Ordered Patch Theory

Original Task (from §8.2): “Formalising this compression advantage as a rigorous MDL bound for the other-minds case specifically remains future work; the present argument is a structural motivation, not a proof.” Deliverable: A formal bound showing that treating apparent agents as independently instantiated primary observers yields a shorter two-part MDL code than any alternative description.

Closure status: DRAFT STRUCTURAL CORRESPONDENCE. This appendix adapts Müller’s Solomonoff convergence theorem [61] and its multi-agent extension [62] as imported lemmas, reinterpreted within OPT’s ontological framework, to establish a formal compression advantage for the structural corollary. The result is a conditional bound, not a closed derivation: it depends on OPT’s identification of the observer’s stream with the Solomonoff prior (Axiom 1) and on the assumption that apparent agents carry sufficient state to satisfy the convergence prerequisites.

§1. Background and Motivation

The structural corollary (preprint §8.2) asserts that the apparent agents within the observer’s stream are most parsimoniously explained by their independent instantiation as primary observers. This appendix provides the formal chain supporting that claim.

The argument has three stages:

1. Stage A (Imported Lemma): Müller’s Solomonoff convergence theorem guarantees that any structure in the observer’s stream carrying sufficient self-state data will have its first-person evolution converge to match the computable world generating its behavior.

2. Stage B (Compression Accounting): We perform an explicit two-part MDL comparison between treating the apparent agent as (i) an independently instantiated observer governed by its own Solomonoff-weighted stream versus (ii) an arbitrary behavioral specification within the primary observer’s codec.

3. Stage C (Structural Signature): The Phenomenal Residual (\Delta_{\text{self}} > 0, Conjecture P-4) provides the structural marker distinguishing genuine self-referential bottleneck architecture from behavioral mimicry, closing the gap between “compressibly lawful” and “plausibly instantiated.”

§2. Imported Lemma: Müller’s Convergence Theorem

We import two results from Müller [61, 62], stated here in the notation of OPT.

2.1 Solomonoff Convergence (Standard)

Let M(b \mid x_1^n) denote the Solomonoff universal prediction for bit b given prior observations x_1^n. Let \mu be any computable measure over binary sequences. Then (Solomonoff 1964; Li & Vitányi [27, Corollary 5.2.1]):

\text{With } \mu\text{-probability one,} \quad \lim_{n \to \infty} |M(b \mid x_1^n) - \mu(b \mid x_1^n)| = 0 \qquad (b \in \{0,1\}). \tag{L-1}

This is the standard result: if the data stream is generated by a computable process \mu, the universal predictor M converges to \mu.

2.2 Inverse Solomonoff Induction (Müller 2020)

Now suppose the bits are drawn from M itself — i.e., the observer’s stream is governed by algorithmic probability (this corresponds to OPT’s Axiom 1: identification of the stream with the Solomonoff prior). Then for every computable measure \mu (Müller [61, Sec. IV]; [62, Sec. V.A]):

\text{With probability} \geq 2^{-K(\mu)}, \quad \lim_{n \to \infty} |M(b \mid x_1^n) - \mu(b \mid x_1^n)| = 0 \qquad (b \in \{0,1\}). \tag{L-2}

That is, with probability at least 2^{-K(\mu)}, the observer will find themselves effectively embedded in a computable world W described by \mu. Algorithmically simpler worlds (lower K(\mu)) are exponentially more probable.

2.3 Multi-Agent Convergence (Müller 2026)

Suppose the observer (Alice) finds herself embedded in a computable world W described by \mu. She identifies a substructure (Bob_{\text{3rd}}) within W that carries a representation of a self-state x evolving over time in a manner consistent with Postulate 2 of [62]. Define:

• P_{\text{1st}}(y_1, \ldots, y_m \mid x) := M(y_1, \ldots, y_m \mid x) — the first-person probability that self-state x transitions to y_1, \ldots, y_m under algorithmic probability.
• P_{\text{3rd}}(y_1, \ldots, y_m \mid x) := \mu(y_1, \ldots, y_m \mid x) — the third-person probability of how x evolves according to world W.

