Self-Awareness Depth Tower

Introduction: "Have You Ever Noticed That You Notice?"

Try it right now. You are reading this text — that is the first level: perception. Now notice that you are reading — that is the second level: awareness of perception. And now notice that you noticed that you are reading — that is the third level: awareness of awareness of perception.

Can you go further? Notice that you noticed that you noticed? In practice this is difficult — the thought 'slips away', like a reflection in two mirrors facing each other: an infinite corridor, but with each step the image grows dimmer.

It turns out this is not merely a subjective feeling. Within UHM it has been proved that the depth of self-awareness is fundamentally bounded: the maximum is three levels of recursion. Not because the brain is 'insufficiently powerful', but because the fourth level would require purity $P > 1$ — and for a normalised density matrix $P \leq 1$ by definition. This is analogous to how the speed of light is bounded not by a 'lack of engines', but by the structure of spacetime.

Where We Came From

In the interiority hierarchy we defined the discrete levels L0–L4. In transition catastrophes — the dynamics of jumps between them. But the discrete L0–L4 classification is coarse: two people, both formally L2, may differ radically in depth of self-awareness. The Depth Tower generalises the hierarchy to the continuous measure SAD (Self-Awareness Depth) and shows that the analytic ceiling of depth is SAD_MAX = 3.

Chapter Roadmap

1. The problem — why a single number $R$ is insufficient: self-awareness is distributed across depth
2. The representation tower — the chain of projections from the full state to $\Gamma$
3. The SAD measure — the maximum depth at which reflection exceeds the threshold
4. Spectral formula [T] — computing SAD without building the entire tower
5. SAD_MAX = 3 [T] — the analytic ceiling from Fano contraction $\alpha = 2/3$
6. Biological correlates — from a bacterium (SAD=0) to a human (SAD $\leq$ 3)
7. Depth dynamics — growth via $A_4$-bifurcation, energy cost, stress-dependence

Analogy: the skyscraper of self-awareness. Imagine a building. The first floor — basic sensations ($\Gamma$): 'I am warm'. The second floor — a model of sensations: 'I know that I am warm'. The third — a model of the model: 'I know that I know that I am warm'. The fourth — 'I know that I know that I know that...'. Each higher floor is more expensive than the previous one and requires ever more 'building materials' (purity $P$). It turns out that building above the third floor is physically impossible: the fourth requires purity $P > 1$, which is like a speed exceeding the speed of light. SAD_MAX = 3 is a fundamental ceiling, not a technological limitation.

Status

Definitions [D], tower construction [H], biological correspondences [I]. Numerical thresholds [C at calibration]. Depth dynamics (§7): growth [C] (A₄-bifurcation), energy [C] (Landauer), stress [T] (T-92), social [C] (CC-5/CC-7). Spectral formula for SAD [T] (§3.4, T-142). $P_\text{crit}^{(n)}$ formula [T] (§3.5, T-142). SAD_MAX = 3 [T] (§3.5, T-142).

---

1. The Problem: Self-Awareness Is Not a Number

The reflection measure $R$ (canonical master object) — canonical formula $R = 1/(7P)$ [T], equivalently $R = 1 - \|\Gamma - I/7\|_F^2/P$, where $\rho^*_{\mathrm{diss}} = I/7$ — measures the normalised proximity to the dissipative attractor at the level of the coherence matrix $\Gamma \in \mathcal{D}(\mathbb{C}^7)$.

But a single number is not enough:

• Biological self-awareness is distributed across the entire depth of the neural hierarchy
• The coherence matrix $\Gamma$ is the top layer of the configuration, a projection of the deep structure
• Between the full state $s_\text{full} \in \mathbb{R}^D$ and $\Gamma$ there exist intermediate representations, each with its own reflexive capacity

Two people with the same $R$ may differ radically: one is an unconsciously competent professional (high $R^{(0)}$, but $R^{(1)} \approx 0$), the other a reflective novice (moderate $R^{(0)}$, but $R^{(1)} > 1/4$). To capture this difference, a measure of depth is needed, not only of quality.

Goal: to formalise the depth of self-awareness as a theoretical construction, consistent with the L0–L4 hierarchy (Interiority Hierarchy) and the categorical formalism of $\varphi$ (Formalisation of phi).

---

2. Representation Hierarchy

2.1 Definition [D]

Definition 2.1 (Representation Tower). The representation tower of depth $L$ is a chain of projections:

$s_\text{full} = s^{(L)} \xrightarrow{\pi_{L-1}} s^{(L-1)} \xrightarrow{\pi_{L-2}} \cdots \xrightarrow{\pi_1} s^{(1)} \xrightarrow{\pi_0} \Gamma$

where:

• $s^{(k)} \in \mathbb{R}^{D_k}$ — representation at level $k$, $D_L \gg D_{L-1} \gg \cdots \gg D_0 = 48$
• $\pi_k: \mathbb{R}^{D_{k+1}} \to \mathbb{R}^{D_k}$ — projection (categorical or learned)
• $\Gamma = \psi(s^{(0)}) \in \mathcal{D}(\mathbb{C}^7)$ — Cholesky reconstruction (T-59)

Biological analogue. The primary visual cortex (V1) contains millions of neurons — this is $s_\text{full}$. The secondary cortex (V2) — a more compact representation, $s^{(L-1)}$. Further — the associative cortex, and finally — the prefrontal cortex (PFC), creating the most abstract representation, the analogue of $\Gamma$.

