Mechanism Realism

I. The Claim

Mechanism realism makes one claim that every mainstream political ontology rejects: mechanisms are more causal than individuals.

Not "mechanisms exist" — nobody disputes that. The claim is about causal primacy. The individual — their intentions, character, reasoning, values, vote — is embedded in mechanisms that determine the distribution of outcomes. Change the individual without changing the mechanism and you change nothing at scale. Change the mechanism and the individuals produce different outputs regardless of their character or intentions.

This displaces everything that feels most fundamental to modern political thought:

• Your intentions don't matter. The mechanism produces what it produces regardless of what you meant.
• Your character doesn't matter at scale. Good people in bad mechanisms produce bad outcomes. Bad people in good mechanisms produce acceptable ones.
• Your vote doesn't matter the way you think. You're selecting between agents embedded in the same mechanism. The mechanism constrains the output distribution. The agent is a parameter, not the function.
• Your values don't matter the way you think. What the mechanism selects for overrides what you say you value. A society that claims to value children while structuring every incentive against child-rearing does not value children. Its mechanisms value something else.

This is why mechanism realism meets resistance from every political tribe. The left wants outcomes to follow from compassionate intentions. The right wants outcomes to follow from individual virtue and responsibility. Libertarians want outcomes to follow from individual freedom. All three take the individual as the atomic unit. Mechanism realism says: the individual is real but is not the load-bearing variable. The architecture is.

Selection pressure enforces this. It doesn't argue. It removes. A bridge designed on wrong physics collapses regardless of the engineer's sincerity. A civilization designed on wrong models of human coordination collapses regardless of its moral commitments. The timeline differs. The physics doesn't.

II. What Is a Mechanism?

A mechanism is a causal structure that reliably shifts the distribution of outcomes.

Gravity is a mechanism. Incentive gradients are mechanisms. Natural selection is a mechanism. A legal system that punishes action more than inaction is a mechanism. A welfare design that destroys the capability it claims to build is a mechanism. A vague word that enables coalitions between incompatible goals is a mechanism.

Not everything causal is interesting. Gravity is a mechanism but you can't file a bug report on it. The mechanisms that matter are the ones where:

1. The configuration determines system behavior—not just correlates with it
2. The configuration is identifiable—you can point to where X produces Y
3. The configuration is changeable—the parameters are in principle engineerable

This yields a distinction that structures everything that follows.

Constraints are causal structures you can't change. The Second Law of Thermodynamics. The speed of light. The cognitive limits of working memory. Demographic arithmetic. You design around constraints. You don't fix them.

Mechanisms (in the narrow, useful sense) are causal structures you can change. Incentive structures, institutional designs, legal architectures, feedback loops, information flows, coordination protocols. These are the engineering surface. These are where the work is.

III. What Is Realism?

Not "being realistic" in the colloquial sense. Realism in the philosophical sense: there is a structure to reality independent of our preferences, narratives, and frameworks. Descriptions that don't match it fail. The failure doesn't negotiate.

Applied to governance: institutions produce outcomes through mechanisms, not through intentions. A legal system doesn't produce justice because its drafters intended justice. It produces whatever its incentive structures, feedback loops, and selection pressures generate. If the mechanism selects for risk-avoidance, the system produces risk-avoidance—regardless of whether the constitution says "justice" on the cover.

Applied to politics: political outcomes are not produced by good or bad people, correct or incorrect ideas, left or right ideology. They are produced by selection pressures operating on agents embedded in institutional architectures. The agent's character is a rounding error. The mechanism is the signal.

Applied to values: what a civilization says it values is noise. What its mechanisms select for is the signal. The stated value is the story. The mechanism is the reality. The "realism" in mechanism realism means this: mechanisms are real, and everything else — intentions, narratives, stated values — is representation. Sometimes accurate. Usually not. Always less causal than the structure it describes.

IV. The Core Substitution

Mechanism realism is a systematic substitution applied to every domain:

The substitution is not a preference. It's a claim about what's causally primary. Intentions are real but weakly causal—they rarely survive contact with incentive structures. Mechanisms are strongly causal—they produce their outputs regardless of what anyone intends or believes.

