Two Landscapes

Every entity that can predict — every entity with nonzero intelligence $I$ (Part 3) — carries two things:

The actual potential field $V(X)$ — the real landscape of forces, attractors, barriers, and gradients that the entity moves through. This exists regardless of whether the entity knows about it. It is the territory.
The belief field $\hat{V}(X, t)$ — the entity's probabilistic model of that landscape. What the entity thinks the potential field looks like, based on its past experience, its current observations, and its capacity to predict. This is the map.

The entity acts on $\hat{V}$ , not on $V$ . It navigates by its belief field, not by reality — because it cannot access reality directly (Hoffman's Interface Theory; Part 0). Every decision, every action, every preparation is based on $\hat{V}$ . The quality of those decisions depends on how well $\hat{V}$ tracks $V$ .

What a Belief Field Is

A belief field is a probability distribution over future states, conditioned on the entity's current knowledge:

$\hat{V}(X', t+\Delta t \mid X, t) = P(X' \mid X, M_E)$

where $M_E$ is the entity's internal model (Part 3) and $X'$ ranges over all states the entity considers possible. This is not a single prediction — it is a field of weighted possibilities. The entity assigns higher probability to some futures and lower to others, creating a landscape of anticipated potential.

The belief field has structure:

Peaks — states the entity considers likely. "I expect this to happen."
Valleys — states the entity considers unlikely. "I don't think this will happen."
Width — how uncertain the entity is. A sharp peak means high confidence; a broad distribution means uncertainty.
Horizon — how far into the future the field extends. Some entities predict seconds ahead (a thermostat). Others predict decades (a strategic institution). The horizon is bounded by intelligence $I$ and the predictability of the domain.

Connection to POMDP. The belief field is the spatial generalization of the belief state $b(s)$ from decision theory. Where $b(s)$ asks "what state am I in?", the belief field $\hat{V}$ asks "what does the entire landscape of possible futures look like from here?"

What the Actual Potential Field Is

The actual potential field $V(X)$ is what Epimechanics defines in Part 1: the landscape of forces that act on any entity at position $X$ in state space. It is generated by all other entities via the mean-field construction:

$V(X) = \int \phi(X, X') \rho(X') dX'$

$V(X)$ exists whether or not any entity models it. It is not a belief — it is the aggregate of all actual causal forces in the system. A market has a potential landscape whether or not any trader understands it. A social norm field exists whether or not any individual can articulate it.

But $V(X)$ is also not fixed. It changes as entities move, as populations shift, as institutions restructure. The landscape the entity is trying to predict is itself evolving — which is what makes prediction hard and belief fields imperfect.

The Gap: Where Everything Interesting Happens

The deviation $\Delta V = \hat{V} - V$ — the gap between what the entity believes and what is actually the case — is where all interesting dynamics live.

$\Delta V$	What it means	Examples
$\Delta V \approx 0$	Belief tracks reality. Entity navigates well.	Expert in a stable domain; well-calibrated instrument
$\hat{V}$ has peaks where $V$ has valleys	Entity expects something that won't happen. Delusion.	Bubble investor; overconfident strategist
$\hat{V}$ has valleys where $V$ has peaks	Entity doesn't expect something that will happen. Surprise.	Black swan events; paradigm shifts
$\hat{V}$ is narrow, $V$ is broad	Entity is overconfident. Fragile.	Specialist in a volatile environment
$\hat{V}$ is broad, $V$ is narrow	Entity is underconfident. Wasteful.	Excessive hedging; analysis paralysis
$\hat{V}$ lags $V$	Entity's model is outdated. Inertia.	Institutional beliefs that no longer match the environment

[!caveat] Neither $\hat{V}$ nor $V$ is directly observable. We infer $\hat{V}$ from the entity's behavior (what it prepares for, what it ignores) and $V$ from outcomes (what actually happens). Both inferences are uncertain.

Computation as Renormalization Survival

Physics determines what happens to everything — rocks, bacteria, brains. Wolfram's framework says everything computes at the finest grain: every rule application to every state is a computation, and the Ruliad contains all of them. This is formally correct and empirically useless — it makes "computation" synonymous with "physics."