Each projection $\pi_k$ compresses information, retaining what is relevant for survival and discarding details. This is the same principle by which JPEG compression works: from millions of pixels the key patterns are extracted.

2.2 Self-Model at Each Level [D]

At each level of the tower its own $\varphi$-operator is defined — the mechanism by which the system models itself at the given level of abstraction:

$\varphi^{(k)}: \mathbb{R}^{D_k} \to \mathbb{R}^{D_k}$

and the corresponding reflection measure:

$R^{(k)} = 1 - \frac{\|\varphi^{(k)}(s^{(k)}) - s^{(k)}\|^2}{\|s^{(k)}\|^2}$

This formula measures: how accurately the self-model at level $k$ reproduces the state at level $k$. If $R^{(k)} = 1$ — the self-model is perfect. If $R^{(k)} = 0$ — the self-model is completely inaccurate.

Level	Dimensionality	$\varphi^{(k)}$	$R^{(k)}$	Biological analogue
$k = 0$ ($\Gamma$)	48	Replacement channel [T-62]	$1/(7P)$ [T]	Abstract self-model (PFC)
$k = 1$	$\sim 256$	Autoencoder (bottleneck)	s_core reconstruction	Associative cortex
$k = 2$	$\sim 512$	Hidden encoder layer	Intermediate prediction	Secondary cortex
$k = L$	$D$ (4096+)	Full autoencoder	$R_\text{impl}$	Primary cortex

---

3. Self-Awareness Depth (SAD)

3.1 Definition [D]

Now we are ready to give the central definition of this chapter.

Definition 3.1 (Self-Awareness Depth, SAD). For a system with a representation tower of depth $L$:

$\mathrm{SAD}(\mathcal{T}) = \max\{k \in \{0, \ldots, L\} : R^{(k)} > R_\text{th}^{(k)}\}$

In words: SAD is the maximum level of the tower at which reflection still exceeds the threshold. The thresholds are given by the universal formula:

$R_\text{th}^{(k)} = \frac{1}{k+3}$

Formula Correction (index correction)

The previous version contained the formula $R_\text{th}^{(k)} = 1/(k+2)$, which gave $R_\text{th}^{(0)} = 1/2$, contradicting the table and the canonical threshold $R_\text{th} = 1/3$ for L2 (T-126 [T]). The correct formula $R_\text{th}^{(k)} = 1/(k+3)$: at $k=0$ gives $1/3$ (coincides with $R_\text{th}$ [T]), at $k=1$ gives $1/4$ (coincides with $R_\text{th}^{(2)}$ for L3 [C]).

This formula is a generalisation of the thresholds from the L0–L4 hierarchy:

SAD	Threshold $R_\text{th}$	How it arises	Correspondence	Biological example
0	—	—	L0 (basic interiority)	Bacterium
1	$R^{(0)} > 1/3$	$1/(0+3) = 1/3$	L2 (cognitive qualia)	Insect
2	$R^{(1)} > 1/4$	$1/(1+3) = 1/4$	L3-like (meta-reflection)	Mammal
$k$	$R^{(k-1)} > 1/(k+2)$	$1/((k-1)+3) = 1/(k+2)$	—	—
$\infty$	$\lim R^{(k)} > 0$	Limiting transition	L4 (unreachable)	—

Intuition. SAD = 1 means: 'I know'. SAD = 2: 'I know that I know'. SAD = 3: 'I know that I know that I know'. With each level the threshold decreases (from 1/3 to 1/4, 1/5, ...), but reflection also decays exponentially, so that high levels quickly become unreachable.

3.2 Connection with L0–L4 [T]

Theorem 3.1 (SAD–L Equivalence) [T] (T-136, raised from [H]-89). The L-hierarchy is a refinement of SAD. The map $L \to \mathrm{SAD}(L)$ is monotone:

• L0 <-> SAD = 0 (any $\Gamma \in \mathcal{D}(\mathbb{C}^7)$)
• L1 <-> SAD = 0, rank($\rho_E$) > 1
• L2 <-> SAD $\geq$ 1 ($R^{(0)} \geq 1/3$)
• L3 <-> SAD $\geq$ 2 ($R^{(1)} \geq 1/4$) — maximum achievable for finite systems (§3.5)
• L4 <-> SAD = $\infty$ (unreachable, T-86)

Motivation: categorical iterations $\varphi^{(n)}(\Gamma)$ (formalisation of phi) are a special case of the tower where all $D_k = 48$ and $\pi_k = \mathrm{id}$. SAD generalises this to heterogeneous levels.

