The Spectrum of Epistemic Fidelity: A Thermodynamic Audit of the Agent
The Spectrum of Epistemic Fidelity: A Thermodynamic Audit of the Agent
By Andraž Đurič
With Claude (Anthropic) as collaborator
Draft, May 2026
This paper develops the epistemic-fidelity dimension within the framework of Thermodynamic Realism (Đurič, May 2026). It treats one axis of agent evaluation in detail: the degree to which an agent's internal models track the external structure they are coupled to. The framework's full architecture (operator-bound ought-semantics, K-coupling across nested embeddings, horizon-as-exogenous, affect as telemetry, compatibilist agency) is assumed; the parent paper provides the derivation.
1. The Basal Premise
Every agent, from the basal lineage of self-maintaining replicators to the modern human, is a localized engine of negentropy maintenance. To persist against the Second Law, an agent must maintain an internal Map of the external Territory it is embedded in.
The agent's epistemic fidelity is a measure of Map-Territory correspondence. Agents can be described on a fidelity axis:
- Left limit: Zero Fidelity. Map and Territory have no correlation. The agent cannot predict and cannot harvest energy. The agent does not persist. The left limit is asymptotic toward agent dissolution.
- Right limit: Asymptotic Fidelity. Map approaches Territory in resolution and accuracy. The right limit is also asymptotic. By P4.1 of the parent framework, an embedded agent cannot losslessly compress its own embedding; the agent is part of the system it would compute. The right limit is not a reachable point but a direction of motion.
The epistemic project is the asymptotic approach to the right limit, given the agent's finite hardware constraints.
2. The Compression Hack
Total mapping is impossible for a finite agent. The territory is too large. The recursive bottleneck (the agent is in the territory) is permanent.
But reality is compressible. The agent does not need a 1:1 map of every atom. It needs the basal laws of physics, which act as compressed source code: if the agent possesses the small set of generative rules, it can unpack reality in real-time at the resolution its hardware allows.
Moving toward the right limit is the process of stripping inherited user-interface representations (folk metaphysics, tribal categories, motivated post-hoc rationalizations, comfortable narratives) and replacing them with representations closer to the generative substrate. Not because the interface representations are evil, but because they impose error-cost wherever they diverge from the substrate, and the divergence accumulates over time and across coupled systems.
3. The Distribution of Fidelity
Agents vary in fidelity. The variation is not a single linear scalar; it is multi-dimensional, with different error profiles across different domains.
- Universal cognitive biases. All human agents exhibit confirmation bias, motivated reasoning, planning fallacy, false-consensus effects, in-group favoritism, hyperbolic discounting, sunk-cost reasoning, and the rest of the standard cognitive-error catalogue. The fidelity axis is not where a species member sits on a line. It is where a species member sits on each of many dimensions, with model-quality varying across domains within a single agent.
- Median behavior. The bulk of agents occupy a region of the multi-dimensional space where models are good enough to support persistence in their actual embedding but carry significant model error in domains the embedding does not stress. The cost of these errors is paid locally and is partly absorbed by cooperative scaffolding: shared institutions, division of cognitive labor, social distribution of error-cost across coupled agents.
- Tail behavior. At both tails of any specific error dimension, the metabolic and social cost of maintaining the model rises sharply. Heavy motivated reasoning, sustained denial of well-tested empirical claims, ideological commitment that overrides evidence: these are ordinary high-error states present at scale in neurotypical agents. The framework treats them as positions on continuous dimensions, not as categorical separation from the rest of the distribution.
- Right-tail effort. Science, philosophy, rigorous epistemic discipline, and contemplative traditions are intentional efforts to shift agents along specific fidelity dimensions. They reduce error in target domains at the cost of metabolic investment in modeling. The right tail is not a population type. It is a sustained orientation under specific institutional conditions.
The framework does not rank agents on a single inherent-value scalar. It describes the distribution of model-error across dimensions, in agents whose embeddings impose different selection pressures on different error types.
4. The Consistency Tax and Epistemic Profit
The Consistency Tax is the thermodynamic cost of mismatch between an agent's model of reality and reality itself. It comes in two forms:
- The cost of lies. An agent that intentionally maintains a false map must also maintain the true map to know what it is lying about. Holding two maps at once and reconciling them in real time imposes ongoing cognitive load. The cost scales with the complexity of the deception and the number of agents and domains the false map has to be defended against.
- The cost of being mistaken. An agent with a model that diverges from the territory through missing data, mismodeling, or inherited error pays a different but real cost: failed predictions, recovery work, friction with the structure of the embedding. The agent does not need to know it is wrong for the cost to accrue. The territory imposes the cost whether or not the agent recognizes it.