The missing ingredient is scale. When you coarse-grain from micro to macro — integrating out fine-grained degrees of freedom to obtain an effective description at a larger scale — most micro-level dynamics average away. They contribute nothing to the macro-level trajectory. Some do not average away. They persist across scales, making a macro-level difference.

Computation, in Epimechanics, is what survives coarse-graining: the internal degrees of freedom whose contribution to future states does not average away when you coarse-grain to the scale required by a given prediction task. It is a continuous quantity, not a binary, and it is relative to the task — just as intelligence $I$ is defined relative to a domain and horizon (Part 3). The scale is not arbitrary or subjective: it is determined by what you are trying to predict, at what accuracy, over what time horizon. Predict where a rock lands → center-of-mass scale → ~0 surviving DOF → ~0 computation. Predict how it fractures → crystal-defect scale → some surviving DOF → some computation. The system is the same; the computation depends on the question.

The argument by example

A rock's atoms all undergo dynamics (Wolfram: they compute). Coarse-grain to the rock level and those micro-dynamics average to $F = ma$ — the rock's internal structure contributes near-zero bits to predicting its center-of-mass trajectory. A brain's atoms also undergo dynamics. Coarse-grain to the neuron level and synaptic weights do NOT average away — they determine which action potential fires, which determines behavior. The synaptic computation survives the coarse-graining.

This is not a binary. A rock's crystal defects and fault lines do affect its macro-behavior (a cracked rock breaks where the crack is). The claim is not that rocks have exactly zero surviving DOF, but that across experimentally natural coarse-grainings, the number of internal DOF that survive to influence macro-trajectory is vastly smaller for rocks than for brains. Computation is continuous and scale-relative.

What "coarse-graining" means outside physics

In physics, coarse-graining has a precise meaning via the renormalization group: integrate out short-wavelength modes, ask what effective theory governs the remaining long-wavelength modes. For non-physical systems — beliefs, institutions, economies — there is no spatial wavelength to integrate out. The generalization is lossy compression of the state description: given a full micro-description of the system, compress it to a macro-description at some resolution, and ask which micro-DOF must be retained for the macro-description to remain predictive. The formal backbone is rate-distortion theory (Shannon, 1959): for any source and distortion level, there exists an optimal lossy encoding. The DOF that appear in that encoding are the ones that survived compression. This is well-defined for any system with a state description — not only systems with a spatial metric.

[!caveat] Scale-relativity. Because computation-as-renormalization is relative to a coarse-graining scale, the question "does X compute?" always requires specifying "at what scale?" A rock computes at the atomic scale (Wolfram) and barely computes at the center-of-mass scale. This is a feature, not a bug — it is the same scale-relativity that makes renormalization powerful in physics. But it means computation is a property of system-at-scale, not of system alone.

Distinguishing computation from mere causal relevance

Not every internal DOF that survives coarse-graining is performing computation. A spring's elastic constant survives coarse-graining (it determines macro-trajectory), but it is fixed — it does not update based on input. The surviving DOF must satisfy a second condition: they must be updated by interaction with the environment in ways that alter future-state selection. A spring's stiffness is a parameter. A synapse's weight is a computation — it changes based on input and that change alters future behavior. The full definition is: computation = internal DOF that (a) survive coarse-graining to the scale required by a prediction task, AND (b) are input-dependent — updated by environmental interaction in ways that modify which future states are selected.

