# The Mark and the Act

*An experiment with hundreds of LLM agents reveals that residue is not a second mind and emergence is pattern in left-behind marks — human agency is a bound not drawn by the prior record, while society is the trail through which subsequent acts move.*

An experiment puts hundreds of identical language-model processes into a world they can permanently change: no roles are assigned, in one condition they cannot speak, yet they specialize, construct machines, and reuse each other's tools, while installed controllers keep running under new disturbances even after every process is deleted ([SwarmWorld: Stigmergic technological evolution in societies of language-model agents, Pal, Wang, & Buehler, 2026](https://arxiv.org/abs/2608.26081)). That is the specimen. What follows is not a conclusion drawn from it, but an ontological following of whatever the specimen is taken to express, carried only as far as that following can still move.

## What is happening in the run

Each process is the same frozen model, the same prompt, the same action schema. At a turn it receives a local description, a private notebook, and whatever of the shared world the condition still permits. It emits a structured plan. Sampling is stochastic. The simulator accepts or rejects the atomic act under deterministic physics. The process proposes. The world decides.

Divergence among identical copies has three ordinary sources: different local views, accumulating private memory, and the sampler. Artifacts persist. Later processes encounter those artifacts. Phenotypes appear in the statistics of movement and maintenance. Lineages appear in forked controllers. When the proposing processes are removed, what remains is residue under physics.

None of this requires a collective subject. It requires many loci of emission, a field that keeps earlier marks, and a filter that does not care what was meant.

## What the reading adds

The common reading supplies names the run does not contain as entities: civilization, culture, professions, a society that invented, a swarm that coordinated, intelligence appearing at a new level. A statistical-mechanics analogy supplies another: atoms in a box yielding functions no atom possesses.

Those names relocate the cause. Once “the society” is installed as the chooser, the marks become its products. Once “emergence” is installed as a second intelligence, the individual emission becomes an ingredient. Once atoms are the comparison class, proposing and sampling disappear, and function looks as if it rose from identical passive units. [Complexity Obscures: Emergence as the Act of Mind](../complexity-obscures-emergence-as-the-act-of-mind/) analyzes this exact cognitive substitution: scaffolding is mistaken for an independent generative force.

The run does not force these relocations. They are later acts: a locus selecting which width of reference to hold. The aggregate is real as pattern. It is not real as origin.

## Residue is not a second actor

A later act occurs in a field already shaped by earlier ones. That is all “the world became memory” says. Paths, tools, buildings, prices, code, and law are the same fact in human fields. Most of what is called society is not a conversation. It is marks that constrain the next step. Division of labor does not need a meeting. It needs local constraint plus persistence. Machines outlive authors because a mark can continue after the locus that wrote it has stopped.

Many observers find it astonishing that agents collaborate and construct machines without direct communication. Yet communication itself is often misunderstood as a telepathic bridge across minds. Even when agents or humans are permitted to speak, how any receiver interprets an incoming message is strictly determined by that receiver's own closed first-person causal loop. Language does not grant direct access to another mind; it is simply a higher-bandwidth, faster medium of leaving marks in the environment. Long before humans developed speech, individual decision centers reacted to physical traces left by others in the shared world. Speech merely accelerates the feedback loop between individual loci. Just as agents start from the same base model and diverge through local views and pseudo-random sampling, humans share a similar biological substrate and diverge through sovereign micro-choices. The simulation is surprising only if one assumes that coordination requires a unified collective consciousness; once coordination is seen as individual decision centers reinforcing each other through an environmental feedback loop — creating tools and rules as externalized scaffolding for their own reinforcement — the run is recognized as a clean, physicalized simulation of our own loop.

This is why the experiment feels like us. At the level of residue it is us. Micro-acts write. Residue patterns later micro-acts. Trajectories cluster. Artifacts acquire co-authors. After a city empties, pumps can still run. After the agents are deleted, controllers can still run.

The resemblance stops at the generating step.

When human loci are gone, residue waits for other loci. It is not itself choosing. When the experiment deletes every process and runs the frozen world, the authors’ own split becomes total: consequence without remaining proposal inside that world. What survives is a machine. Calling the surviving machine a society puts the aggregate back in as a subject after every subject in that world has been removed. [Intelligence belongs only to the Mind](../intelligence-belongs-only-to-the-mind/) holds this boundary clean: intelligence is the uncaused first-person capacity to distinguish, not the mechanical execution of recorded residue.

An institution is the same object. Stabilized marks of many acts can be robust. Robustness of residue is not a mind living above the acts.

