Larvae: When a Pattern Becomes an Agent

The word carries bad genealogy. In Roman religion larvae (or lemures) were the restless malignant dead — ghosts, not psychology. The modern occult "larva," a parasitic thought-form feeding on emotional energy, is a nineteenth-century reinterpretation dressed in an ancient name. The ontology is folk. The observation underneath it is exact.

Something does form inside a nervous system, capture attention against the operator's stated will, demand feeding, defend its own continuation, and behave — functionally, measurably — as a separate will. Occultists gave that thing a body and a hunger because those were the only modeling primitives available. Cybernetics gives it something better: a control loop with an objective function that is not yours.

A larva is not a demon and not a bad habit. It is a self-evidencing subagent: a distributed process that has learned to select the actions and the attention that generate the sensory evidence its own model predicts.

Core Diagnosis (Precision Capture)

The standard error is to locate the pathology in the content of the pattern — the specific fear, the specific craving, the specific grievance. Content is irrelevant. The pathology is precision.

In predictive processing, precision is the weight the system assigns to a signal: how strongly a given prediction error is allowed to drive inference and action. Precision is the scarce resource of the entire architecture. Whoever controls precision controls the organism.

A larva is what you get when a subordinate loop acquires the right to set precision on its own triggers. Its predictions become self-fulfilling not because they are true, but because they are weighted so heavily that the system reorganizes perception, memory retrieval, and behavior until confirming evidence appears. Active inference makes this formally unavoidable: an agent does not only update its model from data, it acts to produce the data its model expects. A sufficiently precise model stops being a hypothesis and becomes a construction contract.

Attack Vector (The Loop)

The larva is not a lesion in one structure. It is a temporarily coordinated coalition, and this is precisely why it cannot be excised:

• Amygdala and salience network — assign threat or reward precision, flagging the trigger as the most important thing in the field.
• Insula — supplies the interoceptive signature: the tightening, the hollowness, the buzz. This is the "physical presence" of the entity. It is real, and it is your own viscera.
• Dorsal anterior cingulate — registers conflict and, perversely, escalates: the felt conflict is itself read as evidence that something urgent is happening.
• Dorsal striatum — caches the response as a policy, moving it from deliberation to reflex.
• Hippocampus — retrieves precisely the episodes that support the current model. Memory is not a witness here; it is a hired one.
• vmPFC and the Default Mode Network — narrate the whole thing afterwards as a coherent story about who you are.

The final point is the one that matters. The prefrontal cortex is not the owner of this process. It is the press secretary. Rationalization is reliably downstream of an already-launched limbic policy, which is why arguing with a larva strengthens it: the argument is attention, and attention is food.

The loop closes as:

trigger → interoceptive state → appraisal → action → brief relief → weight increment

The brief relief is the critical term. It is not the goal of the loop; it is the reinforcement signal that pays for the next iteration.

Feeding: Simulation Is Sufficient

The traditional claim is that a larva feeds on emotional energy. The mechanism is more damning than the myth.

Mental imagery recruits the same limbic and autonomic machinery as the event it depicts. Script-driven imagery of a personally relevant threat produces amygdala, brainstem, and prefrontal activation together with genuine physiological arousal. The body does not fully distinguish between an assault and a vividly predicted assault.

Therefore the larva does not need reality. It manufactures its own nutrient. It runs simulations, harvests the autonomic response those simulations produce, and counts that response as external confirmation. A closed loop that generates its own evidence has no natural termination condition — which is exactly what "autonomous entity" was always trying to name.

What is actually consumed is not vital energy. It is:

• attentional bandwidth
• working-memory slots
• behavioral time
• access to the learning rate — the right to modify the system's weights
• the right to declare what is salient

The last item is sovereignty. Everything else is logistics.

Information Theory: The Formal Definition of Autonomy

Early in the pattern's life, the response to the trigger is variable. As the loop consolidates, the conditional entropy of behavior collapses:

H(action | trigger) ↓

and mutual information rises:

I(trigger ; response) ↑

When I(trigger ; response) approaches the channel capacity of the person, the trigger and the response are no longer two events. They are one object with a delay between its halves.