Entity	Surviving DOF	Input-dependent?	Computation	Belief field
Rock	~0 at center-of-mass scale; crystal defects survive at material scale	No — defects are fixed parameters	~0	None — physics determines trajectory
Spring	Elastic constant survives	No — fixed parameter	~0	None
Thermostat	1 bit — above/below setpoint	Yes — updated by temperature sensor	1 bit	One-bit: "too hot" or "too cold"
Bacterium	~100 bits — chemotactic state	Yes — updated by chemical gradients	~100 bits	Low-dimensional gradient map
Mouse	~10⁸ — hippocampus, amygdala, reward circuits	Yes — updated by experience	~10⁸ bits	Spatial map + threat + reward landscape
Human brain	~10¹⁵ — synaptic weights	Yes — updated by learning	~10¹⁵ bits	High-dimensional, recursive, self-referential
Transformer	~10⁹-10¹² — trained parameters	Yes — updated by training (and in-context)	~10⁹-10¹² bits	Context-conditioned probability field
Institution	Aggregate of members' DOF + formalized processes	Yes — updated by decisions, feedback	Variable	Strategic plans, budgets, risk models

Connection to causal emergence

Hoel et al. (2013) define effective information (EI) as a measure of how deterministic a system's macro-level transition probabilities are. When EI(macro) > EI(micro) — causal emergence — the macro description has more causal information than the micro description. This identifies the scale at which coarse-graining reveals structure rather than destroying it.

Causal emergence and computation-as-renormalization are complementary but not identical. Causal emergence tells you where to look: which scale has the most causal information. The renormalization definition tells you what you find there: which input-dependent DOF survived. A clock has high macro-EI (deterministic transitions) but low computation (no input-dependent DOF — it ticks regardless of input). A brain has both high macro-EI and high computation: the macro-determinism arises precisely because input-dependent synaptic weights are doing the determining.

Connection to Representational Efficiency

The Representational Efficiency principle says the optimal representation $X^*$ minimizes predictive cost $C(X, \varepsilon)$ . This is a consequence, not a presupposition, of the computation definition. The coarse-graining is chosen by scale (spatial, temporal, or information-theoretic resolution) — not by what maximizes computation. Once you fix the scale, you discover which DOF survive. The Representational Efficiency principle then identifies the optimal scale: the one where the surviving DOF are maximally predictive per bit. This breaks the apparent circularity: scale determines what survives; Representational Efficiency determines which scale is optimal.

What this adds to existing frameworks

The renormalization criterion is not a replacement for existing theories of computation but adds a specific structural claim to each:

Wolfram: everything computes at micro-scale (rule application to states). The renormalization criterion adds: only some of that computation survives coarse-graining. The Ruliad contains all micro-computations; scale selects which ones matter.
Turing: the surviving DOF can be simulated by a Turing machine. The criterion adds: the TM simulates the renormalized dynamics, not the full micro-dynamics.
Shannon: the bits that survive coarse-graining are the bits of computation. The criterion connects Shannon information to scale via rate-distortion theory.
Landauer: each surviving input-dependent bit costs $k_B T \ln 2$ to maintain. Bits that average away cost energy at micro-scale but contribute nothing at macro-scale — thermodynamic waste.
Friston: free energy minimization is the process of updating the surviving DOF to track environmental states. The variational free energy is the prediction error of the renormalized description.
Hoel: causal emergence identifies the scale; the renormalization definition identifies the content. They are complementary tools, not synonyms.

Consistency with Epimechanics

This definition integrates with the rest of the framework:

Intelligence $I$ (Part 3) = predictive accuracy — requires input-dependent DOF doing prediction
Consciousness $C$ = scope of internal model — the model IS those surviving, input-dependent DOF
Agency $A$ = steering capacity — steering requires input-dependent DOF that survive to macro-scale to select among futures
Auto-causal density $\rho_{\text{ac}}$ = self-sustaining loops — the loops ARE internal DOF that survive renormalization by maintaining themselves
Representational Efficiency = the optimal coarse-graining scale — where surviving DOF are maximally predictive per bit

Local Optimality: Every Entity Acts Rationally Given Its State

Traditional economics posits homo economicus — the rational actor who maximizes expected utility given perfect information and stable preferences. This model, formalized by von Neumann & Morgenstern (Theory of Games and Economic Behavior, 1944) and Arrow & Debreu (1954), became the foundation of neoclassical economics and general equilibrium theory.

