Peter David Fagan

The Core Hypothesis

While often treated as abstract, substrate-independent algorithmic properties, intelligence and computation are ultimately physical processes constrained by conservation laws. The core hypothesis of this framework is that information processing emerges when open systems undergo irreversible transitions, carving out stable macroscopic states from underlying reversible micro-dynamics. Under this physical lens:

• Intelligence (χ) measures the physical efficiency of goal-directed computation—specifically, the amount of useful causal work extracted per unit of irreversible (distinction-destroying) information processing.
• Consciousness (κ) is operationalized as the physical efficiency of internal organization—specifically, the amount of goal-directed work supported per unit of preserved (reversible) internal informational structure.

Grounding Information in Physical Substrates (CCE)

To measure information physically rather than abstractly, we utilize the Conservation-Congruent Encoding (CCE) framework. Traditional information theory (e.g., Shannon entropy) measures logical uncertainty, but is blind to equilibrium asymmetries and reservoir exchanges. CCE grounds information in matter by defining a macroscopic distinction as a family of protected regions (metastable basins of attraction) in state space whose boundaries are stabilized by conservation laws.

Under CCE, informational state changes are mapped directly to physical fluxes:

• Irreversible Computation (I_irr): Many-to-one logical operations (such as resetting a register or merging states) compress macroscopic support. This collapse of distinctions forces a proportional export of entropy—measured in nats—through specific conserved channels (e.g., heat dissipation in a thermal bath, satisfying Landauer-type bounds).
• Reversible Manipulation (I_rev): One-to-one logical operations transport or advect the state within protected world-tubes without crossing boundaries or destroying distinctions. This preserves internal structure, incurring zero thermodynamic penalty in the quasistatic limit.
• Controlled Expansion: One-to-many retrodictive operations reopen macroscopic support, importing conjugate conserved loads (e.g., absorbing heat) to re-instantiate the branch variables.

From Physical Information to the Core Identity

By defining both intelligence (χ) and consciousness (κ) as ratios of goal-directed work (W) to their respective informational currencies (I_irr and I_rev), we isolate how an agent utilizes physical resources to influence its environment.

Solving the consciousness relation for work yields W = κ I_rev. Substituting this into the definition of intelligence gives the identity:

This identity shows that an agent's operational intelligence is the product of its structural consciousness and its processing efficiency (the ratio of preserved to destroyed distinctions). A system can achieve high intelligence either by executing massive, energy-intensive irreversible computations (high I_irr) or by relying on complex, pre-compiled internal models that advect information reversibly (high κ and I_rev).

AI Safety

AI safety is traditionally treated as a normative alignment problem. However, developing a formal framework to measure intelligence shifts the focus to physical boundaries and resource constraints. By defining intelligence as a measurable capacity to perform goal-directed work, we can establish the physical limits within which an agent operates.

Safety constraints can be studied mathematically via boundary dynamics and symbiotic coupling rather than isolated value alignment.

The Fallacy of the Detached Observer

In classical physics and computational theory, observers frequently treat macroscopic equations of motion as absolute and decoupled from the microscopic substrates they describe. This assumption—which we call the Platonic Observer Fallacy—imagines that coarse-grained states and continuous field variables are objective entities that can be measured, tracked, and simulated without physical cost. By neglecting that observation is a physical act bounded by conservation laws, standard modeling hides a growing informational deficit.

1. Information Accounting: Shannon vs. CCE

The difference between Shannon's formulation and the Conservation-Congruent Encoding (CCE) framework lies in the distinction between pure logical accounting and dynamical accounting.

Under Shannon's formulation, information is logical and substrate-independent. For a discrete random variable X representing symbolic states, the Shannon entropy measures uncertainty purely from the state probabilities:

If a bit transitions from an initial distribution p to a final distribution p_{\text{final}}, Shannon's logical accounting reports the information change purely as the difference in logical entropy:

This accounting does not care about the physical characteristics of the underlying states, making it symmetric and blind to the physical effort required to maintain, transition, or merge them.