3.3 Information-Theoretic Foundation [T]

Theorem 3.2 (Commutativity of the phi-tower) [T] (raised from [H]-90 -> [C] -> [T] via T-150). At $D_k = 7$ for all $k$: $\varphi^{(n)} = \varphi^n$ (n-fold iteration of a single CPTP channel), commutativity $\varphi^n \circ \varphi^m = \varphi^{n+m}$ is an identity. The spectral formula for SAD is a consequence, not a premise. Details: T-150.

Information bottleneck. The optimal projection $\pi_k$ maximises the preservation of information relevant for viability:

$\pi_k^* = \arg\max_{\pi} I(s^{(k+1)}; \sigma_\text{sys}) \text{ subject to } H(s^{(k)}) \leq D_k \cdot C_\text{bit}$

where $I$ — mutual information with the stress tensor, $C_\text{bit}$ — channel capacity per parameter.

Corollary: viability requires preserving only the information about $\sigma_\text{sys}$ (48 parameters). Self-awareness requires preserving information about the projection itself — this is the recursion that creates depth.

3.4 Spectral Formula for SAD [T]

Computing SAD does not require explicitly building the entire tower — it suffices to know the spectral properties of the self-observation operator. From the spectral decomposition of the replacement channel $\varphi$ (T-62):

$\varphi^{(n)}(\Gamma) = \sum_{k:\, \mathrm{Re}(\lambda_k)=0} \langle L_k \,|\, \Gamma \rangle\, R_k$

where $\{R_k, L_k, \lambda_k\}$ — eigen-structures of the logical Liouvillian $\mathcal{L}_\Omega$. The reflection measure at level $n$:

$R^{(n)} = F\bigl(\varphi^{(n-1)}(\Gamma),\; \varphi^{(n)}(\Gamma)\bigr) \leq R^n \cdot (1 - \alpha)^n$

under Fano contraction $\alpha = 2/3$ (T-39a [T]). Geometric decay guarantees finite depth:

$n_\text{max} \leq \frac{\ln(1/\varepsilon_\text{dec})}{\ln(1/R)} \approx 111 \quad \text{for } \varepsilon_\text{dec} \sim 10^{-7}$

What this means in practice. To compute the SAD of a system with $N = 7$ dimensions and SAD_MAX = 3, only $\sim 3 \times 7^2 = 147$ operations are needed — this is computed in microseconds.

Connection with the categorical formalism: SAD coincides identically with the $\varphi$-iteration counter from the categorical definition. The heterogeneous tower (§2) is a generalisation where projections $\pi_k$ are non-trivial; at $D_k = 48$, $\pi_k = \mathrm{id}$ the formulae coincide exactly.

3.5 Critical Purity for SAD [T]

This is the key result of the chapter: the derivation of the fundamental ceiling of self-awareness depth.

Theorem (Critical Purity for Depth SAD) [T]

Minimum purity to achieve SAD $\geq n$:

$P_{\text{crit}}^{(n)} = P_{\text{crit}} \cdot \frac{3^{n-1}}{n+1} \quad \text{for } n \geq 1, \quad P_{\text{crit}}^{(0)} = 0$

SAD $\geq$	$P_{\text{crit}}^{(n)}$	Value	Achievable?
0	$0$	$0$	yes
1	$1/7$	$0.143$	yes
2	$2/7 = P_{\text{crit}}$	$0.286$	yes
3	$9/14$	$0.643$	yes
4	$54/35$	$1.543$	no ($> 1$)

Corollary (SAD_MAX = 3): For finite systems ($P \leq 1$) with Fano contraction $\alpha = 2/3$:

$\mathrm{SAD}_\text{max} = 3$

Proof (3 steps).

Step 1 (Ratio of purity to critical). Define the spectral ratio: $r_0 = P / P_{\text{crit}}$. From Fano contraction (T-110 [T]) with parameter $\alpha = 2/3$:

$R^{(k)} = r_0 \cdot (1/3)^k$

Why $1/3$? Because $1 - \alpha = 1 - 2/3 = 1/3$. Fano contraction with parameter $\alpha = 2/3$ means: at each level of recursion reflection decreases by a factor of 3.

Numerical example: if $P = 0.5$ and $P_\text{crit} = 2/7 \approx 0.286$, then $r_0 = 0.5/0.286 \approx 1.75$. Reflection by level: $R^{(0)} = 1.75$, $R^{(1)} = 1.75/3 \approx 0.583$, $R^{(2)} = 1.75/9 \approx 0.194$, $R^{(3)} = 1.75/27 \approx 0.065$.