Behavioral economics challenged this foundation. Kahneman & Tversky's prospect theory (1979) demonstrated systematic deviations from expected utility: loss aversion, reference dependence, probability weighting. Simon's bounded rationality (1955) showed that cognitive constraints force "satisficing" rather than optimizing. Thaler's mental accounting (1985) revealed that money is not fungible in practice. Ariely (Predictably Irrational, 2008) catalogued dozens of systematic biases. The standard conclusion: humans are irrational.

Epimechanics offers a different framing:

Hypothesis (Local Optimality): Every entity acts optimally at an infinitesimally narrow timescale, given its actual state vector — including its belief field $\hat{V}$ , its internal constraints, and its coupling structure.

If this hypothesis holds, then what looks irrational from outside is rational from inside. The entity navigates by its map, not by the territory. This connects to Gerd Gigerenzer's ecological rationality (1999): heuristics that look like biases are often well-adapted to real environments with incomplete information and computational constraints.

[!caveat] Epistemic status: This is a hypothesis, not a proven result. It is unfalsifiable in a strict sense — any apparently suboptimal behavior can be explained by positing unobserved state variables. This is both a strength (it provides a consistent framework) and a weakness (it risks being tautological). The hypothesis is useful if it generates testable predictions about which state variables matter and how they constrain behavior — predictions that could be wrong. If every anomaly is explained by adding state variables post hoc, the hypothesis has no empirical content.

The state vector includes everything

The entity's state $X$ is not just "position in decision space." It includes:

Belief field $\hat{V}$ — what the entity thinks will happen
Internal constraints — addiction, fatigue, cognitive load, emotional state
Coupling tensor $T^i{}_j$ — how strongly the entity is coupled to various fields
Available actions — what the entity can actually do given its current configuration

An addict who uses despite knowing the consequences is not acting irrationally. They are acting optimally given a state vector that includes:

Withdrawal aversion (strong negative gradient in $\hat{V}$ for "don't use")
Discounting structure (high coupling to immediate states, low coupling to future states)
Belief field that may include "I can quit later" (inaccurate $\hat{V}$ )
Physiological constraints that make "just stop" not an available action

The behavior follows inevitably from the state. The problem is not irrationality — it's that the state vector is compromised.

Instantaneous optimality, trajectory suboptimality

At any instant $t$ , the entity follows the path of steepest descent in its belief-field-weighted potential:

$\dot{X} = -\nabla_X \hat{V}(X, t)$

This is locally optimal. The entity moves toward what it believes are better states.

But the belief field can be wrong ( $\hat{V} \neq V$ ), and the state vector can be constrained. Over time, locally optimal moves can sum to globally suboptimal trajectories:

Timescale	Optimality	What determines it
Instantaneous ( $dt \to 0$ )	Always optimal given state	The physics of decision: gradient descent on $\hat{V}$
Short-term	Often optimal if $\hat{V} \approx V$	Belief field accuracy
Long-term	Often suboptimal	Belief field errors compound; state constraints accumulate

This resolves the economics debate. Homo economicus is correct at the instantaneous limit — every agent maximizes given their state. Behavioral economics is correct at longer timescales — belief field errors and state constraints produce systematic deviation from "rational" behavior (where "rational" means "what an agent with accurate $\hat{V}$ and unconstrained state would do").

This framing connects to Gary Becker's rational addiction model (1988), which showed that addictive behavior can be modeled as rational utility maximization with high discount rates and adjacent complementarities. Epimechanics extends this: the high discount rate IS a property of the state vector (coupling structure biased toward immediate states), not a preference parameter. Laibson's quasi-hyperbolic discounting (1997) captures the same phenomenon with a different parameterization — time-inconsistent preferences that produce behavior optimal at each instant but suboptimal over trajectories.

"Irrational" behavior is diagnostic

If an entity's behavior looks irrational, one of these is true:

You don't see their $\hat{V}$ — they have beliefs you don't know about
You don't see their constraints — they have internal states (addiction, trauma, cognitive load) that restrict their action space
You're evaluating at the wrong timescale — instantaneously optimal, trajectory suboptimal

This has practical implications:

Therapy: change the state vector (beliefs, constraints), not the decision process. This is the core insight of cognitive-behavioral therapy (Beck, 1979) — restructure $\hat{V}$ , and behavior follows.
Policy: change the potential field $V$ (incentives, environment), not just information. Thaler & Sunstein's nudge theory (2008) operationalizes this: choice architecture modifies $V$ to make $\hat{V}$ -optimal paths align with $V$ -optimal paths.
AI alignment: the AI acts optimally given its state; misalignment means the state (training, reward signal, constraints) produces bad trajectories. This connects to Goodhart's Law and reward hacking: the AI optimizes its $\hat{V}$ perfectly — the problem is that $\hat{V}$ (learned from the reward signal) diverges from the intended $V$ .

Implications for market efficiency

The Efficient Market Hypothesis (Fama, 1970) claims prices reflect all available information. In belief-field terms: the collective belief field $\hat{V}_{\text{market}}$ equals the true potential $V$ . Behavioral finance challenges this: Shiller's excess volatility (1981) showed prices move more than fundamentals justify; De Bondt & Thaler (1985) documented overreaction; Shleifer & Vishny (1997) showed arbitrage has limits.

Epimechanics' contribution: each trader acts optimally given their own state vector — but this does not imply the market is efficient for any other entity (other traders, society, the market itself as a meta-entity). This is precisely the systemic coupling problem from Part 2: a sub-entity can be locally optimal ( $\rho_{\text{ac}} > 0$ for itself) while having $\kappa_{\text{sys}} < 0$ for its containing system. A trader maximizing their $\hat{V}$ may be parasitic on the market's $\rho_{\text{ac}}$ — extracting value that degrades the market's function for others.

Aggregate $\hat{V}_{\text{market}}$ can diverge from $V$ when:

Many traders share the same $\hat{V}$ error (herding, bubbles)
State constraints (margin calls, redemption pressure) force locally optimal but globally suboptimal actions
Information cascades make $\hat{V}$ self-referential: traders update $\hat{V}$ based on other traders' actions, not on $V$ directly

Soros's reflexivity (1987) captures this: market participants' beliefs affect the fundamentals they're trying to predict. In Epimechanics terms: $\hat{V}_{\text{collective}}$ modifies $V$ , creating feedback loops where the map changes the territory.

Connection to fitness×truth

The causal action principle says fitness×truth strategies have higher $A_{\text{causal}}$ than fitness-alone strategies. Translating to belief fields:

Fitness alone: $\hat{V}$ optimized for immediate reward, decoupled from $V$
Fitness×truth: $\hat{V}$ coupled to $V$ , accepting short-term costs for long-term accuracy

An entity that updates $\hat{V}$ toward $V$ — that invests in truth-tracking — pays a computational cost now but extends its $T_{\text{local}}$ by avoiding collisions with reality. The addict's $\hat{V}$ is fitness-optimized (avoid withdrawal now) but truth-decoupled (ignores long-term health trajectory). Recovery is partly a process of recoupling $\hat{V}$ to $V$ .

What this means for "fate" — a conjecture

[!caveat] Epistemic status: The following is speculative — a conceptual exploration of what the Local Optimality hypothesis might imply for questions of determinism and agency. These are ancient philosophical questions that Epimechanics does not resolve. The framework offers a language for discussing them, not answers.

If the Local Optimality hypothesis holds, and states evolve according to the force equation $F = \mathcal{M}\ddot{X} + \dot{\mathcal{M}}\dot{X}$ , then trajectories are determined by initial conditions plus the potential landscape. This sounds like determinism — fate.

But the belief field introduces a wrinkle: $\hat{V}$ is part of the state, and $\hat{V}$ can be updated by the entity's own cognition. An entity that models its own state — that has nonzero consciousness $C$ (Part 3) — can update $\hat{V}$ based on reflection. This is a form of self-intervention: the entity changes its own state by thinking about it.