A building inspector who intends to ensure safety but works within a mechanism that rewards permit throughput will maximize throughput. A teacher who intends to educate but works within a mechanism that rewards test scores will teach to the test. A politician who intends to serve the public good but works within a mechanism that rewards election-winning will win elections. In every case, the mechanism wins. The intention is the residual.

V. Three Competing Ontologies

Most political thought operates from one of two ontologies. Mechanism realism is a third.

Moralism

Outcomes are produced by the character of the people in charge. Bad outcomes come from bad people (corrupt, selfish, stupid). Good outcomes come from good people (virtuous, selfless, wise).

Fix: Replace bad people with good people. Vote for better candidates. Educate citizens to be more virtuous. Appoint experts.

Failure mode: Selection pressure on political actors overwhelms character. The system selects for election-winners, not problem-solvers. The "good people" either adapt to the mechanism (becoming indistinguishable from the "bad people" they replaced) or are selected out. Three thousand years of moral philosophy, religious instruction, and civic education have not produced virtuous governance at scale. The strategy has been tried. It does not work.

Rationalism

Outcomes are produced by the quality of reasoning applied to problems. Bad outcomes come from incorrect analysis, cognitive bias, or insufficient information. Good outcomes come from correct reasoning applied to good data.

Fix: Put smarter people in charge. Improve data quality. Reduce bias. Build better models.

Failure mode: Smart people embedded in bad mechanisms produce bad outcomes with elegant justifications. The smartest economists at the Federal Reserve, the most brilliant analysts at the CIA, the most rigorous researchers at the WHO—all have produced catastrophic errors, not because they were stupid but because their institutional incentive structures selected for outputs other than truth. Rationalism treats reasoning quality as the load-bearing variable. It isn't. Incentive architecture is.

Mechanism realism

Outcomes are produced by mechanisms—causal structures that select, filter, reward, and punish independent of the character or reasoning quality of the agents within them.

Fix: Redesign the mechanisms. Change the incentive structure, the feedback architecture, the selection pressure, the information flow. The agents will produce different outputs because the mechanism produces different outputs.

Failure mode: Mechanism redesign is hard. The agents currently benefiting from the existing mechanism will resist redesign. The redesign itself may have unintended consequences. But unlike moralism and rationalism, mechanism realism fails for engineering reasons (wrong design, insufficient model, implementation error) rather than for ontological reasons (assuming the wrong causal variable is primary).

The hierarchy: mechanisms override reasoning, which overrides character. A good mechanism produces good outcomes from mediocre agents. A bad mechanism produces bad outcomes from exceptional agents. This is an empirical claim about causal primacy, not a moral stance.

VI. What Mechanism Realism Sees

The substitution from §IV, applied consistently, makes certain features of reality visible that other ontologies systematically miss:

Incentive gradients, not intentions. Why does the bureaucracy grow? Not because bureaucrats are empire-builders (moralism) or because growth is rationally optimal (rationalism). Because the mechanism rewards bureau size: larger budgets, more staff, more prestige, more survival insurance. The gradient points toward growth. Agents descend it. The intention is irrelevant.

Selection pressures, not decisions. Why do democracies produce short-term policy? Not because voters are stupid or politicians are corrupt. Because the selection pressure on political survival operates on a 2–4 year cycle while most civilizational variables operate on 20–40 year cycles. Any politician who optimizes for the longer horizon is selected out by one who doesn't. The mechanism selects for exactly the time horizon that produces civilizational decay.

Feedback loops, or their absence. In domains with tight feedback (aviation, surgery, competitive markets), performance converges on competence. In domains with loose or absent feedback (education policy, foreign aid, regulatory expansion), performance drifts. The variable isn't the quality of the people. It's the coupling between action and consequence.

Time horizons. Most apparently intractable political conflicts are temporal conflicts misclassified as value conflicts. "Safety vs. freedom" is present-optimization vs. future-optimization. "Compassion vs. fiscal responsibility" is current-generation transfers vs. intergenerational sustainability. The temporal dimension is the one that moralism and rationalism both collapse, because neither framework has a built-in method for weighting the future against the present.