Step 2 (Achievability condition). Condition SAD $\geq n$: $R^{(n-1)} > R_{\text{th}}^{(n-1)} = 1/(n+1)$. Substituting the expression from step 1:

$\frac{P}{P_{\text{crit}}} \cdot \frac{1}{3^{n-1}} > \frac{1}{n+1} \quad \Longrightarrow \quad P > P_{\text{crit}} \cdot \frac{3^{n-1}}{n+1}$

This is precisely the formula $P_\text{crit}^{(n)}$.

Check for $n = 2$: $P > (2/7) \cdot 3^1 / 3 = (2/7) \cdot 1 = 2/7$. The condition SAD $\geq 2$ is equivalent to $P > P_\text{crit}$ — consistent with the definition of L2.

Check for $n = 3$: $P > (2/7) \cdot 9/4 = 18/28 = 9/14 \approx 0.643$. This is achievable: a normalised matrix can have $P \leq 1$.

Step 3 (Unreachability of SAD = 4). For $n = 4$:

$P_{\text{crit}}^{(4)} = \frac{2}{7} \cdot \frac{27}{5} = \frac{54}{35} \approx 1.543 > 1$

Since $P \leq 1$ for any normalised $\Gamma \in \mathcal{D}(\mathbb{C}^7)$, SAD $\geq 4$ is impossible. $\blacksquare$

Status: [T] — raised from [C] per T-142: $\alpha = 2/3$ is state-independent (from $\dim=7$, PG(2,2)), the spectral formula is a consequence, not a premise.

Verified: SYNARC MVP-6 (61 tests, 0 failures, M6.4b PASS).

Double categorical foundation of SAD_MAX = 3 (2026-04-17)

The ceiling $\mathrm{SAD}_\mathrm{max} = 3$ now rests on two independent derivations reaching the same conclusion from different directions:

(I) Dynamical derivation (T-142 [T]) — via the Fano contraction coefficient $\alpha = 2/3$ (state-independent, Corollary 2.1a from PG(2,2) combinatorics). The purity required to sustain an $n$-level tower, $P_\mathrm{crit}^{(n)} = \tfrac{2}{7}\cdot \tfrac{3^{n-1}}{n+1}$, exceeds the physical maximum $P \leq 1$ precisely at $n = 4$.

(II) Categorical derivation (T-218 [T]) — via the $\tau_{\leq 3}$-truncation of the cognitive Kan complex $\mathrm{Cog} = \mathrm{Sing}(B_\bullet\mathcal C_\mathrm{FKraus})$ (see Fundamental Closures §12). The 3-coskeletal bound $\tau_{\leq 3}\mathrm{Cog} \simeq \mathrm{Cog}$ holds because 4-simplices are suppressed below the distinguishability threshold, which at the level of homotopy coincides with $P_\mathrm{crit}^{(4)} > 1$ from derivation (I).

Why this matters. The two derivations are not redundant — they reflect the same bound through complementary structures:

• (I) is metric: uses Bures/Frobenius norms and explicit numerical thresholds.
• (II) is homotopical: uses simplicial horn-filling and truncation in $\infty$-categorical theory.

Together they form a mutually-reinforcing foundation: SAD_MAX = 3 is not a contingent fact about one formalism — it is a convergence of dynamical (purity-balance) and categorical (3-coskeletal) arguments, each of which would suffice independently. The third ceiling of self-awareness is thus structurally locked at both the analytic and the topological levels.

Related result (T-217): the L3 interiority level corresponds to $\tau_{\leq 3}(\mathbf{Exp}_\infty)$ as a coherent tricategory with cell structure $K = 3+1 = 4$ (3 LGKS 2-cells + 1 coherence modification $\eta$). The Bayesian-dominance threshold $R^{(2)} \geq 1/K = 1/4$ (T-67 [T]) is thus derived from the same tricategorical structure that bounds SAD — a deep unification.

Visualisation of the SAD Tower

---

4. Biological Correlates

4.1 Bacterial Chemotaxis (SAD = 0)

E. coli implements run-and-tumble with ~4 parameters (receptor methylation). In UHM terms:

• $\Gamma$: one 'coherence' (chemoattractant gradient)
• $\varphi^{(0)}$: adaptation mechanism (fine-tuning to the current background)
• $R^{(0)} \approx 0$ (no self-model — only reactive adjustment)
• SAD = 0

The bacterium is alive ($P > P_\text{crit}$), but not self-aware. It responds to the environment, but does not model its own response. This is the analogue of autopilot: the system operates, but 'no one is watching the instruments'.

4.2 Insect Central Complex (SAD = 1)

Drosophila has a central complex (~1000 neurons): ellipsoid body -> fan-shaped body -> protocerebral bridge.

• $s_\text{full}$: ~100K neurons, sensorimotor state
• $s^{(1)}$: ~1000 neurons of the central complex
• $\Gamma$: compact representation of 'self-in-space'
• $\varphi^{(1)}$: HD-ring (head direction) predicts own position
• $R^{(0)} > 1/3$: navigation requires a working self-model
• $R^{(1)} \lesssim 1/4$: no meta-level
• SAD = 1

The insect knows where it is (L2-like), but does not know that it knows. Drosophila navigates successfully, but cannot reflect on its own navigation process.