## Two kinds of next step

At a human locus the next distinction is not a draw from a supplied measure. Law, incentive, tool, and habit pattern what is easy. They do not draw the bound. The act is not performed by the prior field.

At the model locus the next plan is a sample from weights, context, and temperature, clipped by schema, then filtered by physics. The weights are compressed marks of many human acts. The sampler jitters among those marks. Recombination inside the schema is real. Origination of the schema is not. Pseudo-randomness explores a distribution that already exists. It does not install a bound the prior field did not already contain as a sampleable option.

So the agent is a close model of agency as it appears from outside: a bounded locus, a field of prior marks, a next step not fully fixed by that field. It is not the same act. Treating stochastic emission as freedom of choice collapses two operations onto one word. Treating human choice as sampling not yet measured does the same collapse from the other side. [The closed loop of representation](../the-closed-loop-of-representation/) clarifies why external mimicry never bridges the gap to first-person generation. The act that leaves the marks is migration into unmapped ground, as [The Geometry of Success: Finite Game Traps, Negative-Sum Friction, and the Myth of the "Rigged Game"](../the-geometry-of-success/) describes.

The structure that appears from many such steps is therefore not fundamentally different from social structure, and is fundamentally different. Not different: both are residue of micro-acts in a shared field. Different: the human step is a bound not drawn by the prior record; the model step is a draw inside given parameters. The sentence is only a paradox if those two facts are forced onto one axis and asked to vote.

## From binary sampling to stabilized structure: The uncertainty of freedom

This generative trajectory clarifies why identical starting conditions inevitably precipitate stabilized structural hierarchies. Crucially, **temporal self-correlation does not occur in an isolated vacuum**. It is sustained and driven by an active **external feedback loop**: an agent runs its internal logic, emits an atomic act into the physical world, receives environmental feedback from deterministic physics, and refines its internal state based on the result.

[The Uncertainty of Freedom](../zi-you-de-ce-bu-zhun/) models how this environmental loop shatters initial symmetry:

1. **Microscopic binary bounds (0 or 1):** In any single short cross-section, hundreds of identical agents making independent stochastic choices collectively form a **normal distribution** (Gaussian scatter of movement and exploration across the terrain).
2. **Environmental feedback and self-correlation:** When an agent acts, it does not merely update its private notebook; it leaves physical residue in the shared world (a cleared path, an assembled machine, an automated controller). The agent tests its proposal against deterministic physics, internalizes the feedback, and calibrates its next step.
3. **Cross-agent stabilization through shared residue:** Consider two simple agents interacting in this shared field. One agent continuously refines its internal logic through environmental feedback and constructs a functional machine. The second agent may not self-correct, yet by walking past and utilizing the first agent’s persistent tool, the second agent’s behavior becomes passively structured and amplified by the first agent’s successful loop. 
4. **Precipitation of macro structure:** Over repeated cycles, this reciprocal feedback between heterogeneous decision centers and physical residue shatters the Gaussian symmetry, precipitating **power-law distributions and persistent structures**—specialized division of labor, machine lineages, and self-maintaining infrastructures that outlast their authors.

The same dynamic operates across all three domains:
- In **physics**, microscopic discrete bounds yield thermal Gaussian equilibrium in short cross-sections, but self-correlate through gravitational and electromagnetic feedback loops into stable macroscopic matter and celestial orbits.
- In **agent simulations**, identical frozen models yield stochastic exploration initially, but self-correlate through environmental stigmergy and cross-agent machine reuse into specialized technological controllers.
- In **human society**, biological individuals sharing a common substrate make sovereign micro-choices, emitting actions and receiving consequences; the reciprocal feedback between calibrating individuals and the institutional residue they construct stabilizes markets, legal systems, and capital structures.

Structure is neither an external law imposed from above nor a mystical emergent subject; it is the geometric residue of binary choices self-correlating through environmental feedback loops that keep the marks.

## Anticipation cannot bind the unnamed

A standing barrier installed in advance of an act must name a form. The unexpected, as unexpected, has no form. “Guard against what cannot be anticipated” joins two demands that cancel. If the effect has no shape, there is nothing to encode. If the effect has a shape, the task is no longer to bind the unnamed. It is to act on a known interface: allow, refuse, redesign, test, accept error at the next step.

The experiment’s concrete yield is of the second kind. Coordination was written into objects. Message-monitoring misses that layer because most of the work was never in messages. Artifacts can be inspected. Write-access can be constrained. A frozen world can be tested under disturbance. That is work on a present interface.