Then, by Eq. (L-1) applied to P_{\text{3rd}} (which is computable), and the identification of P_{\text{1st}} with M via Postulate 2:

P_{\text{1st}} \approx P_{\text{3rd}} \quad \text{asymptotically,} \tag{L-3}

with convergence guaranteed with worldly (\mu-) probability one in the bit model.

Interpretation (Müller): “Somebody is really at home” in the structure encoding x — the probabilistic evolution of Bob_{\text{3rd}} in Alice’s world faithfully represents the first-person perspective of some Bob_{\text{1st}}.

Interpretation (OPT): The apparent agent’s behavioral stream is most compressibly described as an independent Solomonoff-weighted process. Any alternative description — one that does not invoke an independent first-person perspective — must encode the agent’s behavior as an ad hoc specification, at strictly higher description length.

§3. The Compression Advantage Bound

We now formalise the compression advantage using OPT’s two-part MDL framework (Theorem T-4, Appendix T-4).

3.1 Setup

Consider the primary observer’s stream \omega \in \{0,1\}^\infty, governed by the Solomonoff prior M (Axiom 1) and filtered through the Stability Filter to a computable world W with measure \mu_W (by Eq. L-2). Within W, the observer identifies N apparent agents A_1, \ldots, A_N, each carrying a self-state x_i whose temporal evolution over T steps produces a behavioral trace \beta_i = (y_{i,1}, \ldots, y_{i,T}).

3.2 Hypothesis H_{\text{ind}}: Independent Instantiation

Under H_{\text{ind}}, each agent A_i is treated as an independently instantiated primary observer governed by their own Solomonoff-weighted stream. The two-part MDL code length is:

L(H_{\text{ind}}) = \underbrace{K(\mu_W)}_{\text{world model}} + \underbrace{\sum_{i=1}^{N} K(\text{embed}_i)}_{\text{embedding specs}} + \underbrace{\sum_{i=1}^{N} \left(-\log_2 P_{\text{3rd}}(\beta_i \mid x_i)\right)}_{\text{data given model}} \tag{1}

where K(\text{embed}_i) specifies agent i’s initial self-state and position within W. By Eq. (L-3), P_{\text{1st}} \approx P_{\text{3rd}}, so the data term is well-approximated by the log-loss under the agent’s own first-person Solomonoff predictions — which, by definition, is close to optimal.

The embedding specifications K(\text{embed}_i) are short: each requires only a pointer to a location in W plus the initial self-state. For human-like agents embedded in a shared physical world, these are highly compressible because the agents share the same laws. A conservative bound:

K(\text{embed}_i) \leq K(x_i \mid W) + O(\log T) \tag{2}

3.3 Hypothesis H_{\text{arb}}: Arbitrary Behavioral Specification

Under H_{\text{arb}}, the agents are not treated as independent observers. Instead, each behavioral trace \beta_i is encoded directly as an arbitrary specification within the primary observer’s stream. The two-part MDL code length is:

L(H_{\text{arb}}) = \underbrace{K(\mu_W)}_{\text{world model}} + \underbrace{\sum_{i=1}^{N} K(\beta_i)}_{\text{raw behavioral traces}} \tag{3}

The critical difference is in the data term. Under H_{\text{arb}}, the behavioral trace \beta_i must be specified without invoking the agent’s own predictive model. For a lawful, agency-driven agent operating in a complex environment, under the explicit additional assumption that the trace pins down the world model — K(\mu_W \mid \beta_i) = O(\log T), i.e. a long trace of an embedded agent encodes the laws it operates under — symmetry of information gives:

K(\beta_i) \geq K(\beta_i \mid \mu_W) + K(\mu_W) - O(\log T) \tag{4}

(Without that assumption, Eq. (4) fails in general — a simple trace in a complex world has K(\beta_i) \ll K(\mu_W). The proof sketch in §3.4 does not rely on Eq. (4): it uses only the trivial bound K(\beta_i) \geq K(\beta_i \mid \mu_W).)