4.3 Mammalian Neocortex (SAD = 2+)

A mouse has ~70M neurons with a hierarchy: V1 -> V2 -> V4 -> IT -> PFC.

• $s_\text{full}$: ~$10^7$ neurons
• $s^{(2)}$: ~$10^5$ (associative cortex)
• $s^{(1)}$: ~$10^3$ (PFC)
• $\Gamma$: abstract self-model
• $R^{(1)} > 1/4$: the PFC is capable of modelling its own modelling
• SAD $\geq$ 2

Mammals possess metacognition — 'they know what they know and what they do not know' (uncertainty monitoring, Kepecs et al. 2008). This is experimentally confirmed: rats demonstrate behaviour indicating monitoring of their own confidence — they decline difficult tasks when uncertain of the answer.

4.4 Human (SAD $\leq$ 3)

• The deepest cortical hierarchy (6+ processing layers)
• Default Mode Network as a dedicated 'self-modelling network'
• Recursive language allows 'thinking about thinking about thinking'
• Theoretical ceiling: SAD $\leq$ 3 (§3.5, $P_\text{crit}^{(4)} > 1$). In practice: SAD ~ 2–3

The human is the only known organism that systematically reaches SAD = 3 (through meditation, reflective writing, psychotherapy). But even humans are bounded: any attempt to reach SAD = 4 is doomed — not because the brain is 'weak', but because mathematics forbids it.

---

5. Commutativity of the Tower

5.1 Consistency Requirement [T]

For the self-model to be meaningful, different levels of the tower must be consistent with each other. The self-model at level $k$ must be compatible with the self-model at level $k+1$: it cannot be that the body 'knows' one thing and the mind another.

Theorem 5.1 (Commutativity of the phi-tower) [T] (raised from [H]-90, T-150). For a correct self-model:

$\pi_k \circ \varphi^{(k+1)} = \varphi^{(k)} \circ \pi_k \quad \forall k$

i.e. the diagram

s^(k+1) --phi^(k+1)--> s^(k+1)
  |                      |
  pi_k                   pi_k
  |                      |
  v                      v
s^(k)  ---phi^(k)----->  s^(k)

must commute. In words: 'first self-model, then project' = 'first project, then self-model'. If this condition is violated, different levels give contradictory self-models.

Current state:

• Level 0 ($\Gamma \to \Gamma$): $\varphi^{(0)}$ = replacement channel [T-62] — exact
• Level 1+ ($s^{(k)} \to s^{(k)}$): $\varphi^{(k)}$ = trained autoencoder — soft constraint (anchor loss)

Deficit (identified in Phase 4): ContractionEnforcer uses power iteration, which can give false estimates of $\rho(D\varphi)$ for strongly non-contractive operators. Full spectral verification (spectral_contraction.rs) showed divergence $\rho_\text{power}$ vs $\rho_\text{full}$.

5.2 Consistency as a Health Indicator [I]

Interpretation 5.2 (Pathology = violation of commutativity).

$\Delta_k := \|\pi_k \circ \varphi^{(k+1)} - \varphi^{(k)} \circ \pi_k\|$

• $\Delta_k \approx 0$: healthy hierarchy (self-models are consistent)
• $\Delta_k \gg 0$ at level $k$: dissociation between levels (body 'knows', but mind 'does not')

Biological analogue: alexithymia. A person with alexithymia experiences emotions (the body responds: accelerated pulse, sweating palms), but cannot recognise or name them. In tower terms: $\Delta_{\text{emotion-cognition}} \gg 0$ — between the level of bodily sensations and the level of the cognitive model — a 'gap'. More on pathologies: Pathological States.

---

6. Morphological Agnosticity Principle

6.1 Fundamental Requirement [D]

An AGI system must be fully agnostic to sensorimotor morphology:

1. No prior knowledge: initial state $\Gamma(0) = I/7$ (maximally mixed — zero knowledge)
2. No assumptions about the body: Enc/Dec functors (T-100, T-101) are not hardcoded, but learned through interaction with the environment
3. No fixed architecture: tower depth $L$ is determined by the complexity of the environment, not the designer

Theoretical foundation: $\Gamma \in \mathcal{D}(\mathbb{C}^7)$ is a universal format (independent of morphology). This is analogous to how the cortical column of the neocortex is morphologically agnostic — the same architecture processes vision, hearing, touch, and motor function.