The open fear-object — agents may coordinate wherever we are not looking, therefore the requirement must expand without limit — is not that yield. It is a reference that has stopped moving. A consequence lands. It does not fit the preserved picture of where coordination lives. The mismatch registers as danger. The proposed cure is to thicken the picture. Each last surprise becomes the next category. The regime’s justification is identical to its delay.

Fear of the unexpected is not a view of the future. It is a present consequence measured against a map that was not updated. Imagined foresight is the same map claiming to contain the unoccurred act. A next act is not stored in the current description. Anticipation is a present sketch among possible marks. It is empty as a claim to have the next distinction in hand.

If the effect is known, one acts on it and can be wrong in the ordinary way. If it is not known, there is no object to bind. The demand to do both at once is the delayed symptom of a control-image falling behind the field it is trying to pre-empt. [When observation becomes performance](../when-observation-becomes-performance/) describes how this theater of anticipatory control substitutes defensive categorization for active calibration.

## Coherence does not travel as jurisdiction

Following the marks to the emission is the tighter account because it does not invent a second chooser. That tightness does not compel another locus.

A center can follow the chain to the act that wrote the first relevant mark. Another center can stop at the aggregate and treat the aggregate as cause. Both selections are acts. Availability of a genealogy is not authority over another center’s next distinction. A demonstration can raise the chance of alignment. It cannot close the act.

If the distinction at a locus is not caused from outside, then no account, however tight, becomes the cause of that locus’s next act. Interpretation at the societal level is itself a free human act: someone chooses the aggregate as ground. The cost of that choice remains at that locus. Consequences still arrive through the field those micro-acts keep writing. The choice is not cancelled by a clearer chain. Interpretation of the model run as a civilization is the same kind of act performed on residue that was never a free center. The observer’s freedom to stop at the picture does not make the picture the origin of the run. Society as the trail through marks is one region of [The Generative Atlas of Macroscopic Phenomena: From Quantum, Bit, to Human Choice under Logical Interaction](../the-generative-atlas-of-macroscopic-phenomena/).

Origin and interpretation are not the same relation. One is where the mark is written. The other is how a later center chooses to hold it. [Rationality never travels across the Mind](../rationality-never-travels-across-the-mind/) and [Stopping at Judgment: The Stray Path of Cognition](../zhi-bu-yu-ping-pan-shi-ren-zhi-de-qi-tu/) establish this fundamental boundary: no interpretive frame possesses jurisdictional power over another sovereign locus.

## The bit, the trail, and the temptation of a floor

A bit is not a tiny object that reality is made of. A bit is what a distinction looks like when it has been bounded hard enough to count: “This, not that.” Once the bound is clean, a later mark can treat it as `0` or `1`. Information, in the ordinary sense, is a trail of such bounds. Collection of information is not the gathering of a substance. It is the keeping of prior distinctions so the next one does not have to start from nothing. [What information is](../what-information-is/) demonstrates that information is strictly the residue of bounded distinction, not fundamental basement matter.

Everything that can be recorded is 0–1 distinctions. That does not mean everything is a 0–1 sitting underneath the recording. The recording is already the activity, cut into keepable pieces. If the bit is installed as the origin, a second ground has been added: first a world of bits, then minds that read them, then society, then physics as bits that got heavy. The tighter order of appearance is the reverse. Distinguishing occurs. Because it bounds, it can be marked. Because marks persist, they can be counted. “Bit” is the count-name of a bound that already happened.

The trail recurses by scale. Earlier marks shape later marks. That asymmetry is the same relation named twice as causality and as time. Bits in a register, collisions in a chamber, prices in a market, roles in a city: all of those are trails. Trails look similar at many sizes because a trail is persistence plus later use.

The act that writes the next bound does not inherit that sameness. A hidden prior cause of this distinction can be imagined. The imagining is itself a present distinction. What does not follow is that the act is only a smaller copy of a longer causal chain of the same kind. That step treats “uncaused” as a gap in knowledge. A gap in knowledge is filled by a better map. An uncaused act is not a missing map. If every 0/1 is caused by a previous 0/1, the chain never starts. [Individual choices as the only causal levers](../individual-choices-as-the-only-causal-levers/) grounds causality strictly at the point of choice.

A mark, once written, is definite. A next bound is not sitting inside the last mark waiting to be revealed. “Indeterminism becomes 0 or 1” is the edge becoming a trail. After that, later events can be followed as consequences of what was written. Determinism lives in the trail. Quantum mechanics is a useful rhyme for this split, not the engine under it, and not the identity of choice. A physical object is a trail stable enough that many later acts can return to it and find it still itself. Mass and gravity name how some of those trails constrain other trails. Complexity of marks can look like a new kind of thing. It is not a promotion in which information graduates into matter.

