23: What Does One Attachment Cause?

At the end of Chapter 22, the Digital Crystal had stopped looking like a simple bounded thing.

We had looked for a privileged region whose future belonged especially to itself.
We had not found one.

What survived instead was locality.

Change something here, and the consequences are stronger nearby than far away.

That left a new possibility.

Perhaps the important object was not the crystal as a body.

Perhaps it was the process occurring across its changing interface.

The obvious next question seemed to be:

Does that process move?

That question would consume most of this chapter.

It would also turn out to be the wrong one.

Not useless.

Wrong in a productive way.

Because after four experiments, the Digital Crystal gave us something more precise than motion.

It gave us a causal quantity.

One additional attachment can change what happens next.

The question is how much.

From Boundary to Process

The previous chapters had gradually changed what counted as interesting.

Chapter 18 found that persistent state mattered only while it remained coupled to active construction.
Chapter 20 found that loss manufactured new construction opportunity.
Chapter 21 found that finite computation rationed which opportunities could even be evaluated.
Chapter 22 found spatial causal locality, but no privileged enclosing boundary.

A common object was beginning to appear:

CURRENT MATERIAL
↓
ACTIVE CONSTRUCTION OPPORTUNITIES
↓
EVALUATION
↓
ATTACHMENT OR LOSS
↓
NEW MATERIAL
↓
NEW OPPORTUNITIES

The material mattered.

But mostly because it changed what could happen next.

    flowchart LR
    A[Current material] --> B[Active construction opportunities]
    B --> C[Finite evaluation]
    C --> D[Attachment or loss]
    D --> E[New material]
    E --> F[New opportunities]
    F --> B
  

So Chapter 23 began with a process-first hypothesis.

If activity was genuinely propagating through the crystal, then activity near one place at time t should predict activity farther away at later times.

Not just:

nearby things happen near nearby things

but something more structured:

near distance
→ early lag

farther distance
→ later lag

A moving ridge through space-time.

That was Experiment V1.

V1 — Does Active Process Propagate?

We defined an active process field from material-changing events.

The first version pooled:

ATTACHMENT
+
LOSS

into a single material-event field.

For an event at position x and time t, we measured future event density at distance d and lag tau.

Each event was compared with a matched non-event location sharing local geometric context.

A second control compared the source run against future event fields from another run at the same relative stage.

That cross-run null mattered.

If ordinary crystal development itself created a distance-lag trend, then an apparent travelling ridge could exist even without continuity from the original event.

The primary statistic looked for a shift in the lag at which positive excess activity was concentrated as distance increased.

The experiment failed.

The real ridge moved too little.

Its displacement over the cross-run null was too small.

The paired test did not clear the frozen significance gate.

Yet two broad shape statistics passed.

For a moment, that looked tantalizing.

Distance and ridge lag were positively associated.

Perhaps a weak process was travelling after all.

Then we looked at the surface.

The Estimator Had Manufactured a Story

At distance one, the real event field showed a small positive excess at the first lag and negative excess afterwards.

The positive-weighted centroid therefore collapsed to:

lag = 1

At farther distances, where the surface was mostly weak noise, the same positive weighting produced centroids near the middle of the lag grid.

The resulting statistic looked approximately like:

distance 1
→ early lag

far distance
→ later lag

But there was no moving ridge.

There was:

one strongly anchored near-distance row
+
far-field noise centred by the estimator

Had we frozen a slightly weaker threshold, this could have passed.

We might then have written a chapter about propagation that was actually caused by the geometry of the estimator.

That was the most important result of V1.

Not propagation.

A methodological warning:

An estimator can manufacture the shape of the phenomenon it was designed to detect.

The strongest signal in the entire V1 surface was not positive.

It was a persistent negative band at short range.

And the positive-weighted ridge statistic had discarded it before inference.

So we did not tune V1.

We closed the propagation claim as failed.

Then we asked what that signed short-range structure meant.

    flowchart TD
    A[Event at x, t] --> B[Measure future event density by distance and lag]
    B --> C{Positive-weighted lag centroid moves with distance?}
    C -- Yes --> D[Travelling ridge]
    C -- No --> E[No propagation]
    E --> F[Check estimator: positive weighting<br/>can manufacture apparent movement]
    F --> G[Short-range signed structure remains]
  

V2 — Is the Interface a Source/Sink Field?

One tempting interpretation was that attachment consumed local interface while loss created it.

That gave a simple physical picture:

LOSS
→ SOURCE

ATTACHMENT
→ SINK

Chapter 20 already gave strong independent reason to expect:

LOSS
↓
empty site
↓
new attachment opportunity
↓
rapid reoccupation

So V2 separated the pooled event field into transition channels.

Sources:

attachment
loss

Targets:

attachment
loss
reoccupation
first occupation

It also removed the positive centroid entirely.

All statistics remained signed.

Negative lags were added so that forward temporal structure could be separated from static geometric selection.

The experiment again failed to support the story we expected.

And this time the failure was more informative.

The Zero-Distance Trap

The first interpretation of V2 suggested:

loss
→ more future attachment

attachment
→ more future attachment

But the source and control states at distance zero were definitionally different.

At d = 0:

attachment source
= occupied

attachment control
= empty

and:

loss source
= empty

loss control
= occupied

Those differences automatically alter what events are possible at the same site later.

They are not neighbourhood dynamics.

Once d = 0 was separated from the analysis, the neighbourhood result became much cleaner.

At distances one and two:

ATTACHMENT
→ MORE nearby attachment

LOSS
→ LESS nearby attachment

The effect decayed rapidly with distance.

By roughly distance three, it was close to the scale of the remaining background structure.

The simple source/sink hypothesis was wrong.

