beneath the surface of belief

Before you have opinions about taxes or immigration or whether the president is doing a good job, you're a biological organism dropped into an environment. Your nervous system reacts to threats at a speed you didn't choose. Your dopamine system seeks or avoids novelty in ratios you didn't set. Your capacity for empathy was largely determined before you could speak.

None of this sits at the surface of your politics, but all of it shapes what your politics will become. Everything upstream of explicit politics is really just the problem of an agent encountering a world and learning within that world.

a stranger in a room

There's a field called reinforcement learning that studies how agents learn from interactions, and could be the gold standard for thinking about politics. The setup is that an RL gets dropped into an environment, takes actions, gets feedback, and over time develops a policy — when I'm in this kind of situation, here's what I do — and a value function, a learned sense of how good or bad each situation is.

So the policy is what the agent does. The value function is why. Another way to think about this is to distinguish between the algorithm and the data. The algorithm is your built-in learning architecture that's downstream of some innate hardware — your genetics and neurology. The data is the stream of experience the world feeds into that architecture. Politics begins to form where those two meet: neither in the organism alone nor the environment alone, but in the learned policy and value function that emerge from their interaction.

unspooling the feedback loop

The problem with actually trying to model this is that the process is recursive. Your behavior changes your environment, your environment reshapes the options available to you, those options shift which dispositions become salient, and those dispositions generate new behavior — your media diet shapes your politics which shapes your media diet, and so on indefinitely. Causality running in both directions means you're no longer looking at a chain of causes but a system folding back into itself, which is analytically difficult because causal inference depends on acyclic structure. Feedback loops don't just complicate the model; they undermine the kind of inference the model requires.

And even bracketing the feedback problem, the full state of a political agent is staggering in scope — genome, neurology, upbringing, cultural context, economic conditions, the entire history of their interactions with the political system. There's no realistic path to collecting signal on all of it, and collapsing it into something tractable necessarily means throwing things away.

The standard response to this kind of intractability is to find some way of simplifying without losing what matters. One approach that shows up across disciplines is the Markov property: if you can represent an agent's current state in a way that captures everything relevant to predicting what happens next, you can discard the history and work only with the present. The past doesn't disappear so much as compress — it becomes legible through its effects on the current state. Whether that compression preserves what actually matters is a different question.

In real political development, it often doesn't. A person's future political movement isn't determined just by their visible political profile right now. It depends on deeper strata: neurology, formative experience, socialization, accumulated feedback, latent sensitivities. Two people who look politically identical today can diverge sharply tomorrow, because the hidden structure that produced those outward similarities is completely different. The history hasn't really disappeared. It's still living inside the state. We just don't observe it.

So instead of trying to model the full recursive process in real time, I flattened it. Take the feedback loop and cut it at natural joints — developmental periods where the agent's state crystallizes before the next medium acts on it. Each cut gives you a snapshot. String the snapshots together and the intractable loop becomes a directed chain: Genome, Neurology, Crystallized Personality, Political Disposition, Political Alignment, Political Behavior.

Each link in the chain is a refraction. An internal state meets an external medium, and something new comes out. And if you hold the layers below the one you're trying to simulate constant, you get the Markov property back — not as a free assumption, but as something you earn by choosing where to cut.

I find it helpful to think about each layer in terms of what's endogenous, internal to the agent, and what's exogenous, in the environment. At every refraction, an endogenous state encounters an exogenous medium, and the medium reshapes the state into something new. That new form becomes the endogenous input to the next layer.

And at each refraction, there are really only two questions: is this a different algorithm, or different data? When the same genome develops under different conditions, that's different data run on the same algorithm. When the same upbringing meets a fundamentally different neural architecture, that's the same data run on a different algorithm. The endogenous state is the algorithm. The exogenous medium is the data. The output is the learned weights — and those weights become the next layer's algorithm.

Here's what this looks like concretely. Start with neurology — the hardware the agent was born with. Threat sensitivity, novelty-seeking, empathy bandwidth. This is the endogenous state before the environment has meaningfully shaped it.

Now that state encounters its first exogenous medium: childhood socialization. Family structure, religious upbringing, neighbourhood, class. The neurological predispositions don't disappear — they get refracted through this medium. A high-threat-sensitivity child raised in a stable, trusting community develops differently from the same child raised in an unstable one. Same algorithm, different training data, different learned weights. The output becomes the input to the next layer.

Each subsequent layer does the same thing at a different developmental timescale, producing a more politically legible version of the agent, until you reach the disposition profile: the thing we actually measure.

the layers

exogenous activation

There's an important caveat: political dispositions in a vacuum are not 100% predictive, because certain nodes can be activated at different times by the environment. Not all dimensions are fully salient all the time, creating a wrinkle in forecasting how certain political disposition archetypes behave in the future, since you can't hold the environment constant.

A breakdown in institutional trust activates one structure. Economic precarity activates another. Rapid demographic change, war, elite failure, cultural humiliation, technological upheaval — each changes which latent dimensions become salient. The agent is not rebuilt from scratch. Different parts of the same underlying structure are being activated — made more salient — by different exogenous conditions.

Consider the election of 1860. The exogenous condition was the fracturing of the political order over slavery — a moral and economic crisis that made cultural and moral dimensions suddenly dominant for people who had previously voted on other grounds entirely. Or consider the Depression: someone whose politics in 1928 was organised around social issues or cultural identity might have shifted into FDR's coalition by 1932, not because their underlying disposition changed, but because the economic catastrophe activated their material interests so powerfully that those dimensions overwhelmed everything else. The disposition was always there. The environment decided which part of it mattered.

political labels as principal components

The word liberal has meant so many different things in so many different places that it barely functions as a descriptor anymore. In nineteenth-century Europe it named market openness and civil liberty. In mid-century America it named redistribution and labour. In France it still carries classical-liberal connotations closer to what Americans would call libertarian. In Russia it became an insult. And in the United States right now, you almost never hear it. Why?

These terms are better understood as principal components than as real dimensions. They're summary bundles of correlated views, identities, and intuitions — not singular underlying axes that explain who someone is. Like PCA in statistics, they compress a high-dimensional disposition space into convenient labels that capture the most variance in a given era. What liberal captures in 1972 — Great Society, labour unions, anti-war — is not what it captures in 2024 — identity politics, credentialism, technocratic governance. Same label, different loading vector. The components rotate when the coalitional sorting changes.

This is why the labels feel increasingly strained. When the coalition structure is stable, the first principal component — the liberal-conservative axis — explains a lot of variance and the labels feel natural. When coalitions are realigning, as they are now, the loading vector is rotating, and people who haven't changed at all suddenly find themselves without a label that fits. They didn't move. The axis did. The word stopped being useful not because people stopped having politics, but because the coalition it summarised is in the process of flying apart.

The fourteen-dimensional PRISM space doesn't compress to a line. It preserves the cross-pressures that the PCA projection destroys: the person who is economically redistributionist and culturally traditional, the person who is procedurally conservative and morally universalist, the person who is deeply engaged but refuses partisan identity. These combinations are common in the population and invisible on a left-right spectrum.

You are a stranger in a room. You have hardware you didn't choose. That hardware was calibrated by an upbringing you didn't choose. That calibration crystallized into a personality you only partially chose. That personality, encountering the cultural and economic conditions of your era, produced a set of political dispositions you experience as beliefs. Those dispositions, compressed into ideological language and filtered through the affordances of the political moment, produce the behaviour the world sees and calls your politics.

The political spectrum is a projection of this process onto a line. PRISM is an attempt to map the process itself.