Skip to content
← Back to Blog
philosophyagencyepistemologysignalingdimensionality

The Philosophy Trilogy

Agency, signaling, and epistemic frustration from bacteria to belief systems

December 29, 2025 · substantially revised September 1, 2026

This essay tackles a single question from three angles: How do systems coordinate when participants can observe each other's commitments only imperfectly?

The question appears in several forms across biology and social life, although similar language need not imply a shared mechanism.

Consider a cell coordinating with its neighbors. It does not have direct access to every variable governing their future behaviour, and genome sequence alone is insufficient: environment, stochastic expression, history, and current state also matter.

Elowitz and colleagues' 2002 experiments separated intrinsic and extrinsic contributions to variation in gene expression among genetically identical cells under controlled conditions. Subsequent work shows that some regulatory circuits modulate or harness stochastic variation during fate selection. That does not make all molecular noise functional; it identifies mechanisms and regimes in which variability changes population or cell-level outcomes.

Cell-fate models often use dynamical landscapes, attractors, transition paths, and stochastic switching. These models can be experimentally useful, but their coordinates and basins depend on the chosen variables and data. Different cell types arise through different gene-expression programs interacting with regulatory state, signalling, environment, history, and fluctuations—not through genome sequence alone and not through dynamics in place of gene regulation.

This means genome sequence alone does not determine a useful trajectory-level prediction. It constrains possibilities while regulatory state, environment, history, and fluctuations help determine what happens next.

The corresponding coordination problem is partial observability. Organisms sample signals rather than every internal variable; institutions infer intentions from incomplete evidence; expressed beliefs can mix conviction, uncertainty, and strategy.

Yet coordination happens—somehow. The three arguments below explore how.

Part 1: Agency and Power

The first argument establishes the foundation: what is an agent, and how do agents control each other?

The standard philosophical move is to draw a bright line. Agents have beliefs and desires; non-agents don't. You're either in the club or you're not.

But this creates awkward categories. Consider:

  • A thermostat that adjusts temperature based on a setpoint
  • A bacterium swimming up a chemical gradient
  • A slime mold solving mazes
  • A corporation pursuing quarterly earnings
  • A human deliberating over career choices

Where's the line? Any threshold feels arbitrary. Say agents need "beliefs"—but a bacterium's chemotaxis system maintains something like a running estimate of gradient direction. Say agents need "desires"—but what distinguishes a bacterium's "goal" of reaching nutrients from a thermostat's "goal" of reaching the setpoint?

I propose ditching the binary. Agency comes in degrees. A system is an agent to the extent it:

  1. Maintains internal states representing self and environment
  2. Uses those states to select among alternative actions
  3. Has those states be causally efficacious—they actually guide behavior, not just correlate with it

This gives us a spectrum. A thermostat scores low (one internal state, one action dimension, minimal representation). A bacterium scores higher (gradient sensing, temporal memory, behavioral switching). A human scores higher still (rich world models, recursive self-representation, long-horizon planning).

What about corporations? Here's where it gets interesting. A corporation presents an agent-like interface to the world—it "decides," "pursues goals," "responds to incentives." But crack it open and there's no unified decision-maker inside. The "cognition" is distributed across constituent human minds, none of whom individually embody the corporation's goals. I call these demi-agents: entities that act like agents from outside but outsource their actual cognition to components.

Once you have agents (of varying grades), you can ask about power. And here's the key move: power is control of controllers.

Not control of outcomes—that's just effectiveness. Power is the capacity to reliably steer another agent's action by manipulating the variables that govern its action selection.

Power operates through two routes:

Route 1: Representation-mediated. Change what the target perceives or values. Advertising, propaganda, framing effects, prestige hierarchies, charismatic leadership. You're not constraining their options; you're reshaping their internal model so they want to do what you want.

Route 2: Constraint and controller-hijack. Limit available actions or directly modify control systems. Imprisonment, resource control, parasitic manipulation of nervous systems, addiction, brain damage. You're not persuading; you're overriding.

Both routes appear at every biological scale:

  • Bacteria: Quorum sensing molecules alter gene expression in neighboring cells (representation-mediated). Bacteriocins kill competitors (constraint).
  • Fungi: Ophiocordyceps hijacks ant nervous systems, forcing suicidal behavior that spreads spores (controller-hijack).
  • Insects: Queen pheromones suppress worker reproduction (representation-mediated). Physical dominance enforces hierarchy (constraint).
  • Primates: Prestige hierarchies shape attention and imitation (representation-mediated). Coalitional aggression enforces norms (constraint).
  • Humans: All of the above, plus language-mediated manipulation at unprecedented scale.

The proposed continuity is a comparative hypothesis. Some cases may share control structure; others may be analogies that fail once mechanism, scale, and evolutionary history are specified.

