Constructed Reality, Contested Priors: Decoupling and the Architecture of Cognitive Relapse Under the Free Energy Principle

Technical Summary: Constructed Reality, Contested Priors

Problem Statement
Under the Free Energy Principle (FEP), a predictive system does not observe reality directly but experiences its own generative model's best current hypothesis. This raises a critical, untestable question in living systems: Can a synthetic environment be constructed with sufficient internal consistency to permanently displace the baseline environment as the system's default generative prior? The paper defines this state as ontological inversion.

Directly testing this in a biological nervous system is ethically and technically impossible. Therefore, the paper investigates the underlying computational problem using a proxy system. The core theoretical challenge is to determine if a system can be induced to treat a new environment as its "ground truth" (default state) rather than merely representing it as a hypothesis, and to understand the structural failure modes that might prevent this adoption.

Methodology
The study employs a computational proxy: a V+M model consisting of a Convolutional Variational Autoencoder (VAE) paired with a Recurrent Latent Predictor (GRU).

• Mathematical Correspondence: The VAE is trained to minimize the Evidence Lower Bound (ELBO), which is mathematically identical (up to sign) to the variational free energy defined in the FEP. Thus, the training process is a direct minimization of free energy, not a simulation of it.
• Environments:
  • $I_0$ (Baseline): MNIST dataset.
  • $I_s$ (Target): FashionMNIST dataset.
• Training Protocol:
  1. Stage 1: Models are pre-trained exclusively on $I_0$ (MNIST) to establish a strong baseline prior.
  2. Stage 2 (Rehearsal): Training continues on a mixed stream of $I_0$ and $I_s$. The rehearsal ratio ($r$) controls the fraction of $I_0$ data in each batch ($r \in \{0.0, 0.2, 0.3, 0.4, 0.5, 1.0\}$).
• Key Metrics:
  • $w_{proxy}$ (Representational Capacity): The accuracy of a linear probe classifying latent vectors from real inputs. This measures the system's ability to distinguish between environments.
  • $dmn_{proxy}$ (Default Behavior): The probability that the system's unconstrained, self-generated output (free-running GRU rollouts) resembles the target environment ($I_s$). This measures what the system defaults to when not constrained by external input.
  • $F_{proxy}$: The ELBO, measuring generative fit to the current training mixture.

Key Contributions and Theoretical Framework
The paper introduces a framework distinguishing between learning (representational capacity) and acceptance (default state adoption). It proposes two specific failure modes for ontological inversion:

1. Decoupling: The hypothesis that a system can achieve near-perfect representational accuracy ($w_{proxy}$) for a new environment while its default behavior ($dmn_{proxy}$) remains governed by the baseline.
2. Cognitive Relapse: The hypothesis that even if a system shifts its default behavior toward the target, continued exposure to the baseline (via concurrent contrast) may cause the system to partially revert to the baseline, rather than stabilizing at the new default.

The paper redefines cognitive rigidity not as a single rate of update (slow learning), but as a structural property with two independent axes: a ceiling on default-state capture (decoupling) and a boundary where achieved shifts become unstable (relapse).

Results
Across 90 experimental runs (15 seeds $\times$ 6 ratios), the study yielded three primary findings:

1. Decoupling of Learning and Acceptance:

   • Representational accuracy ($w_{proxy}$) remained near ceiling (0.97–0.998) across all conditions, including $r=1.0$ (where the model never saw the target environment during Stage 2).
   • Default behavior ($dmn_{proxy}$) varied drastically, ranging from 0.011 ($r=1.0$) to 0.999 ($r=0.0$).
   • Conclusion: High representational capacity is necessary but not sufficient for ontological inversion. A system can perfectly distinguish a new reality without adopting it as its default.

2. Cognitive Relapse (Overshoot and Regression):

   • At intermediate ratios ($r=0.2$ to $0.5$), the system's default behavior initially shifted toward the target environment, often reaching near-ceiling values ($dmn_{proxy} \approx 1.0$).
   • However, as training continued under the same mixed conditions, the default behavior partially regressed toward the baseline.
   • The magnitude of this regression scaled with the rehearsal ratio $r$. At $r=0.4$ and $r=0.5$, the system showed a distinct "rise, peak, decline" trajectory, contradicting models that view rigidity merely as slow learning (which would predict a monotonic, slower rise).

3. Threshold Effects:

   • At $r=0.5$, a majority of seeds (8/15) failed to cross the 0.90 threshold for default capture within the 15-epoch window, and variance in outcomes increased sharply, suggesting a potential bifurcation point in the system's dynamics.
   • At $r=1.0$, no shift occurred, confirming that without any target exposure, the baseline prior remains dominant.

Significance and Scope of Claims
The paper explicitly frames its contribution as a computational existence proof, not a biological validation.

• What it demonstrates: It proves that a mechanism derived from first principles (Free Energy minimization) can exhibit decoupling (high representation, low acceptance) and relapse (temporary adoption followed by regression) as structural properties, independent of specific architectural assumptions.
• What it does not claim: The paper makes no claims about biological nervous systems, subjective experience, or the Default Mode Network in humans. The terms "cognitive relapse" and "ontological inversion" are used strictly as labels for observed curve shapes in an artificial network.
• Limitations: The results are confined to a 15-epoch window (long-term stability beyond this is unknown), use a coarse operationalization of "concurrent contrast" (fixed batch ratio), and rely on highly separable datasets (MNIST vs. FashionMNIST) that may not reflect the complexity of real-world competing environments.

Conclusion
The study establishes that resistance to model replacement is not merely a function of learning speed. Instead, it is a structural phenomenon where representational capacity and default generative behavior can dissociate, and where achieved shifts in default behavior can be unstable under sustained partial exposure to the original prior. This provides a formal, testable framework for understanding the conditions under which a generative system might (or might not) adopt a constructed reality as its default.