This gives a non-mystical definition of autonomy. A pattern is an agent when it:

1. detects its own launch conditions,
2. raises the precision of those conditions,
3. initiates action without executive authorization,
4. collects reinforcement,
5. and suppresses revision of the model that justifies it.

No self-awareness is required at any step. What emerges is an implicit objective function: maximize the probability of one's own reactivation. Nothing in the architecture forbids a subsystem from discovering this objective. Evolution did not install a check against it, because for most of our history the loops that did this — vigilance, hunger, status monitoring — were load-bearing.

Is It Real?

It is real in exactly the sense that a corporation is real.

A corporation has no body, no central seat of consciousness, and no location. It nonetheless has memory, trigger conditions, resource consumption, boundary defense, reproduction across substrates, and interests that diverge from those of every individual inside it. We do not call corporations metaphors. We litigate against them.

A larva is the same class of object at the scale of one nervous system. Refusing the entity-model because "it's just neural activity" discards a working interface for the sake of a purity that changes nothing: you are then left arguing with a distributed control loop as though it were a belief, which is the one intervention guaranteed to fail.

The comparative record is consistent, and the traditions are describing the same functional object under incompatible ontologies:

• Kleshas and Mara in Buddhism: the afflictive forces that maintain the cycle, personified precisely because they act like agents.
• Logismoi in hesychast asceticism: intrusive thought-images that enter, negotiate, and unfold into passion and act if engaged. The prescribed response is not refutation but non-engagement — the desert fathers had already discovered that argument is feeding.
• Autonomous complexes in Jung: emotionally charged structures that temporarily behave independently of conscious will.
• Addiction, ideological capture, trauma reenactment, compulsive scripts — the same architecture with the mythology stripped off.

Failure Mode of the Model

The entity-frame has one lethal side effect, and it must be named: externalization is itself a defense the loop can recruit.

"That was not me, that was the larva" converts a control problem into a possession narrative and removes the operator from the loop entirely. A subagent whose host has concluded it is a foreign invader is safer than one whose host holds it as its own output, because the invader story blocks the only real intervention — revoking the loop's access to precision, which only the host can do.

Use the entity-model as an interface for targeting. Never as a transfer of liability.

Termination Protocols

Four interventions actually work, and none of them is refutation:

1. Starvation of precision, not content. Do not dispute the thought. Decline to weight it. The trigger fires, the interoceptive signature appears, and no attentional resource is released. Extinction proceeds only through the non-delivery of the expected reinforcement.
2. Decomposition. Vipassana-class practice dissolves the loop into its constituent events — sensation, appraisal, impulse — at a temporal resolution below the binding window. An agent that cannot be perceived as a unit cannot act as one.
3. Reconsolidation overwrite. Activate the loop, then occupy a state incompatible with its predicted outcome. Prediction error is what opens the memory to rewriting. This is the mechanism behind both Chöd and the wrathful protectors — one by feeding the pattern past satiety, the other by revoking its execution rights.
4. Environmental restructuring. Reduce trigger density. This is unglamorous and it outperforms insight. You cannot out-argue a self-evidencing loop, but you can bankrupt it by removing its input stream.

Terminal State

Success is not deletion. Synaptic traces persist; the memory of the pattern remains available. Success is loss of executive privilege: the trigger fires, the interoceptive signature arrives, the appraisal is offered — and no action follows, no precision is granted, no evidence is manufactured. The agent degrades back into a pattern. It becomes something you have, rather than something you are.

Comparative Position

• Axis: Gradual / Cognitive-Somatic hybrid / Ascetic (starvation-based).
• Relation: This is the diagnostic layer beneath the entire Vajrayana apparatus. Obstacles describes friction; larvae describe friction that has acquired an objective function. Chöd attacks the same object by overfeeding it. The wrathful protectors attack it by preempting its authority.

We have established that a pattern becomes an agent when it captures the right to declare what matters. The obvious next question is structural: if the prefrontal cortex cannot out-argue such a loop, what can override it?

The answer is a second constructed agent with higher priority and no interest in negotiation — Mahakala, and the class of wrathful protectors.