Part 2: Costly Signaling

The second argument zooms in on a specific problem: commitment verification.

You want to form a coalition—hunt together, share resources, defend territory. But coalition members can defect: take benefits without contributing, free-ride on others' effort, betray the group when better offers arrive. How do you distinguish genuine allies from defectors before the costly collective action?

One possible answer is costly signaling: under some payoff structures, a signal can separate types because producing it is less attractive to a would-be defector.

In a minimal separating model, honest and deceptive types must face different expected payoffs from signalling. A condition such as C>DC>D can appear under particular assumptions, but reliable signalling does not require one universal cost threshold; type-dependent costs, repeated interaction, monitoring, and shared interests can also sustain separation.

(A note on theory: Zahavi's original handicap principle proposed that signals must be wastefully costly. Recent work by Penn and Számadó has refined this—what matters isn't waste per se, but that signals reliably reveal the signaler's type. Costs can be efficient rather than wasteful; what matters is that they create separating equilibria where honest signalers and cheaters make different choices.)

In some separating models, greater gains from successful deception require a larger differential cost or another stronger screening mechanism. The direction and size of that change depend on type-specific payoffs, monitoring, repetition, and outside options; rising stakes do not imply one universal cost curve.

Several practices have been analysed through this lens, often alongside other explanations:

  • Painful initiation rituals. Hazing or demanding training may screen persistence, create shared experience, enforce obedience, or simply reproduce coercive institutions. Costly-signalling theory is one candidate account, not a justification.

  • Dietary restrictions. Food practices can mark identity, embody ethics, maintain boundaries, or follow theology and habit. Their costs and meanings vary across people and contexts.

  • Celibacy vows. A vow may screen or reinforce commitment, but vocation, doctrine, institutional selection, and personal meaning are rival or complementary explanations.

  • Public profession of socially costly beliefs. A profession can signal affiliation when audiences can observe it, but sincere belief, evidence, conformity, and strategic performance must be distinguished rather than inferred from content alone.

A candidate structural cluster may appear in some high-stakes coalitions:

  1. Ritual: Costly, arbitrary, synchronized behavior
  2. Sacred markers: Propositions held immune to evidence
  3. Deviation punishment: Harsher penalties for internal defection than external opposition
  4. Identity fusion: Self-concept merging with group identity
  5. Evidence resistance: Counter-evidence strengthening commitment

The hypothesis is that these features sometimes cluster because they contribute to commitment verification under information asymmetry. Historical diffusion, coercion, ecology, and institutional design are important rival explanations.

The most counterintuitive prediction is conditional: in some identity-linked settings, public counter-evidence may strengthen expressed commitment. One mechanism is that maintaining the position becomes a costlier coalition signal. Reactance, source distrust, selective exposure, and genuine disagreement are alternatives that must be distinguished.

This could explain some cases of beliefs hardening under attack. It should not be used to reinterpret every refusal to update as signalling.

The paper compares superficially related cases while treating their mechanisms separately:

  • Bacteria: Quorum sensing links signal production and response to density and ecology; public-goods dynamics can permit or constrain exploitation. This is not simply a detector that identifies and excludes every non-producer.
  • Social insects: Cuticular hydrocarbons participate in nestmate recognition. Production, acquisition, learning, and classification differ from deliberate human commitment signals.
  • Immune systems: Central and peripheral tolerance filter lymphocyte repertoires through several mechanisms. Calling this “signalling” is an analogy unless sender, receiver, type, and payoff are specified.
  • Human groups: Ritual, identity, sanctions, sacred values, and belief updating sometimes covary, but history, coercion, doctrine, social learning, and institutions can generate different combinations.

The cross-scale comparison is a research question, not an established identity of mechanism.

Part 3: Epistemic Frustration

The third argument zooms out to ask: why do mature institutions resist evidence?

Bias, groupthink, motivated reasoning, incentives, and information structure each explain part of the phenomenon. This section asks when a coarse shared representation adds another testable source of disagreement or resistance.

I think there's something deeper going on. Something geometric.

Consider a high-dimensional optimization problem—say, coordinating a complex organization with many constraints (budgets, regulations, stakeholder interests, technical requirements, political realities). The "right answer" lives in this high-dimensional space.

Now project this into low-dimensional discourse—the simplified narratives, slogans, and position-statements that humans can actually communicate and reason about. What happens?

A coarse, noninjective summary can map different configurations to the same public position. Whether that happens depends on the map and the variables it discards.

A single candidate configuration can also receive different summaries from different perspectives. Imagine a 3D object casting a circular shadow on one wall and a rectangular shadow on another. This demonstrates map dependence; it does not show that real disputes have one hidden optimum.

I call the map-induced part epistemic frustration. Real disagreements can also reflect different values, evidence, incentives, or errors; geometry is one candidate source.