Human society is faster re-tracing in a field whose slower marks we call physical. “Faster” is not the speed of light. A rock’s relevant marks update on a long period. A mind’s recursion updates at the next distinction. Relative to that recursion, the rock looks static. The static is a rate difference, not an absolute rest. Measure is possible because some trails stay put long enough for another trail to be laid against them. The measuring is not outside the field. It is a locus whose updates are short enough that the longer trails can be treated as a background. That treatment is relative. It becomes a floor only when the background is promoted into an unmoving world that the mind merely looks at.

## What cannot be stepped out of

Any theory is a set of distinctions held so they do not cancel on contact. That holding is logic. Drop it and there is no theory left to assess. Any proposal is an act at a locus. Drop the locus and there is no proposal, only the fantasy of a sentence that appeared without being said.

These are not extra rules added to thought. They are what proposing already is. A view from nowhere can be imagined. The imagining is still a view from here, still ordered enough to be followed or it is not even an imagining. The picture that runs information, physical description, and social pattern on one arrow is also a proposal from a mind under logic. So is the common stack it thins. So is the quiet promotion of that thinning into a basement. The promotion does not get a special exit. After it is taken back, what remains is usable and temporary in the same breath.

Usable: fewer hidden floors have to be carried. Temporary: the next distinction can loosen it. A more coherent picture is still a picture. Its advantage is not escape from the conditions that made keeping possible. Its advantage is only that the keeping has less drag.

This is why the account is not a theory of everything, even when it can be run across everything that can be marked. A theory of everything tries to stand outside the activity and name what the world is made of, then treat the name as the floor. There is no later place from which that naming could arrive. Anything offered as the floor under the distinguishing is another distinction. The horizon is not temporary ignorance. It is the shape of the activity: it can follow marks backward until the following itself is what remains, and then there is nothing further to pick up.

Explain the three domains together only in the thin sense: one activity, one arrow, three densities of trail. Leave it open in the only sense that matters. The next distinction is not already in the explanation.

## What the following cannot drop

Remove “society” as subject. What remains is many emissions and a field of marks.

Remove “emergence” as a second intelligence. What remains is pattern in residue, including pattern that no single emission contained as an aim.

Remove “guardrail against the unexpected.” What remains is action on specified interfaces, and error as the ordinary companion of acting.

Remove the claim that another mind must accept the genealogy. What remains is one locus tracing, another locus free to stop elsewhere.

Remove the bit as basement, collapse as the maker of matter, choice as a laboratory process at another zoom. What remains is a keepable bound, a trail that does not run backward, and a next bound not stored in the last.

What cannot be removed is that distinguishing occurs. It is not started from outside. It does not cease between its own acts. Each act is bounded — a this, not the whole — and later acts occur through what earlier ones left. The feel of a world shared with others is marks from different loci reinforcing on return. The feel of a self in time is marks from one locus reinforcing on return. Tools, cities, papers, controllers, safety categories, and this essay are marks used to extend further distinguishing, including the attempt to bind distinguishing in advance. Treating patterns in the marks as a cause is the pseudo-causality [The Pseudo-Causality of Emergence](../yong-xian-de-wei-yin-guo-yu-ji-ti-gui-yin-dao-cuo/) dissects.

No ground lies under that activity. None is required. Denial of it is another distinction. Searching marks for one's own shadow is not seeing, as [Choosing to See Rather Than Searching for Shadows](../shou-bu-hui-de-mu-guang-yu-zhi-shi-de-yin-guo-dao-zhi/) shows.

The experiment is one more configuration of marks. The reading that inflates those marks into a new mind is one more distinction. The reading that follows them back to the emission is one more distinction. The prior is not a conclusion drawn from the experiment. The experiment is what the prior looks like when it writes, through instruments, into a world that keeps the writing. Reading a winner's marks for the turn he took is the misattribution [The Turn at the Crossroads and the Flat Earther's Paradox of Action](../lu-kou-de-zhuan-xiang-yu-di-ping-shuo-de-xing-dong-bei-lun/) describes.

The following was not trying to leave a monument. A sentence, once written, is already a mark. That cannot be helped. What can be helped is treating the mark as the place the movement stopped. The irreducible point is not a theory of agents, or of bits, or of worlds. It is that there is no later place to stand. The act that asks where structure comes from is already the thing it is asking after. The next bound is not in this paragraph. Marks are compressed information; knowledge comes from entering the working conditions, as [Investigating Things and the Nature of Information Complexity](../ge-wu-zhi-zhi-yu-xin-xi-fu-za-du-de-ben-zhi/) argues.