CCE, by contrast, relies on dynamical accounting. It grounds information in matter by mapping logical states to protected phase-space volumes stabilized by conservation laws. Let the unconstrained thermal equilibrium distribution of the physical substrate be \pi. In CCE accounting, the physical information content of a macroscopic state distribution p is measured using the Kullback-Leibler (KL) divergence relative to this equilibrium distribution:

This divergence represents the physical distinction held by the observer's encoding relative to the ambient substrate dynamics. The minimum irreversible cost I_{\text{irr}} of transitioning the system from distribution p to p_{\text{final}} is determined by the difference in their physical distinctions:

To see the impact of this dynamical ledger, consider an asymmetric bit with two macroscopic basins: Left (L) and Right (R). Due to differences in energy levels, boundary geometry, or coupling to the environment, the phase-space volumes corresponding to these states are unequal: V_R = \alpha V_L (with \alpha \neq 1). The prior equilibrium probabilities of the basins are:

If we prepare the system in a distribution where p(L) = p and p(R) = 1-p, and subsequently execute a logically irreversible Reset-to-Left operation (yielding p_{\text{final}} = (1, 0)), the logical Shannon cost is H_S(p). However, the CCE dynamical cost is:

The CCE dynamical accounting recovers the logical Shannon term H_S(p) plus a physical asymmetry penalty (1-p)\ln\alpha.

Crucially, the asymmetry parameter \alpha is a fixed physical property of the hardware (the geometric or energetic ratio of the basins), whereas how this asymmetry penalizes or discounts the transition is determined by the direction of the operation. Here, because \alpha > 1 (meaning the Right basin is larger), sweeping probability out of the larger basin into the smaller Left basin requires compressing the phase space. This direction-dependent physical effort is why the operation pays a penalty of +(1-p)\ln\alpha.

If we instead executed a Reset-to-Right operation (sweeping the system into the larger basin, so p_{\text{final}} = (0, 1)), the CCE dynamical cost would be:

In this direction, the penalty becomes a thermodynamic discount of -p\ln\alpha. Because the operation allows the system to expand from the smaller, constrained Left basin into the larger Right basin, the physical substrate assists the transition. Under pure Shannon accounting, both operations cost the same logical H_S(p) nats. Under CCE, the logical ledger is bound to the physical hardware: compressing data against the substrate's natural asymmetry costs extra work, while expanding with it recovers work.

2. From Micro-Reality to Macro-Equations

We now scale this concept from a single bit to a macroscopic equation modeling a high-dimensional microscopic reality. Let the true microscopic state of a system be x in a high-dimensional phase space \mathcal{M}, with its probability distribution evolving under the microscopic Liouville or Fokker-Planck flow as P_{\text{micro}}(x, t).

A macroscopic observer uses a coarse-graining projection operator \Pi: \mathcal{M} \to \mathcal{C} to map these microstates to a reduced, continuous field variable \phi(t) \in \mathcal{C}. The observer models the system using a macroscopic equation of motion:

By relying solely on \phi(t), the observer implicitly assumes a microscopic density constructed via the maximum-entropy (or local equilibrium) lift operator \Pi^*:

However, the true microscopic distribution P_{\text{micro}}(x, t) evolves under the full chaotic and coupled microscopic laws. As time progresses, microscopic interactions generate fine-grained fluctuations, correlations, and gradients across the boundaries of the coarse-grained cells. These details are neglected by P_{\text{macro}}(x, t), which assumes local equilibrium within the macroscopic states.

3. Nats of Divergence: A Concrete Example

The difference between prediction and retrodiction under a macroscopic model can be illustrated using a classical physical system: a particle of mass m moving in a viscous fluid with friction coefficient \gamma.

Let the true microscopic dynamics of the particle's velocity v be stochastic due to collisions with fluid molecules (thermal noise), described by the Langevin equation:

where D is the diffusion coefficient and W_t is a standard Wiener process. The true microscopic probability density P_{\text{micro}}(v, t) evolves under the Fokker-Planck equation, tending toward the thermal equilibrium variance \sigma_{\text{eq}}^2 = D/\gamma.

A macroscopic observer ignores the thermal fluctuations and models the velocity using the deterministic decay equation:

The observer's measurement instrument has a finite resolution represented by a narrow Gaussian of variance \sigma_0^2 (where \sigma_0^2 \ll \sigma_{\text{eq}}^2).

A. Prediction Divergence

Suppose the observer prepares the particle in a highly localized velocity state v(0) = v_0 with variance \sigma_0^2 at t = 0. After a time interval T, the true microscopic distribution spreads due to diffusion:

Meanwhile, the macroscopic model predicts v_{\text{macro}}(T) = v_0 e^{-\gamma T} and assumes the measurement resolution remains \sigma_0^2, yielding P_{\text{macro}}(v, T) = \mathcal{N}(v_0 e^{-\gamma T}, \sigma_0^2).

The predictive divergence in nats is the Kullback-Leibler divergence between these two distributions:

where \sigma_T^2 = \sigma_0^2 e^{-2\gamma T} + \sigma_{\text{eq}}^2(1 - e^{-2\gamma T}). As T becomes large, the true variance approaches the thermal variance \sigma_{\text{eq}}^2, and the predictive divergence plateaus at a constant level:

This represents the information lost by ignoring environmental fluctuations, which is bounded by the ratio of thermal noise to observer resolution.

B. Retrodictive Divergence

Now suppose the observer measures the velocity at time T to be v(T) = v_T. To reconstruct the past state at t = 0, the macroscopic modeler runs the deterministic equation backward in time, yielding the retrodictive estimate v_{\text{macro}}(0) = v_T e^{\gamma T} with resolution variance \sigma_0^2.

However, the true microscopic retrodiction is the Bayesian posterior distribution of the initial state given the measurement. Under a thermal prior P(v(0)) = \mathcal{N}(0, \sigma_{\text{eq}}^2), the posterior distribution is:

Because high-energy states are exponentially rare under the thermal prior, observing a velocity v_T today does not mean the particle started with a massive velocity v_T e^{\gamma T} that slowly dissipated. Rather, it is overwhelmingly more probable that the particle was near thermal equilibrium (near zero) in the past, and a recent random thermal fluctuation pushed it to v_T. The macroscopic model completely ignores this thermal prior, hallucinating a physically exorbitant, high-energy history.

The retrodictive divergence in nats between this true historical distribution and the macroscopic model's backward projection is:

As T grows, the difference between the true posterior mean and the macroscopic model's retrodiction grows exponentially. The divergence is dominated by the mean mismatch:

Unlike the prediction divergence which plateaus, the retrodictive divergence grows exponentially with T. This asymmetry represents the informational debt of coarse-graining: the macroscopic model throws away phase-space volume in the forward direction, which requires exponential precision (information) to reconstruct in reverse.

C. Physical Analogy: Singularities and Heat Death

These informational divergences provide a direct bridge to physical dynamics and energy under CCE. If we take the macroscopic equations as they are and project them backward in time under an assumption of costless resolution, the observer must pack an exponentially growing amount of energy to track and re-instantiate the lost historical details, mirroring the divergence of a physical singularity. Conversely, in the forward predictive limit, the complete export of the observer's organized energy and signal into the thermal bath matches the dynamics of heat death.

Although these asymptotic limits diverge from physical reality—because real observers are finite and bounded by resource constraints—the analogy highlights a key CCE principle: thermodynamic events like singularities and heat death can be understood as the operational boundaries of macroscopic coarse-graining when pushed to its limits.

Conclusion

Rather than combining these divergences, it is the individual terms and the deep asymmetry between them that expose the Platonic Observer Fallacy. The predictive divergence, D_{\text{pred}}(T), plateaus at a constant level, indicating that the observer's loss of forward predictive capacity is bounded by the finite scrambling capacity of the thermal bath. Conversely, the retrodictive divergence, D_{\text{retro}}(T), grows exponentially over time, representing the severe informational debt incurred when trying to reconstruct history from a dissipative macroscopic representation. This asymmetry demonstrates that macroscopic models are not self-contained descriptions of reality; their forward evolution silently discards microscopic coordinates, demanding an exponentially growing amount of information to resolve in reverse.