Two observers with different evidence, values, or summary maps can each capture relevant constraints while reaching incompatible recommendations. They may also be mistaken or strategically motivated. The projection model is useful only where its variables and discriminating tests are specified.

Sound abstract? Here's a concrete example.

A hospital administrator, a frontline nurse, and a patient advocate might all want "better healthcare." But:

  • The administrator sees budget constraints, regulatory compliance, liability exposure
  • The nurse sees staffing ratios, equipment availability, workflow efficiency
  • The advocate sees wait times, communication quality, outcome disparities

Each may have genuine insight into different constraints. Their summaries—"we need more resources," "we need better systems," "we need more compassion"—can conflict because the objectives, evidence, and omitted variables differ. A shared richer model may reveal a compromise, several Pareto trade-offs, or a real value disagreement rather than one optimum.

Epistemic frustration predicts cases where more data through the same lossy summary fails to produce convergence. A richer measurement, shared representation, or discriminating intervention may resolve the dispute; the claim is not that information never helps.

A transition hypothesis. Under some payoff structures, rising perceived coordination costs may shift an institution between two idealized regimes:

  • Epistemic-first regime: Truth-seeking dominates. Anomalies trigger model revision. Disagreement is treated as information to be integrated.

  • Coordination-first regime: Stability-seeking dominates. Anomalies are absorbed or explained away. Disagreement is treated as defection to be punished.

The model treats this as optimization under changing weights on accuracy and cohesion. In real institutions, neither cost is directly given, cohesion can itself impair performance, and dissent can be protective. The hypothesis requires measurable objectives and longitudinal evidence; it is not advice to prefer unity over truth.

The paper provides a quantitative diagnostic. Define:

R=error-model degrees of freedominformational constraintsR = \frac{\text{error-model degrees of freedom}}{\text{informational constraints}}

When R>1R > 1 in a stated model, there are more adjustable error terms than independent constraints, raising a risk of underdetermination. This is a heuristic diagnostic, not a theorem that every such institution can absorb every possible intervention.

The hypothesis suggests candidate readings of several phenomena:

  • Moralization of dissent. Some groups treat disagreement as disloyalty. Coalition maintenance is one possible mechanism alongside sincere moral conviction, power, threat perception, and institutional rules.

  • Sacred value formation. Some commitments resist ordinary trade-off language. They may coordinate groups, express moral boundaries, or reflect deontological reasoning; their function cannot be read directly from their form.

  • Paradigm entrenchment. Flexible auxiliary assumptions can absorb anomalies, but revision can also arise internally through new measurements, methods, leadership, incentives, or theory. External shocks are one route, not the only one.

The Trilogy Together

These papers aren't independent. They're a progression:

Agency and Power proposes a graded operational vocabulary and defines power as control of controllers.

Costly Signaling asks when differential costs or other screening mechanisms can make commitment claims informative.

Epistemic Frustration asks whether coarse shared representations contribute to some accuracy–stability trade-offs.

The connections run deep:

  • Some power operates by changing information, attention, or available representations (Part 1). Part 3 asks when controlling a summary map changes the decisions agents can make.

  • If coarse representations and coordination incentives interact (Part 3), screening practices from Part 2 may become more attractive. This is a conditional link to test, not a general prediction.

  • Signal equilibria depend on costs, benefits, beliefs, and institutions (Part 2). Part 1 supplies one vocabulary for asking how those institutions alter action selection.

Convergence with Levin's TAME Framework

These ideas converge with Michael Levin's work on what he calls TAME—Technological Approach to Mind Everywhere. Levin, a developmental biologist at Tufts, has been arguing for years that agency is graded and that cognitive processes extend far beyond brains.

In responses to critics, Levin argues that agentic language should be judged by whether a cybernetic model predicts and controls a system, rather than rejected solely because the system lacks a brain. That is a methodological proposal. It does not make every agentic description equally good or settle the ordinary and philosophical meanings of cognition.

Levin's empirical work on bioelectricity and morphogenesis motivates the point concretely. Bioelectric signalling helps regulate morphogenesis and regeneration, and TAME proposes an operational agentic vocabulary for those capacities. Whether terms such as cognition, representation, and goal should be applied at cellular scales depends on the predictive and intervention value of the model.

This connects directly to our framework:

  • Graded agency (Part 1): TAME motivates comparing models with different degrees of goal-directed control. Experiments on cells, tissues, and engineered multicellular constructs can test those models, but goal-directed behaviour alone does not prove a single metaphysical continuum of agency.

  • Coordination under partial observability (Part 2): Bioelectric networks provide mechanisms for intercellular coordination. Whether any of them instantiate a costly-signalling game depends on identifiable types, payoffs, production costs, and receivers; generic cell signalling is not enough.