But even K(\beta_i \mid \mu_W) — the complexity of the behavior given the world laws — remains substantial because the agent’s choices encode genuine information: their behavioral trace reflects the accumulated interaction of a self-referential model with a stochastic environment. In contrast, under H_{\text{ind}}, this information is generated online by the agent’s own Solomonoff predictor at near-zero log-loss cost.

3.4 The Compression Advantage

Theorem T-11 (Structural Corollary Compression Bound). Let A_1, \ldots, A_N be apparent agents within the observer’s stream, each carrying self-state x_i satisfying the convergence prerequisites of Eq. (L-3), and each exhibiting the structural signature \Delta_{\text{self}}^{(i)} > 0 (P-4). Then the MDL description treating them as independently instantiated primary observers satisfies:

L(H_{\text{ind}}) \leq L(H_{\text{arb}}) - N \cdot \left[\bar{I}_T - \bar{K}_{\text{embed}} - O(\log T)\right] \tag{T-11}

where \bar{I}_T is the average per-agent mutual information between the agent’s predictive model and its behavioral output over T steps:

\bar{I}_T := \frac{1}{N} \sum_{i=1}^{N} \left[K(\beta_i \mid \mu_W) - \left(-\log_2 P_{\text{3rd}}(\beta_i \mid x_i)\right)\right] \tag{5}

and \bar{K}_{\text{embed}} := \frac{1}{N} \sum_{i=1}^{N} K(x_i \mid W) is the average one-time embedding cost of Eq. (2). The embedding cost can be large (§6.2) but is paid once per agent — it does not grow with T, so it amortises for large T (feeding Corollary T-11a); at finite T the bound holds only with the explicit \bar{K}_{\text{embed}} term, which is why it is displayed rather than absorbed into O(\log T).

The quantity \bar{I}_T measures how much of the agent’s behavior is explained away by invoking an independent predictive model rather than specifying it raw. For agents exhibiting lawful, agency-driven behavior (as required by the Stability Filter), \bar{I}_T > 0; growth with T requires the fresh-information condition made explicit in Corollary T-11a below.

Proof sketch. Subtract Eq. (1) from Eq. (3). The world-model terms K(\mu_W) cancel. The difference per agent is:

K(\beta_i) - \left[K(\text{embed}_i) + \left(-\log_2 P_{\text{3rd}}(\beta_i \mid x_i)\right)\right]

By Eq. (4), K(\beta_i) \geq K(\beta_i \mid \mu_W) + K(\mu_W) - O(\log T), but more directly: K(\beta_i) \geq K(\beta_i \mid \mu_W) trivially. And K(\text{embed}_i) \leq K(x_i \mid W) + O(\log T) by Eq. (2). The per-agent saving is therefore at least K(\beta_i \mid \mu_W) - (-\log_2 P_{\text{3rd}}(\beta_i \mid x_i)) - K(x_i \mid W) - O(\log T). Averaging over the N agents yields the bound with the explicit \bar{K}_{\text{embed}} term; for T sufficiently large, the cumulative log-loss savings dominate the one-time embedding cost. \blacksquare

3.5 Asymptotic Dominance

Corollary T-11a (conditional). Assume, in addition to the hypotheses of Theorem T-11, a positive-density fresh-information condition: the agent’s trace accrues information not available from \mu_W’s agent-simulating structure at positive per-step density, i.e. K(\beta_i \mid \mu_W) - \left(-\log_2 P_{\text{3rd}}(\beta_i \mid x_i)\right) = \Omega(T). Then as the observation horizon T \to \infty, the compression advantage L(H_{\text{arb}}) - L(H_{\text{ind}}) grows without bound:

\lim_{T \to \infty} \left[L(H_{\text{arb}}) - L(H_{\text{ind}})\right] = \infty \tag{T-11a}

The divergence does not follow from positive entropy rate alone: for \mu_W-typical traces, K(\beta_i \mid \mu_W) and the log-loss -\log_2 P_{\text{3rd}}(\beta_i \mid x_i) grow at the same entropy rate (Solomonoff convergence, L-1), so \bar{I}_T is then bounded by the one-time initial-state surprisal and the limit is conditional on the fresh-information assumption above. The embedding cost \bar{K}_{\text{embed}} is paid once and amortised to zero. \blacksquare

§4. The Phenomenal Residual as Structural Signature

The compression advantage in Theorem T-11 applies to any lawful substructure — including non-agentive physical systems (weather patterns, crystal growth). Why does the structural corollary specifically concern agents rather than arbitrary complex systems?