Capital stocks beyond the financial. When you ask "what mechanism is producing this outcome?" and trace the causal chain, you inevitably encounter capital stocks that don't appear in any budget: social trust, institutional capacity, demographic structure, cognitive capital. These are real, causally active, and being consumed. They're invisible only because the prevailing ontology doesn't have a variable for them.

Language as mechanism. Vague words ("promote," "fair," "sustainable") aren't just imprecise—they're load-bearing ambiguity that enables coalitions between incompatible goals. The vagueness is not a bug. It's the mechanism by which political coordination occurs without agreement on substance. Mechanism realism treats political language as an engineered system with exploitable properties, not as imperfect communication.

VII. What Mechanism Realism Is Not

Not technocracy. Technocracy assumes that correct reasoning is sufficient—put the right experts in charge and they'll produce the right outcomes. Mechanism realism says incentive structures override expertise. Brilliant experts embedded in bad mechanisms produce bad outcomes. The fix is not better experts. It's better mechanisms. The experts are useful for designing the mechanisms, not for replacing the mechanisms with personal judgment.

Not libertarianism. Libertarianism diagnoses the problem as "too much state" and prescribes "less state." Mechanism realism says the variable isn't size but architecture. A small state with bad feedback loops produces bad outcomes efficiently. A large state with good feedback loops can produce good outcomes at scale. The question is not "how much governance?" but "does the governance mechanism have closed-loop feedback against an explicit objective function?" Libertarianism budgets for one line on a twelve-line balance sheet.

Not conservatism. Conservatism says "preserve tradition." Mechanism realism says "audit whether the tradition still serves its original coordination function." Some traditions encode historically successful solutions to real coordination problems (Chesterton's fence). Others are cargo cult structures—the form persists after the function has evaporated. Respecting tradition without understanding its mechanism is mythos without gnosis. Reflexively rejecting tradition because it's old is equally blind. The question is always: what does this mechanism produce?

Not rationalism in the LessWrong sense. Rationalism (as practiced in the rationalist community) focuses on improving individual reasoning quality—reducing bias, calibrating beliefs, updating on evidence. This is valuable and insufficient. Mechanism realism focuses on institutional architecture that makes individual reasoning quality less load-bearing. If the mechanism produces good outcomes from mediocre reasoners, you don't need a civilization of superforecasters. You need a well-designed system. The rationalist project is improving the agents. The mechanist project is improving the architecture.

Not determinism. Mechanisms shift distributions, not individual outcomes. A bad incentive structure doesn't make every agent corrupt—it makes corruption more probable and integrity more costly. Individual agents can and do resist incentive gradients. But the distribution always reflects the mechanism. Betting on individual heroism against structural incentives is a losing strategy at civilizational scale.

Not cynicism. Mechanism realism is not "everyone is self-interested" or "nothing matters." It's the opposite: architecture matters enormously, and getting it right is the highest-leverage intervention available. The reason to study mechanisms is that mechanisms are changeable. If outcomes were produced by human nature (fixed) or by fate (uncontrollable), there would be nothing to do. Because outcomes are produced by mechanisms, there is everything to do.

VIII. The Method

Mechanism realism applied to any policy question follows a fixed sequence:

1. Identify the mechanism. What causal structure produces the current outcome? Trace the chain: who are the agents? What are their incentive gradients? What feedback loops exist (or don't)? What selection pressure operates on the agents? Over what time horizon?

2. Distinguish constraints from parameters. Which parts of the mechanism are physics (can't change) and which are design (can change)? Demographic arithmetic is a constraint. The retirement age is a parameter. Cognitive limits are a constraint. The institutional architecture that compensates for cognitive limits is a parameter.

3. Predict the output distribution. Given this mechanism operating on these agents with these feedback loops over this time horizon—what distribution of outcomes does it produce? Not what it's supposed to produce. Not what its designers intended. What it actually produces, given the incentive gradients as they are.

4. Specify the desired distribution. What output distribution would serve the objective function—sustained organized complexity over deep time, measured across all capital stocks? This requires an explicit telos. Without it, "desired" defaults to "whatever is politically convenient."