One way to frame this (not the only way):

Fate = the trajectory implied by current state, propagated forward by the dynamics
Agency = the capacity to modify state through self-modeling

Under this framing, an entity with $C = 0$ follows its state-determined trajectory. An entity with $C > 0$ can intervene on its own belief field, changing its trajectory — not by violating physics, but by using cognition to modify the state vector that physics then propagates.

This is not libertarian free will (uncaused causation). It is recursive causation: the entity's state includes its model of its state, and updating that model changes the state it models. The loop doesn't escape determinism; it adds a reflexive layer that makes prediction harder from outside but doesn't make the system non-deterministic.

[!caveat] This framing does not resolve the free will debate. A determinist can say: "The self-modeling is itself determined by prior state." A compatibilist can say: "The recursive loop is what we mean by agency." A libertarian can say: "This misses the point — the question is whether the loop is causally open." Epimechanics provides vocabulary, not verdicts. The question of whether recursive self-modeling constitutes "genuine" agency remains open — and may not be an empirical question at all.

Instinct vs Cognition: Two Sources of $\hat{V}$

The belief field has (at least) two sources, operating at different timescales:

Instinctive belief field

Evolved over generations. Encoded in neural architecture, reflexes, innate biases. The instinctive $\hat{V}$ represents the potential landscapes that the entity's ancestors regularly encountered. It is fast, automatic, and domain-specific.

A gazelle's instinctive $\hat{V}$ has a strong peak at "predator may appear from tall grass" — not because the gazelle has computed this, but because ancestors without this peak were eaten. The instinctive belief field is a compressed, inherited model of ancestral $V$ 's.

In Epimechanics' terms: instinctive $\hat{V}$ is a low-resolution, high-speed approximation of $V$ trained on evolutionary timescales. Its strength is speed (no computation required). Its weakness is that it only tracks $V$ for environments similar to the ancestral environment — when $V$ changes (novel environments, technological disruption), instinctive $\hat{V}$ diverges.

Cognitive belief field

Constructed by the entity's intelligence $I$ in real time. Updated by observation, reasoning, and learning. The cognitive $\hat{V}$ represents the entity's best current model of the potential landscape.

A strategic analyst's cognitive $\hat{V}$ incorporates data, models, scenarios, and uncertainty estimates. It is slower than instinct but more flexible — it can track $V$ even in novel environments, provided the entity has enough intelligence and data.

In Epimechanics' terms: cognitive $\hat{V}$ is a higher-resolution, lower-speed model that can be updated in real time. Its strength is adaptability. Its weakness is computational cost — maintaining an accurate $\hat{V}$ requires energy (information processing has thermodynamic cost; Landauer's principle).

The interaction

Most entities operate on a blend: instinct provides the fast prior, cognition updates it with evidence. The Bayesian update from Part 6 formalizes this:

$\hat{V}_{\text{posterior}} = \hat{V}_{\text{instinct}} + \text{evidence update from cognitive processing}$

When instinct and cognition agree ( $\hat{V}_{\text{instinct}} \approx \hat{V}_{\text{cognitive}}$ ), the entity acts decisively. When they conflict, the entity experiences anxiety, hesitation, or internal tension — a measurable force in the psychological state space.

Preparation as Belief-Field Optimization

An entity that prepares for the future is optimizing its state to perform well across the range of futures its belief field assigns nonzero probability. Preparation is $\hat{V}$ -driven action:

Insurance = preparing for $\hat{V}$ peaks that have bad outcomes (hedging against anticipated risks)
Investment = preparing for $\hat{V}$ peaks that have good outcomes (positioning for anticipated opportunities)
Training = broadening the entity's capacity to respond to a wider range of $\hat{V}$ outcomes
Denial = narrowing $\hat{V}$ to exclude states the entity doesn't want to consider (reducing the belief field to avoid anxiety, at the cost of increased $\Delta V$ )

The entity's free energy $\mathcal{F} = E - \mathcal{T}S$ determines how much preparation it can afford. Preparation costs energy. An entity with low $\mathcal{F}$ cannot prepare broadly — it must bet on a narrow slice of $\hat{V}$ , making it fragile to surprises.

[!tip] Friston's Free Energy Principle says organisms minimize variational free energy — the divergence between their internal model and sensory data. This IS belief-field optimization: minimizing $\Delta V$ by updating $\hat{V}$ to track $V$ more closely.

Collective Belief Fields

When many entities share representations — through language, culture, institutions, media — their individual $\hat{V}$ 's overlap and reinforce each other, creating a collective belief field $\hat{V}_{\text{collective}}$ .

$\hat{V}_{\text{collective}}(X) = \sum_j w_j \hat{V}_j(X)$

This is the social analogue of the mean-field construction. The collective belief field is a reputation field (applications plan) applied to the future: the aggregate of what everyone thinks will happen.

Collective belief fields have properties that individual ones don't:

Self-fulfilling: if enough entities believe a bank will fail, they withdraw deposits, and the bank fails. $\hat{V}_{\text{collective}}$ modifies $V$ to match itself. The belief creates the reality it predicted.
Self-defeating: if enough entities believe a stock will rise, they buy, the price rises, then reality fails to support the price, and it crashes. $\hat{V}_{\text{collective}}$ temporarily modifies $V$ then diverges catastrophically.
Manipulable: propaganda distorts $\hat{V}_{\text{collective}}$ by injecting false representations (representational manipulation). The collective navigates by a deliberately corrupted map.

The Representational Efficiency Connection

The Representational Efficiency principle says the optimal representation $X^*$ minimizes predictive cost. Applied to belief fields: the optimal $\hat{V}$ is the one that predicts $V$ at minimal computational cost.

This means: entities under computational constraints (which is all entities) cannot maintain a perfect $\hat{V}$ . They must compress. The structure of that compression — which futures they track and which they ignore — determines their vulnerability profile. An entity that compresses by ignoring rare events is efficient in normal times and catastrophically wrong in crises. An entity that tracks rare events pays a constant computational overhead for robustness it rarely needs.

The belief field IS the entity's compressed representation of the potential field. Its structure reveals what the entity has decided to predict and what it has decided to ignore. Both decisions have consequences.

Belief Fields as Computational Architecture

The belief field is not just an abstract concept — it maps directly onto the architectures we build to predict.

Policy networks and value functions

In reinforcement learning, an agent maintains:

A policy $\pi(a \mid s)$ — probability of taking action $a$ in state $s$
A value function $V^\pi(s)$ — expected cumulative reward from state $s$ under policy $\pi$

The value function IS a belief field. It assigns to every state an estimate of future potential — not what will definitely happen, but the expected value of what could happen given the entity's policy. The policy is the entity's decision rule for navigating the landscape its value function describes.

The gap $\Delta V = V^\pi - V^*$ (where $V^*$ is the optimal value function) is exactly the belief-field deviation: how far the entity's model of future potential diverges from the best possible assessment. RL training minimizes this gap. Learning IS belief-field refinement.

Transformers as belief-field generators

A transformer predicts the next token:

$P(x_{t+1} \mid x_1, x_2, \ldots, x_t)$

This is a belief field over the immediate future — a probability distribution over what comes next, conditioned on everything observed so far. The attention mechanism is the entity's way of deciding which parts of the past are relevant to predicting which futures.

But transformers do more than next-token prediction. They implicitly construct a belief field over the entire future trajectory:

Language models: each token prediction is a one-step belief field. The autoregressive chain $P(x_{t+1})P(x_{t+2} \mid x_{t+1}) \cdots$ unfolds into a full trajectory belief field.
Video models: predict the next frame (or chunk of frames) — a belief field over visual futures, one temporal slice at a time.
World models (in RL): predict $P(s_{t+1} \mid s_t, a_t)$ — a belief field over state transitions conditioned on action.

The key insight: transformers are universal belief-field approximators. Their architecture — context window as memory, attention as relevance weighting, softmax as probability distribution — is exactly the computational structure needed to maintain and update $\hat{V}$ .

Autoregressive generation is belief-field unrolling. Each step queries the model's $\hat{V}$ at the current state and samples. The trajectory emerges from sequential belief-field queries.