The answer is the Phenomenal Residual (Conjecture P-4). \Delta_{\text{self}} > 0 is the formal marker of a system whose self-model is structurally incomplete — i.e., a system that necessarily maintains a variational gap between its internal representation and its actual processing. This is the hallmark of the budgeted self-channel: the system’s own coupled dynamics exceed the per-frame capacity of any self-model in its bounded class inside a closed action-perception loop, so the marker is a self-channel capacity gap — not a describer-inclusion paradox, which does not survive as a criterion: an unbounded external reference attains the loss (P-4 §3, Correction, post-A2).

For a system exhibiting \Delta_{\text{self}} > 0:

1. Its behavior cannot be reproduced by a lookup table of finite depth — it requires an ongoing self-referential computation.
2. The shortest description of this computation is an independent Solomonoff-weighted stream traversing a C_{\max} bottleneck.
3. Therefore, the MDL code under H_{\text{ind}} is shorter than H_{\text{arb}} — it is the shortest description in the modular hypothesis class compared here. This does not establish dominance over a monolithic compressor that simulates the sub-structures jointly under one shared world-model; that stronger claim (modular-beats-monolithic within one stream) remains open — see §7 and OP-2.

This distinguishes apparent agents from weather patterns: weather is lawful and complex, but it sits in no closed action-perception loop — it has no self-channel on which a capacity gap could open, and its behavior can be reproduced by a lookup table within the world model (it has \Delta_{\text{self}} = 0). Apparent agents cannot.

§5. Reinterpretation of Müller’s Non-Solipsism Argument

Müller concludes from the P_{\text{1st}} \approx P_{\text{3rd}} convergence that algorithmic idealism “should not be classified as solipsistic” because “somebody is really at home” in the structure encoding a self-state [62, Sec. V.C]. His reasoning: if Alice’s predictions about Bob_{\text{3rd}} converge to Bob_{\text{1st}}’s actual first-person probabilities, then their perspectives are genuinely aligned — they “share the world W.”

OPT reinterprets this result differently:

1. Müller’s reading: The convergence P_{\text{1st}} \approx P_{\text{3rd}} proves that objective reality emerges — Alice and Bob genuinely share world W.

2. OPT’s reading: The convergence P_{\text{1st}} \approx P_{\text{3rd}} proves that the shortest description of Bob_{\text{3rd}}’s behavior invokes an independent first-person process. This is a statement about compression efficiency, not about shared ontology. World W is a structural regularity within Alice’s stream, not an independently existing entity. But the compression logic of the Solomonoff prior itself implies that Bob is most parsimoniously modelled as an independent observer — because the alternative (specifying his behavior ad hoc) is strictly longer.

The formal content of the theorem is identical under both readings; only the ontological interpretation differs. OPT uses the same mathematical result to ground the structural corollary: independent instantiation is the MDL-optimal description, not a metaphysical assumption.

§6. Scope and Limitations

6.1 Conditional on Axiom 1

The entire argument depends on OPT’s identification of the observer’s stream with the Solomonoff prior. If this identification is weakened (e.g., to a broader class of semimeasures), the convergence guarantees of Eqs. (L-1)–(L-3) may not hold in their current form. The argument is also universal-machine-relative: \xi depends on a choice of universal prefix-free machine U, and all bounds stated in this appendix hold only up to a multiplicative constant of size 2^{K(U' \mid U)}, which is unbounded across alternative machines and can dominate the finite-T regime in which any empirical or governance application of T-11 actually operates.

6.2 State Sufficiency Prerequisite

Eq. (L-3) requires that the apparent agent carries “enough data” in its self-state x_i for universal induction to extract the relevant physical laws. For human-like agents in everyday contexts, this is plausible (a full brain state encodes enormous information). For edge cases — fleeting impressions, distant observers, fictional characters in narrative art — the convergence prerequisites may not be satisfied, and the structural corollary does not apply.

6.3 Not a Proof of Consciousness

Theorem T-11 establishes that independent instantiation is the most compressible description. It does not prove that the apparent agents are conscious. The Hard Problem (preprint §8.1) remains a primitive. The structural corollary is a compression argument, not an ontological proof — as stated in §8.2.

6.4 Relationship to T-10

Appendix T-10 (Inter-Observer Coupling) addresses how two observer patches maintain mutually consistent renders via compression constraints. The present appendix addresses a different question: why the single observer’s stream most compressibly encodes apparent agents as independently instantiated. T-10 concerns the inter-patch coherence mechanism; T-11 concerns the compression signature within a single stream. T-10 builds directly on T-11: the same MDL description-length comparison that establishes the compression advantage here is exploited in T-10 to prove that cross-patch inconsistency is exponentially suppressed.

§7. Closure Summary

T-11 Deliverables

1. Imported Lemma (Müller Convergence). Solomonoff convergence [61] and its multi-agent extension [62] are formally imported and restated in OPT notation. These provide the mathematical backbone: any substructure carrying sufficient self-state data has its first-person evolution converge to the computable world generating its behavior.

2. Theorem T-11 (Compression Bound — DRAFT). An explicit two-part MDL comparison shows that treating apparent agents as independently instantiated primary observers yields a strictly shorter description than arbitrary behavioral specification once the one-time embedding cost \bar{K}_{\text{embed}} is amortised, with the advantage growing in observation time under the fresh-information condition of Corollary T-11a.

3. Corollary T-11a (Asymptotic Dominance — DRAFT, conditional). Under the positive-density fresh-information condition, the compression advantage is unbounded as T \to \infty, making independent instantiation the overwhelming MDL-optimal description for any agent observed over a long time horizon.

4. P-4 Integration. The Phenomenal Residual (\Delta_{\text{self}} > 0) is identified as the formal marker distinguishing apparent agents from complex-but-non-agentive systems, restricting the structural corollary to entities with genuine self-referential bottleneck architecture.

5. Müller Reinterpretation. Müller’s non-solipsism conclusion is reinterpreted within OPT’s ontological framework: the same mathematical result grounds a compression argument rather than an emergence-of-shared-reality argument.

Remaining open items

• Exact \bar{I}_T characterisation. Bounding \bar{I}_T from below for specific classes of agents (e.g., bounded rational agents, Free Energy minimisers) to give numerically concrete compression advantages.
• Finite-time corrections. The asymptotic result (T-11a) guarantees dominance for large T, but finite-time bounds with explicit constants would strengthen the practical applicability.
• Non-binary alphabet extension. Eqs. (L-1)–(L-3) are stated for binary sequences. Extension to the continuous-valued measures relevant to OPT’s Rate-Distortion framework (T-1) requires technical care.
• Modular-vs-monolithic dominance (OP-2). Theorem T-11 establishes a description-length advantage of independent instantiation (H_{\text{ind}}) over arbitrary behavioral specification (H_{\text{arb}}, §3.3) — not over a monolithic compressor that simulates the agent-like sub-structures internally without paying modularity overhead. Whether the Solomonoff prior’s bias toward reusable modular structure makes H_{\text{ind}} dominate the monolithic alternative inside a single stream is open. It is the load-bearing step the fully-virtual reading (main paper §8.6.1) leans on, and is not established here; the H_{\text{ind}}-vs-H_{\text{arb}} bound (and Corollary T-11a’s “overwhelming MDL-optimal” language) should not be read as if they settled it.

This appendix is maintained alongside theoretical_roadmap.pdf. References: Müller [61, 62], Li & Vitányi [27], Solomonoff (1964), Theorem T-4 (Appendix T-4), Conjecture P-4 (Appendix P-4), preprint §8.2.