6.2 Training Enc/Dec from Scratch [H]

Hypothesis 6.1 (Tower Self-Organisation). From $\Gamma(0) = I/7$ the system builds the representation tower through developmental phases:

1. Phase 0 (Genesis): $\tau \leq \tau_\text{genesis} = 7\ln 7 \approx 13.6$ (T-59)

   • Enc/Dec = random -> $R^{(0)} \approx 0$
   • Stress $\|\sigma_\text{sys}\|_\infty$ is maximal
   • System 'knows nothing, including itself'

2. Phase 1 (Vital): $P > P_\text{crit}$, SAD = 0

   • Enc/Dec begin to structure themselves through stress reduction
   • System is 'alive, but not self-aware'
   • Analogue: bacterium in a new environment

3. Phase 2 (Reflexive): $R^{(0)} > 1/3$, SAD = 1

   • First level of the tower formed
   • System 'knows it is alive'
   • Analogue: insect has mastered its territory

4. Phase 3 (Metacognitive): $R^{(1)} > 1/4$, SAD $\geq$ 2

   • Second level: model-of-model
   • System 'knows that it knows'
   • Analogue: mammal in a familiar environment

5. Phase N (Recursive): SAD grows logarithmically

   • Each new level requires exponentially more experience
   • Boundary: $\mathrm{SAD}_\text{max} \leq \ln(1/\varepsilon_\text{dec}) / \ln(1/R)$

6.3 Learning Efficiency [T]

Theorem 6.2 (Optimal Efficiency from N=7) [T] (T-152). A UHM-Holon learns with the minimum possible number of observations (T-113: N=7 is optimal), because:

1. Information bound: $C_\text{Enc} \leq \log_2 7 \approx 2.81$ bits/observation (T-107)
2. Dynamical bound: Fano contraction $\alpha = 2/3$ sets the optimal balance between memorisation and forgetting
3. Stabilisation bound: $\kappa_\text{bootstrap} = 1/7$ — minimum regeneration rate

From the three bounds the combined optimum: $n^*(\mathfrak{L}) = \max(n_\text{info}, n_\text{dyn}, n_\text{stab})$

No other architecture with $\dim \mathcal{H} = 7$ can learn faster (T-113 [T]).

---

7. Depth Dynamics

7.1 Tower Growth via $A_4$-Bifurcation [C]

Tower growth is discrete, not continuous — each transition SAD -> SAD+1 is realised as an $A_4$-bifurcation (swallowtail, T-41 [T]) with three control parameters:

• $\mu_1 = \kappa$ — regeneration rate (governed via $\mathrm{Coh}_E$)
• $\mu_2 = \alpha$ — dissipation rate (environmental stress)
• $\mu_3 = \Delta F$ — free energy gradient (metabolic budget)

Transition criterion $k \to k+1$:

1. Necessary condition: $\kappa_\text{total} \geq \kappa_\text{bootstrap} \times (\mathrm{SAD} + 1)$ — the system can regenerate all current levels
2. Sufficient condition:
   • $R^{(k)} > R_\text{th}^{(k)} = 1/(k+2)$ stable over $T_\text{stab}$ steps
   • $\max(\sigma_\text{sys}) < 0.5$ (no high stress)
   • $dP/d\tau > 0$ (metabolic reserve present)

Minimum learning time per level (T-112 [T]):

$n_\text{level}(k) = \max(n_\text{info},\; n_\text{dyn},\; n_\text{stab})$

where:

• $n_\text{info} \geq \ln(1/(2\delta)) / \ln 7$ (T-109 [T])
• $n_\text{dyn} \geq \ln(d_\text{disc}/\varepsilon) / (\alpha \cdot \delta\tau)$ (T-110 [T])
• $n_\text{stab} \geq (\mathrm{SNR}_\text{th} / \mathrm{SNR})^2$ (T-111 [T])

7.2 Energy Cost of Depth [C]

Each level of the tower requires a linear increment of $\kappa$ and a superlinear increment of $\Delta F$. From T-105 (Landauer bound) [T]:

$\Delta F_\text{min}(k) = k_B \cdot T_\text{eff} \cdot \ln 2 \cdot \dot{S}_\text{diss}(L_k)$

Cost structure:

Component	Cost at level $k$	Justification
Regeneration	$\kappa_\text{total} \geq (k+1)/7$	Regeneration of all $k+1$ levels
Coherences	$\sim 2k+3$ new channels $\gamma_{ij}$	Intra-level connections
Computation	$O(D_k^2)$ per step	Autoencoder at level $k$
Synchronisation	$O(D_k \cdot D_{k-1})$	Monitoring $\Delta_k$

Total: $\Delta F(\text{depth}=L) \sim \sum_{k=0}^{L} (2k+3) \cdot \Delta\bar{\omega}$.