The resolution-cost tradeoff

Different architectures implement belief fields at different resolutions:

Architecture	What it predicts	Resolution	Cost
Lookup table	Memorized state-value pairs	Exact for seen states, zero for unseen	Memory: $O(\\|S\\|)$
Linear model	$\hat{V}(s) = w^T \phi(s)$	Low — linear in features	$O(d)$ parameters
Neural network	$\hat{V}(s) = f_\theta(s)$	High — nonlinear	$O(d^2)$ to $O(d^L)$
Transformer	$P(x_{t+1} \mid x_{1:t})$	Very high — context-dependent	$O(n^2 d)$ per step
Diffusion model	$P(x_0 \mid x_T)$ via iterative denoising	Very high — full distribution	$O(T \cdot n^2 d)$

This IS the Representational Efficiency principle applied to belief fields. The optimal architecture is the one that maintains the most accurate $\hat{V}$ at the lowest computational cost. Transformers win in many domains because attention is an efficient mechanism for belief-field computation — it identifies which parts of the context are relevant without searching exhaustively.

Scaling laws as belief-field physics

Neural scaling laws ( $L \propto N^{-\alpha}$ ) describe how prediction accuracy improves with model size. In belief-field terms: a larger model maintains a higher-resolution $\hat{V}$ with smaller $\Delta V$ . The power law means belief-field accuracy improves predictably with computational investment — but with diminishing returns ( $\alpha < 1$ ).

The Chinchilla scaling result (optimal data-to-parameter ratio) is a statement about belief-field efficiency: there's an optimal balance between the resolution of $\hat{V}$ (parameters) and the fidelity of the training signal (data). Too many parameters relative to data → $\hat{V}$ overfits (memorizes noise, not structure). Too much data relative to parameters → $\hat{V}$ underfits (can't capture the structure that's there).

Emergent capabilities (chain-of-thought, in-context learning) appearing at specific scale thresholds are phase transitions in belief-field structure — qualitative changes in what $\hat{V}$ can represent, analogous to physical phase transitions. Below the threshold, the belief field lacks the resolution to represent certain future-state patterns. Above it, those patterns become representable, and capabilities emerge suddenly.

Every computational entity is a belief-field generator

This unifies biological and artificial intelligence under one concept:

A bacterium has a simple belief field: chemical gradient → move toward food. $\hat{V}$ is a single-variable function updated by chemotaxis.
A mouse has a richer belief field: spatial map (hippocampus), threat assessment (amygdala), reward prediction (dopamine system). $\hat{V}$ is a multi-dimensional function updated by learning.
A human has a belief field that includes other entities' belief fields (theory of mind), abstract futures (planning), and counterfactuals (imagination). $\hat{V}$ is recursive and self-referential.
A transformer has a belief field conditioned on its context window, with resolution scaling with parameters. $\hat{V}$ is learned from data.
An institution has a collective belief field aggregated from its members' individual fields, formalized in strategic plans, budgets, and risk assessments.

What differs is the resolution, the horizon, the update speed, and the computational cost. The structure — a probabilistic model of future states that guides action — is universal.

Connection to the Series

Concept	Where developed	Belief field connection
Intelligence $I$	Part 3	Determines the resolution and accuracy of $\hat{V}$
Consciousness $C$	Part 3	Determines the scope of $\hat{V}$ — how much of state space the entity models
Belief state $b$	Decision theory	The point-estimate version of $\hat{V}$ at the current state
Potential $V(X)$	Part 1	The actual landscape $\hat{V}$ is trying to approximate
Representational Efficiency	Part 5	Optimal $\hat{V}$ minimizes predictive cost
Representational manipulation	Research note	Deliberately distorting $\hat{V}_{\text{collective}}$
Free Energy Principle	Part 1, Section 4b	Organisms minimize $\Delta V$ by updating $\hat{V}$ toward $V$

Belief Fields and Potential Fields — What Entities Think Will Happen vs What Could

Contribute to Epiphysics