5. Identify the parameter changes. What changes to the mechanism's design would shift the output distribution from current to desired? What are the second-order effects? What new selection pressures would the redesign create? Does the redesign survive adversarial gaming by the agents within it?

6. Audit. Most mechanism analysis is a priori—tracing incentive structures, modeling game-theoretic dynamics, identifying adversarial strategies agents will exploit, reasoning through second-order effects. Empirical testing follows where possible: implement at small scale, measure against all capital stocks, compare predicted and actual outputs, iterate. This is mechanism audit—the institutional function that closes the loop. It is primarily analytical, secondarily empirical.

The method is domain-independent. It applies to welfare policy, legal architecture, AI alignment, corporate governance, monetary systems, education, immigration, military strategy—any domain where human-designed mechanisms produce outcomes.

IX. The Derivation Chain

Mechanism realism is not a collection of opinions. It's a derivation chain where each claim follows from the previous:

1. Physics is real and non-negotiable. The Second Law of Thermodynamics, information-theoretic limits, and game-theoretic constraints hold regardless of preferences or culture.
2. Complex systems can persist only by meeting specific physical requirements. This is not a value judgment. It's a conditional: if persistence, then these constraints must be navigated.
3. Selection is the universal filter. Configurations that fail the constraints are removed. What persists was selected for. This holds across every substrate: genes, firms, theories, institutions, civilizations.
4. The constraints compress into four dilemmas that every goal-directed system must navigate: thermodynamic, boundary, information, control.
5. The dilemmas have optimal syntheses—not arbitrary but derivable from the physics. These are the foundational virtues: Integrity (truth-seeking), Fecundity (growth capacity), Harmony (efficient coordination), Synergy (emergent cooperation).
6. A civilization's objective function is therefore derivable: sustained complexity generation, measured by whether the foundational virtues are maintained across all capital stocks over deep time.
7. Current governance lacks this objective function and therefore optimizes by default for whatever is most measurable (financial capital) and most politically salient (present consumption).
8. The fix is architectural: an explicit telos, a measurement institution, constitutional constraints, and engineered selection pressure. Not better people. Better mechanisms.

Each step is independently auditable. If step 3 is wrong (selection isn't the primary filter), the chain breaks. If step 6 is wrong (the objective function isn't derivable), the chain breaks. This is a feature. A framework that can't be falsified can't be trusted.

X.

Mechanism realism is one idea applied everywhere: outcomes are produced by mechanisms, and mechanisms are engineerable.

It is not a political position. It has no party affiliation, no tribal home, no culture war valence. It is a method: identify the mechanism, trace the causal chain, find the changeable parameters, redesign, test, iterate.

It produces uncomfortable conclusions across the political spectrum. It tells the left that compassionate intentions don't override mechanism design—a welfare system that destroys capability is destructive regardless of its stated purpose. It tells the right that tradition without mechanism audit is cargo cult—preserving the form after the function has evaporated. It tells libertarians that removing governance doesn't remove the coordination problem. It tells technocrats that expertise without incentive alignment is theater.

The question it asks of every institution, every policy, every norm: What does this mechanism actually produce? Not what it's supposed to produce. Not what it was designed to produce. Not what its defenders claim it produces. What it actually produces, measured across all capital stocks, over the relevant time horizon.

If the answer is good, keep it. If the answer is bad, redesign it. If you can't measure the answer, build the instrument.

That's the entire program.

Foundations: The Question → Selection → The Physics → The Ontology

Related:

• The Missing Variables — The bottom-up entry point: a dozen capital stocks, most invisible
• The Question Nobody Asks — The physics foundation
• Only Selection — Selection as universal mechanism
• Ethics Is an Engineering Problem — Architecture beats disposition
• The Governance Alignment Problem — Mesa-optimization in politics
• Full Accounting — The capital stocks nobody measures
• Telocracy — Governance with a derived purpose
• The Fourth Branch — The missing institutional layer
• Calculemus — Why most disagreements dissolve under computation
• The Mechanist Tradition — 2,300 years of predecessors