Both forms reduce to the same thermodynamic claim:
- Alignment = Efficiency. A model that tracks the territory reduces the energy spent on patching.
- Misalignment = Cost. A model that does not track imposes ongoing maintenance overhead, whether through deception or error.
This is the framework's local claim. It does not extend to literal cosmic-scale effects of individual misalignments. The defensible scope is the metabolic and cognitive cost paid by the misaligned agent and by the cooperative systems they are coupled to. Cosmic-scale framings, where useful, are metaphors riding on top of the local claim, not literal physical predictions.
The accumulated savings of a higher-fidelity agent across the same domains is the Epistemic Profit: the surplus capacity freed by not paying the Consistency Tax.
The felt signature of approaching the right limit is what we call Predictive Calm: the reduction in cognitive load that follows when the agent's models stop fighting the structure of the embedding. The agent no longer pays the friction-cost of expecting the world to be other than it is. This is not a mystical state. On the parent framework's account of affect as telemetry (parent paper §VI), it is the felt signature of low prediction-error operation.
5. The Dual Imperative
Epistemic fidelity alone does not specify the full activity of an agent. The parent framework's dual imperative applies:
- See clearer. Minimize prediction error. Maintain models that track the territory across the dimensions the embedding stresses.
- Act kinder. Reduce friction in coupled systems. Integrate predictive models with other agents. The binding-force structure: when two systems integrate their Markov blankets, their total free energy is less than the sum of their separated free energies.
Kindness-toward-coupled-agents is not a moral addition to epistemic fidelity. It is clarity about K-coupling correctly applied. An agent that ignores the persistence requirements of its coupled systems has a model-error regarding its own persistence conditions, because the agent's persistence is the persistence of the web it depends on (parent paper §IV).
Both imperatives reduce free energy at individual and collective scales. Neither is reducible to the other. The epistemic-fidelity axis is one face of the larger project; the cooperative-coupling axis is the other.
6. The Deterministic Outlier
An agent's position on the fidelity dimensions is a function of two variables:
- Initial hardware. Biological substrate, neurological constraints, computational capacity, sensory bandwidth.
- Input data. Environmental and informational packets received from birth onward, including the modeling traditions the agent is embedded in.
No agent chooses its position. Each agent is a localized calculation running with the inputs it was given on the hardware it was given. This is the parent framework's compatibilist commitment: agency is a particular kind of physical process within lawful causation, not a faculty above it. The agent's modeling activity is real, the cost of misalignment is real, the asymptotic project is real, and all of it is fully lawful.
There is no inherent value attached to a high-fidelity position. There is only Systemic Utility: the agent's predictive resolution across the dimensions its embedding stresses. An agent at very high fidelity in some domain is not "better" in a moral sense; the agent has higher resolution in that domain. The parent framework's agent-and-embedding-relativity (parent paper §VII) prevents any slide from "higher resolution" to "more worthy."
7. Summary
Traditional non-physical normativity is an illusion. Physical ought-facts, within the operator's domain, are real and physically determinate (parent paper §II, §III).
There is only physics, and the agent is an asymptotic project within it: the unending approach to the right limit of Map-Territory correspondence, constrained by the agent's hardware, governed by its embedding, motivated by the cost of misalignment, and accompanied by the dual imperative of clarity and coupling.
The question for the agent is not whether to engage the project. The question is whether the agent's models track the cost-structure of the territory clearly enough to engage it efficiently, or whether the agent pays a Consistency Tax in cognitive load to maintain models the territory does not cooperate with.
8. Visualizing The Spectrum of Epistemic Fidelity
References
Đurič, A. (May 2026). Thermodynamic Realism: In-Principle Determinacy and Ethics as Necessary Modeling. [parent paper]
Friston, K. (2010). The free-energy principle: a unified brain theory? Nature Reviews Neuroscience 11(2).
Jaynes, E. T. (1957). Information theory and statistical mechanics. Physical Review 106(4).
Shannon, C. (1948). A mathematical theory of communication. Bell System Technical Journal 27.
Schrödinger, E. (1944). What Is Life? Cambridge University Press.
Railton, P. (1986). Moral realism. Philosophical Review 95(2).
Foot, P. (2001). Natural Goodness. Oxford University Press.
Acknowledgment: Developed in dialogue with Claude (Anthropic). All theoretical responsibility rests with Andraž Đurič.
Comments
Post a Comment