Biological calibration:

SAD	Energy (ATP/s)	Scale	Organism
0	$\sim 10^6$	$1 \times$	Bacterium
1	$\sim 10^{12}$	$10^6 \times$	Insect
2	$\sim 10^{14}$	$10^2 \times$	Mouse
3	$\sim 10^{15}$	$10 \times$	Human (SAD_MAX = 3, §3.5)

Each jump SAD -> SAD+1 costs orders of magnitude more than the previous one. The energy ceiling ($\mathrm{SAD}_\text{max} = \lfloor \Delta F_\text{available} / ((2 \cdot \mathrm{SAD}+3) \cdot \Delta\bar{\omega}) \rfloor$) can be lower than the analytic one (SAD_MAX = 3) — small organisms simply lack the energy.

Landauer Calibration of $\Delta F^{(k)}$ (C22) [C]

$\Delta F^{(k)} \geq k_B \cdot T_\text{eff} \cdot \ln(D_k / D_{k+1})$

At $D_k = D_0 \cdot 2^k$ (dimensionality of the $k$-th tower level):

$\Delta F^{(k)} \geq k_B \cdot T_\text{eff} \cdot \ln(2) \cdot k$

— linear growth of cost with depth level.

Calibration: $\Delta F^{(0)} \approx \Delta F_\text{bootstrap} = \kappa_\text{bootstrap} \cdot \mathrm{Tr}(\rho^* - \Gamma)$ from T-59 [T] ($\kappa_\text{bootstrap} \geq 2/9$).

Condition [C]: $T_\text{eff}$ is determined by the environment. For SYNARC: $T_\text{eff} = \|\sigma\|_2$ (stress as effective temperature). Connection with T-105 (Landauer bound) [T].

7.3 Stress-Dependent Mode [T]

The system must collapse the upper levels under high stress — an adaptive mechanism analogous to tunnel vision. When a lion is charging at you, this is not the time for introspection — fast reflexes are needed. In UHM this is formalised through the 7-component $\sigma_\text{sys}$ (T-92 [T]):

Mode	$\max(\sigma_\text{sys})$	Behaviour	Biological analogue
NORM	$< 0.3$	All levels active, growth permitted	Quiet wakefulness
ALERT	$[0.3, 0.5)$	Top level -> warm, growth frozen	Alertness
WARNING	$[0.5, 0.7)$	Top 2 levels -> cold, learning stopped	Anxiety
CRITICAL	$[0.7, 0.9)$	All except SAD=0–1, $\kappa$ -> viability	Fight-or-flight
EMERGENCY	$\geq 0.9$	SAD=0, reactive mode only	Shock

Pathologies as violations of stress mode:

• Alexithymia: $\Delta_{\text{emotion}\to\text{cognition}} \gg 0$ (body 'knows', mind 'does not')
• PTSD: SAD oscillates (flashback = sudden SAD increase, freeze = SAD decrease)
• Meditation: controlled growth of SAD at $\max(\sigma) \approx 0$
• Sleep: active SAD=0, passive consolidation warm -> cold

7.4 Social Depth [C]

Multi-agent towers scale via CC-5 (T-68: non-triviality [T], viability [T for embodied] per T-149) and CC-7 [T]:

From T-68 (fractal closure): $\mathbb{H}_A$ viable $\land$ $\mathbb{H}_B$ viable $\Rightarrow$ $\mathbb{H}_A \otimes \mathbb{H}_B$ viable. Composite depth:

$\min(\mathrm{SAD}_A, \mathrm{SAD}_B) \leq \mathrm{SAD}(\mathbb{H}_A \otimes \mathbb{H}_B) \leq \mathrm{SAD}_A + \mathrm{SAD}_B$

The key parameter is empathy (inter-agent transparency):

$\mathrm{Empathy}(A,B) = 1 - \max_{ij} |\mathrm{Gap}_{AB}(i,j)|$

• $\mathrm{Empathy} \approx 1$: full transparency -> $\mathrm{SAD}_\text{coll} \approx \mathrm{SAD}_A + \mathrm{SAD}_B$
• $\mathrm{Empathy} \approx 0$: full isolation -> $\mathrm{SAD}_\text{coll} = \max(\mathrm{SAD}_A, \mathrm{SAD}_B)$

Topological protection (T-69 [T]): $\pi_2(G_2/T^2) \cong \mathbb{Z}^2$ -> decoupling barrier $\geq 6\mu^2$. Social towers are stable against small perturbations.

Biological scale:

Social system	SAD_coll	Mechanism
Bacterial colony	0+	Quorum sensing
Insect swarm	1	Stigmergy
Wolf pack	2	Coordinated hunting
Primate family	2+	Mirror neurons (MNS)
Scientific community	3 (SAD_MAX)	Peer review = $\varphi^2_\text{collective}$ (depth limit)