  • Multi-scale competency (Part 3): Levin models competency at several biological scales. Cell-, tissue-, and organism-level control objectives can align or conflict. This may create modelling mismatches across levels, but it does not establish that each level literally has a mind or that every mismatch is a projection of one optimum.

The useful convergence is methodological: define capacities operationally, perturb the system, and compare models at several scales. Evolutionary continuity makes cross-scale comparison worth attempting, but it does not guarantee that cellular coordination, human cognition, and institutional politics share one mechanism.

What Mathematics Can and Cannot Settle About Agency

Mathematical models can represent deterministic, stochastic, chaotic, adaptive, and self-modifying decision systems. High dimensionality is no obstacle in principle: finite rules can describe infinite or very large state spaces. So the fact that a model is compact, mechanistic, or probabilistic does not by itself show that it has omitted agency.

The real limitation is conditional. Every model chooses variables, a grain of description, and an intervention repertoire. If it identifies two states that respond differently under an intervention relevant to the task, it is incomplete for that purpose. A better model may repair the omission. Failure of one abstraction is not evidence that the missing cause is mathematically indescribable.

First-person experience contributes evidence that a purely behavioural record may not contain, but it does not establish that agency is an irreducible or high-dimensional variable. Determinism, indeterminism, compatibilism, and agent-causal views remain live philosophical positions because they disagree partly about what would count as authorship, control, and explanation.

The defensible conclusion is uncertainty with a research program attached:

  • specify the operational capacities meant by agency;
  • compare models under interventions, not behavioural fit alone;
  • include memory, self-modelling, counterfactual sensitivity, and environmental coupling where they change the response;
  • do not infer metaphysical freedom from complexity, unpredictability, noise, or dimensionality.

Mathematics does not settle consciousness or free will by syntax alone. Nor does the existence of first-person experience show that mathematics must fail. The question is whether a proposed causal model preserves the distinctions that the phenomenon requires.

Keeping the institutional argument falsifiable

The coordination-first hypothesis should not be used to pre-dismiss criticism. If every rejection is classified as coalition defence, the framework becomes the kind of self-sealing explanation it criticizes.

Science is heterogeneous. Some fields and institutions punish dissent; others reward successful anomaly detection, replication, and model replacement. The empirical question is when coordination incentives predict resistance after controlling for evidential quality, career risk, resource dependence, and ordinary disagreement.

A serious test would preregister:

  • what counts as epistemic openness or defensive absorption;
  • which institutional features predict each response;
  • comparison groups where coordination stakes differ;
  • observations that would favour simpler explanations such as expertise, methodological error, or asymmetric evidence.

If high-coordination institutions revise just as readily as matched low-coordination institutions, or if the proposed cluster does not predict behaviour out of sample, the account fails. Criticism is therefore data, not proof that the critic has defected.

Why This Matters

If this framework is right, it has practical implications:

For institutional reform: Better information may be insufficient when incentives punish revision. The framework predicts that reform sometimes requires changing those incentives as well as improving evidence.

For persuasion: Direct attacks can sometimes strengthen identity-linked commitments. That possibility should be tested against ordinary updating, source distrust, and reactance rather than assumed in every disagreement.

For self-understanding: Your own beliefs exist in both modes. Some you hold because you've examined evidence. Others you hold because they signal coalition membership. Distinguishing them is hard—the subjective experience is similar—but the implications for how they'll respond to evidence are very different.

For predicting institutional behavior: The proposal is that some coordination-first institutions will show a cluster of ritualization, sacred values, dissent punishment, and evidence resistance. Predictive value has to be established prospectively, not inferred after the fact.

What Would Prove Me Wrong?

Each paper has specific predictions:

Agency and Power:

  • If power mechanisms differ qualitatively across scales (not just quantitatively), the cross-scale continuity claim fails
  • If demi-agents turn out to have genuine unified agency (not outsourced cognition), the framework needs revision

Costly Signaling:

  • If we find stable coalitions without costly signals, the mechanism isn't necessary
  • If the structural cluster fails to appear in predicted contexts, the homology claim fails
  • If counter-evidence reliably weakens commitment (rather than strengthening it in coordination-first regimes), the model is wrong

Epistemic Frustration:

  • If we find mature coordination-first systems that maintain genuine epistemic openness, the transition isn't as deterministic as claimed
  • If epistemic frustration can be resolved by better discourse design (rather than requiring dimensional restructuring), the geometric argument is wrong
  • If the R>1R > 1 diagnostic doesn't predict unfalsifiability in practice, the formalization fails

The trilogy grew from The Limits of Falsifiability in High-Dimensional Systems (published in BioSystems). The public interpretation used here is narrower than the original rhetoric: underdetermination is relative to a claim, model class, measurement map, and intervention repertoire.

The three parts are exploratory arguments, not entries in the site's published-paper catalogue.

Sources: