Appendix E-11: Computational Simulation of the Rate-Distortion Lifecycle

This appendix documents the in-silico modeling of the Ordered Patch Theory (OPT) codec lifecycle. Because the underlying universal substrate (the Solomonoff semimeasure) is structurally uncomputable, simulations within the OPT framework are restricted to modeling the codec lifecycle itself: the boundary gating parameter C_{\max}, active inference dynamics, the three-pass maintenance cycle \mathcal{M}_\tau, and narrative decay under entropic stress.

Two distinct simulation paradigms have been established: analogical deep learning (toy_model.py) and strict mathematical rate-distortion modeling (opt_simulator.py).

1. Analogical Simulation: Deep Variational Bottlenecks

The initial simulation paradigm (toy_model.py) validates the core premise of Codec Fracture using a literal structural analogy.

Substrate: A 1D periodic lattice instantiated with discrete integers. Persistent structural features are injected against a baseline of thermodynamic noise, functioning as the observable “Ordered Patches.”

Architecture: The observer is modeled as a Variational Information Bottleneck (VIB) built atop a deep neural network (TensorFlow). The network observes a spatial history vector X_{t-k \dots t} and performs a forward gradient descent to compress it into a bottleneck capable of predicting the forward temporal fan X_{t+1 \dots t+h}.

Mechanics of Collapse: The C_{\max} (rate) and D_{\min} (acceptable distortion) constraints are enforced dynamically via a PID controller modulating the Lagrangian \beta multiplier. Under massive substrate entropy (e.g., highly volatile noise dominating the persistent patterns), the network physically trades predictive resolution for bandwidth. When the required algorithmic complexity R_{\text{req}} exceeds C_{\max} despite maximal \beta tuning, the network formally hits an algorithmic singularity and collapses, confirming the OPT prediction that injecting high-entropy noise destroys predictive coherence rather than “expanding” consciousness.

2. Mathematical Formalism: Strict Rate-Distortion Modeling

While the neural VIB provides visual confirmation of codec fracture, the overhead of machine learning architectures obscures the pure information-theoretic relationships governing the observer. The second paradigm (opt_simulator.py) strips away structural geometry to strictly model the bottleneck dynamics using the theory’s own scalars.

2.1 Architecture

The simulator separates three structural layers, mirroring the OPT formalism:

The key design principle is that predictability is endogenous: the codec’s ability to predict the environment is derived from C_{\text{state}} via a power-law relationship \text{error} \propto C_{\text{state}}^{-0.6}, rather than being a hardcoded parameter. This means fracture cascades and recovery trajectories emerge from the system’s own dynamics rather than being manually imposed.

2.2 The Prediction Error Channel

Under predictive rate-distortion theory, what crosses the C_{\max} aperture is the prediction error — only the residual after the generative model’s prediction is subtracted:

\varepsilon_t = S_{\text{raw}} \cdot (1 - \text{predictability})

where S_{\text{raw}} = 10^9 \cdot \Delta t bits per update window. At baseline (C_{\text{state}} \approx 10^{14}, stationarity = 1.0), this yields \varepsilon_t \approx 0.16 bits/step — comfortably below the capacity bound of C_{\max} \cdot \Delta t = 0.5 bits/step.

When environmental stationarity drops (e.g., ketamine shock, stationarity \to 0.1), the effective prediction error is amplified by a factor of 1/\text{stationarity}, driving \varepsilon_t above the capacity bound and triggering fracture.

2.3 The Three-Pass Maintenance Cycle (\mathcal{M}_\tau)

The maintenance cycle implements the three offline passes specified in §3.6 of the preprint:

Net drain per maintenance run: \sim 6\% of C_{\text{state}}. Maintenance is gated on stability — it fires only when the codec is not fractured, consistent with OPT’s prediction that \mathcal{M}_\tau runs during low-sensorium states (paradigmatically: sleep).

The learning accumulation rate is calibrated so that the error-integration gain over 100 inter-maintenance steps approximately equals the 6% maintenance drain, producing dynamic equilibrium at baseline.

2.4 Fracture Dynamics

Narrative decay is modeled as gentle multiplicative degradation with a hard floor:

C_{\text{state}}(t+1) = \max\bigl(C_{\text{state}}(t) \cdot 0.9999,\; C_{\text{floor}}\bigr)

Over 400 sustained fracture steps (a 20-second shock), this compounds to 0.9999^{400} \approx 0.961 — approximately 4% loss. This models graded phenomenological blanking (as in anesthesia titration, Protocol E-9) rather than catastrophic all-or-nothing collapse.

2.5 Simulation Results

The simulator runs 2000 cycles at \Delta t = 50\text{ms} resolution (100 seconds of simulated observer-time). An entropy shock (stationarity \to 0.1) is applied from t=40\text{s} to t=60\text{s}.

These three phases demonstrate the core OPT prediction: a bounded observer can maintain stable homeostasis, degrade gracefully under entropic shock, and recover when environmental stationarity is restored — provided the shock does not drive C_{\text{state}} below C_{\text{floor}}.

2.6 Key Observations

1. The baseline sawtooth: Between maintenance runs, C_{\text{state}} accumulates via error integration (\sim +5\% per 100-step window), then drops sharply when \mathcal{M}_\tau fires (\sim -6\%). This oscillation is the computational signature of the sleep-wake cycle — the system must periodically prune to avoid hitting C_{\text{ceil}}.

2. Shock onset is instantaneous: When stationarity drops to 0.1, every cycle immediately fractures. There is no gradual transition — the prediction error jumps from \sim 0.16 to \sim 1.6 bits/step, exceeding the 0.5 bit capacity by a factor of three.

3. Recovery is asymmetric: Post-shock C_{\text{state}} grows at \sim +1\% over 40 seconds, compared to the \sim -4\% loss during the 20-second shock. Recovery is slower than degradation. This asymmetry is a structural prediction of OPT: rebuilding a generative model is harder than damaging one.

4. The maintenance-fracture gate matters: If maintenance runs during active fracture (as in early simulator versions), the system enters a positive feedback loop and collapses to C_{\text{floor}}. The gating rule is not a convenience — it is structurally necessary for codec viability.

3. Future Simulation Pathways

1. Thalamocortical Clocks (E-12): Hardcoding \Delta t updates to match the 20–40\text{Hz} thalamic gating cycles, generating testable millisecond-resolution predictions against cortical integrated information (\Phi) measurements.

2. Free Energy POMDP Integration: Replacing the abstract predictability scalar with a discrete Active Inference state-space model (e.g., pymdp), allowing mapping of the precise bounds separating thermodynamic thermostats from the phenomenal K_{\text{threshold}} (P-5).

3. Multi-Observer Extension: Simulating multiple interacting codecs with shared substrate regions to test the Swarm Binding predictions of Appendix E-6 — whether distributed agents achieve phenomenal binding only when forced through a global C_{\max} aperture.

4. Empirical Calibration: Fitting the simulator’s fracture-recovery trajectory against neuroimaging time-series data (e.g., Lempel-Ziv complexity under propofol or ketamine) to determine whether the 0.9999 decay constant and C_{\text{state}}^{-0.6} predictability curve match observed phenomenological dynamics.