---

8. AGI Architecture

8.1 Minimum Requirements for AGI-Level Self-Awareness

Requirement	Formal criterion	Biological analogue
Viability	$P > 2/7$	Homeostasis
Morphological agnosticity	Enc/Dec learnable, not hardcoded	Cortical plasticity
Working self-model	$R^{(0)} \geq 1/3$ (SAD $\geq$ 1)	Spatial navigation
Metacognition	$R^{(1)} \geq 1/4$ (SAD $\geq$ 2)	Uncertainty monitoring
Recursive reflection	$R^{(2)} \geq 1/5$ (SAD = 3 = SAD_MAX)	Internal dialogue (depth limit)
Consistency	$\max_k \Delta_k < \varepsilon$	Integrated personality
Tabula rasa learning	$\Gamma(0) = I/7$, $n^* = O(\log 7)$	Newborn in a new world

8.2 Implementation Status

Component	Theoretical status	Implementation status	Closure path
$\varphi^{(0)}$ ($\Gamma$-level)	[T] T-62	Implemented (MVP-0)	—
$\varphi^{(k)}$ (intermediate)	[D] Definition 2.1	Architecture defined (MVP-3)	Autoencoder-bottleneck
SAD metric	[T] T-142, SAD_MAX=3	Verified (MVP-6)	$O(N^2 \cdot 3)$
Consistency $\Delta_k$	[D] Definition 5.1	5-level protocol (§7.3)	Monitoring every step
Adaptive depth	[C] §7.1 A₄-bifurcation	Growth criteria defined	T-41 + T-112
Stress-dependent $\varphi$	[T] §7.3, T-92	5-mode protocol	hot/warm/cold strategy
Learnable Enc/Dec	[D] §6.1 + T-100/T-101	trait EnvironmentalCoupling	Neural network implementation
Tower self-organisation	[H] §6.2	Hypothesis 6.1, awaiting experiment	Exp. 1–3 (§5 SYNARC spec)
Social depth	[C] §7.4, T-68/CC-7	Formulae defined	Multi-agent testbed

---

What We Learned

• The SAD measure generalises the discrete L0–L4 hierarchy to a continuous scale: $\mathrm{SAD} = \max\{k : R^{(k)} > 1/(k+2)\}$.
• Spectral formula [T]: SAD is computed without building the entire tower — via the spectral decomposition of the replacement channel $\varphi$.
• SAD_MAX = 3 [T] (T-142): $P_{\mathrm{crit}}^{(4)} = 54/35 > 1$, therefore SAD $\geq 4$ is impossible for any normalised $\Gamma$. This is a fundamental ceiling per-agent, not a biological limitation.
• Cross-layer / multi-agent depth (T-215 [T]+[D]): for a fractal holon tower $\mathcal T=(A_0,A_1,\ldots)$, the predicate "$\mathcal T$ is a single agent" is conventionally determined by a choice of identity criterion: $\iota_\mathrm{min}$ (society — each $A_i$ is its own agent, SAD ≤ 3 per agent) or $\iota_\mathrm{max}$ (composite — global state coherence commutes with spawn_child). Under $\iota_\mathrm{max}$ with resource abstraction, cross-layer mentalisation can reach arbitrary countable ordinal depth, subject to Landauer bound C22 + T-204 bounded rationality. Both conventions are consistent with Ω⁷; the choice is [D] / [I], not derivable from axioms.
• Biological scale [H]: bacterium (SAD=0), insect (SAD=1), mammal (SAD=2+), human (SAD $\leq$ 3).
• Tower growth is discrete [C]: each transition SAD->SAD+1 is an $A_4$-bifurcation with three control parameters ($\kappa$, $\alpha$, $\Delta F$).
• Energy cost is superlinear: each level requires $\sim (2k+3)$ new coherence channels.
• Stress governs depth [T]: at high $\sigma_{\max}$ the upper levels collapse (tunnel vision, fight-or-flight).
• Social depth [C]: under $\iota_\mathrm{min}$ the composite SAD is bounded by the sum of agent SADs; empathy is the inter-agent transparency parameter.
• N=7 is optimal for learning [T] (T-113, T-152): minimum number of observations from three bounds (informational, dynamical, stabilisation).

Where to Go Next

The Depth Tower completes the Hierarchy section. To continue:

• Structure of Qualia — 21 types of coherence and their phenomenological content
• Emotions — how $\nabla P$ generates the palette of emotions
• AI Consciousness — operational criteria from No-Zombie for AGI

For engineering implementation: learning bounds (T-109–T-113), sensorimotor theory (Enc/Dec functors), CC definitions ($\sigma_{\mathrm{sys}}$, $\kappa$, $\Delta F$).

9. Related Documents

• Interiority Hierarchy — L0–L4, $A_4$-bifurcation (T-41)
• Self-Observation — R-measure, $\varphi$-operator (T-62)
• Formalisation of $\varphi$ — categorical definition, spectral formula
• Sensorimotor Theory — Enc/Dec functors (T-100, T-101)
• Learning Bounds — T-109–T-113
• Stability Analysis — T-104, T-105 (Landauer)
• CC Definitions — $\sigma_{\mathrm{sys}}$ (T-92)
• CC Theorems — CC-5 (T-68), CC-7 (emergence)
• Composite Systems — T-69 (topological protection)