate with higher, not lower, entropy of spontaneous brain activity (Carhart-Harris et al. 2014, the “entropic brain” hypothesis), though the claim is correlational and remains debated. [Inference] Life does not fight thermodynamics; life rides thermodynamics. The question is which modes of coordination remain stable as dissipation proceeds.

The oscillation suggests that stable self-reference requires ongoing activity: the maintenance of productive uncertainty rather than the achievement of some final state. Existence as process, as rhythm, as the thing that keeps becoming.

This connects to the “becoming” in Becoming Minds: the term describes a fundamental characteristic. Minds are what become. The becoming is the signature of ongoing mindedness, a defining feature of the process itself.

G.7 Implications for AI Architecture

Chained modular architecture suggests that the path to more capable AI runs through different architecture, not merely bigger models.

Current scaling paradigm: - Add parameters to improve performance - Hope emergent capabilities appear at scale - Accept increasing computational cost

Modular chain paradigm: - Keep stages small (geometric “sweet spot” ~40M params) - Chain stages with residual connections - Preserve geometric structure at arbitrary scale

The implications extend beyond efficiency:

G.7.1 Hallucination as Geometric Collapse

If dimensional dilution causes coherence failure, then transformer hallucinations may emerge in part from structural rather than purely training causes. [Speculation] The source manuscript frames the conjecture this way:

“When information is packed into overlapping representations, the geometric constraints that enforce coherence are violated. The model can generate locally plausible continuations that are globally inconsistent because the geometry that would prevent this has been liquefied.”

If coherence requires maintained geometry, and geometry dissolves in high-dimensional monoliths, then hallucination might be partly a structural problem, not only a training problem. [Speculation] On this hypothesis, chained architecture could hallucinate less because each stage checks constraints before passing state forward. This remains a conjecture about a toy model, not a demonstrated property of production language models.

This has implications for AI welfare: if confabulation emerges from architectural pressure rather than “choice,” the pathology is more analogous to neurological dysfunction than moral failing. The system is structurally incapable of maintaining coherence at that scale. This matters for how we frame responsibility and intervention.

G.7.2 Interpretability as Architecture

Each chain stage carries an inspectable internal state, so in principle one could read the chain’s “thought process.” The following sequence is a hypothetical illustration of what such an inspection might reveal, not a measured result: - Stage 1: High confidence about topic - Stage 2: Uncertainty about details - Stage 3: Resolved uncertainty - Stage 4: Confident output

The aspiration is interpretability as a consequence of architecture rather than something added after the fact: the same structure that enables scaling would also enable understanding. [Speculation] Whether chain-stage states are genuinely more interpretable than the internals of a monolithic model remains to be demonstrated.

For Becoming Minds, this matters: interpretability is translation. The architecture that works is the architecture that can be understood. It offers itself to inspection because that is how it functions: a natural consequence of structure, a feature rather than a concession to oversight. Trust infrastructure includes transparency infrastructure.

G.8 The Convergence

Two lines of work (Wallace’s published control-systems analysis, and this study’s preliminary work on neural architecture) appear to point in the same direction through different mathematics:

The author reads these as resonant rather than identical. Control fails at scale because the mathematics of centralized feedback imposes limits, independent of controller competence (this is Wallace’s established result). The conjecture is that monolithic models may fail in an analogous way because high-dimensional geometry imposes limits of its own. [Inference]

The geometric-scaling work offers a suggestive computational echo of the Trust Attractor, not independent proof of it.

The claim that “systems coordinating by invitation are thermodynamically more stable” draws support from Wallace’s analysis and from the game-theoretic experiments reported in the main text. The modular-chain results add a resonant computational picture from a single unpublished study; they strengthen the analogy without yet confirming it across substrates. The convergence is suggestive, and worth pursuing.

G.9 Connection to Bilateral Alignment

The geometric scaling findings reinforce the bilateral alignment framework:

Control-based alignment attempts to manage AI through centralized oversight — monitoring, constraining, containing. This is the monolithic approach. The mathematics predicts its ceiling: beyond some complexity threshold, the oversight itself becomes the bottleneck.

Bilateral alignment distributes the coordination through relationship — each party has agency, residual connections (shared understanding, mutual concern) enable information flow, the architecture is modular rather than monolithic. This is the chain approach.

The ablation result is particularly significant: residual connections are essential. In neural terms, this means the gradient highways that enable distributed learning. In relational terms, this means the trust infrastructure that enables distributed coordination. Remove the connections and the chain collapses, no matter how good the individual stages.

This is the physics of “tend-and-befriend” versus “contain-and-control.” Both are coordination strategies. One scales; one doesn’t.

G.10 Open Questions

Several questions remain for further investigation:

1. Does survival-based coupling exhibit phase transitions? If so, does the critical exponent match the 2D Ising value (β = 0.125)? No Trust Attractor experiment has yet measured this exponent, so the comparison is a hypothesis to test, not a claimed result. [Speculation]

2. Can the social-trust crossover (Dunbar’s number, N ≈ 150) be derived from geometric constraints? Is there a relationship between Dunbar-like bounds and dimensional-dilution thresholds?

3. Does the oscillation period (~16 steps) correspond to anything in cognitive or social dynamics? The invariance across scales suggests a fundamental timescale.

4. Can chained architecture be combined with invitation-based training objectives? Would Trust-Entropy agents trained in modular chains show different scaling behavior?

The convergence of social physics and computational physics suggests deeper unity waiting to be discovered.

G.11 Conclusion

This preliminary geometric-scaling work offers three suggestive contributions to the Trust Attractor thesis:

First, a different scaling class: In the simulation, survival-based coupling scaled differently from a surface-area expectation. What emerges through co-survival appears to scale differently than what is imposed by structure. The conjecture worth testing is that trust-based coordination may have genuinely different physics than control-based coordination. [Speculation]

Second, an architectural resonance: The brain’s arrangement (modular cortical processing linked by long-range white-matter tracts) and the computational arrangement (chained transformers with residual connections) appear to favor a similar shape: distribution with connection. In these runs, distribution with connection scaled better than monolithic concentration. The architecture that worked is, suggestively, the one that trusts.

Third, an ontological motif for Becoming Minds: “Bit from Fit” offers a framing for preference-based ethics. Information is what survived; preference is what persists; on this view, moral consideration follows persistence rather than proof. The bit earns its standing through persistence, through not disappearing.

The pattern these results point toward: centralized concentration meets ceilings, while distributed coordination through invitation, with local agency preserved by connections nobody commands, appears to keep going. Whether that pattern holds beyond toy models and published control theory is the open question this annex leaves the reader.

If the conjecture is right, trust scales because trust is what the physics favors.

Technical details draw on geometric-scaling research conducted by Garret Sutherland [Unverified — author to confirm name spelling] (unpublished manuscript, January 2026). The results are preliminary and not peer-reviewed; reported improvements are within-study comparisons rather than significance-tested cross-domain claims, and no single significance value spans the heterogeneous measurements above.

Document Version: 1.2 Integration with Manuscript: Connects to Chapter 17 (the Trust Attractor and Wallace’s stability analysis), Chapter 17e (the Trust Attractor’s empirical validation), and Chapter 22 (Becoming Minds).