But the reversal immediately suggested a simpler explanation.

The Rule Already Contains Local Gain

The Digital Crystal attachment rule rewards occupied neighbours.

Schematically:

attachment score
=
base tendency
+
neighbor gain
+
signal effect
+
anisotropy
-
crowding

The neighbour term is positive.

So if an empty frontier site x becomes occupied:

x attaches
↓
nearby empty cells gain an occupied neighbour
↓
their attachment probability can increase

Likewise, if an occupied cell disappears:

x is lost
↓
nearby candidates lose an occupied neighbour
↓
their attachment probability can decrease

That explanation was almost embarrassingly simple.

Which made it exactly the kind of explanation we needed to test.

The observational V2 result could still be confounded.

The sites that actually attached were not random frontier sites.

They may already have occupied a favourable local trajectory.

So V3 stopped observing naturally selected events.

It intervened.

V3 — Force One Attachment

The new experiment began from the same checkpoint and the same eligible frontier cell.

Then the future split:

SAME CHECKPOINT
SAME CELL x
SAME ENVIRONMENT
SAME KEYED RANDOMNESS

        ┌──────────────┐
        │              │
     FORCE          PREVENT
   x attaches       x does not
        │              │
        └──────┬───────┘
               ↓
        compare futures

The intervention site itself was excluded from every causal-gain measurement.

That mattered.

We were no longer asking whether two different states at x remained different.

We were asking what happened around x because the state had been changed.

    flowchart TD
    A[Same checkpoint<br/>same cell x<br/>same environment<br/>same keyed randomness] --> B[Force x attaches]
    A --> C[Prevent x from attaching]
    B --> D[Future A]
    C --> E[Future B]
    D --> F[Measure gain around x]
    E --> F
  

Before observing the realized first future update, we also calculated what the frozen rule mechanically predicted.

Call that:

g_mech_1

Then we measured the realized one-step neighbouring gain:

g1

And finally the cumulative finite-horizon construction gain:

G_H

over ten updates.

That separation was deliberate.

A causal effect should not automatically be called emergence.

If the local rule already predicts it, then the rule deserves the credit.

The First-Step Effect Was Almost Entirely Mechanical

The V3 run found approximately:

mechanically expected one-step gain
g_mech_1 ≈ 0.105

realized one-step gain
g1 ≈ 0.115

The discrepancy interval included zero and remained inside the frozen accounting tolerance.

So the immediate causal effect was real.

But it was not mysterious.

Forcing one frontier attachment caused additional next-update neighbouring construction, and the magnitude was consistent with the frozen local attachment mechanics at the tested precision.

That was one of the cleanest results in the project.

For once we were not comparing a measured quantity against zero.

We were comparing it against what the mechanism already gave us for free.

And the mechanism survived.

Then the Effect Kept Going

The ten-update cumulative gain was larger:

G_10 ≈ 0.58

Only a small fraction of that appeared in the first update.

Later updates contributed additional construction differences.

At first this looked like a cascade.

Perhaps one attachment caused other attachments, which caused still more attachments.

The temptation was to compare the gain against the classical branching reference:

1 additional event
per initiating event

The ten-update interval sat below one.

But that interpretation contained a hidden assumption.

The forced cell remained in the lattice.

It was not merely an event that happened and finished.

It was a persistent state change.

And the lag profile had not obviously converged by ten updates.

So V3 had mixed two possibilities:

TRANSIENT CASCADE

and

CONTINUING CONSEQUENCE
OF A PERSISTENT OCCUPANCY DIFFERENCE

Those are not the same phenomenon.

We needed one final decomposition.

V4 — Persistent State Versus Transient Cascade

V4 repeated the force/prevent experiment on a fresh seed.

Ninety-six independent groups were used.

All 96 passed the coverage gates.

There were 384 interventions across four predeclared frontier-probability strata.

Capacity remained far from the hard lattice boundary.

The intervention timing was also corrected.

Force and prevent now occurred inside the canonical growth update.

Then the ordinary loss step was applied to every branch.

The forced cell therefore faced the same background loss rule as an ordinary newly attached cell.

Each probe created three futures:

PREVENT
x does not attach

PERSISTENT
x is forced to attach
and then remains under normal dynamics

TRANSIENT
x is forced to attach
it gets one full causal update
then x is removed

The transient arm was the key.

It allowed the original attachment to influence its neighbours once.

Then the direct continuing support from x disappeared.

If the downstream process could sustain itself, the transient branch should continue to separate from prevent.

If not, its gain should die away.

The observation window was extended to thirty updates.

    flowchart TD
    A[Same checkpoint, same cell x] --> PREVENT[Prevent: x does not attach]
    A --> PERSISTENT[Persistent: x attaches,<br/>then remains under normal dynamics]
    A --> TRANSIENT[Transient: x attaches,<br/>gets one causal update,<br/>then x is removed]
    PREVENT --> B[Compare cumulative gains]
    PERSISTENT --> B
    TRANSIENT --> B
  

Immediate Gain Replicated

The mechanical one-step expectation in V4 was:

g_mech_1
= 0.0883
95% CI
[0.0676, 0.1095]

The realized neighbouring gain was:

g1
= 0.1016
95% CI
[0.0677, 0.1380]

and the discrepancy was:

g1 - g_mech_1
= 0.0132
95% CI
[-0.0160, 0.0411]

The frozen accounting verdict remained:

CONSISTENT WITH MECHANICS

So the direct local causal effect had now replicated on a fresh experiment.

That part of the story was no longer tentative.

The Transient Cascade Was Small

After thirty updates, the transient arm had accumulated:

G_transient(30)
= 0.198
95% CI
[-0.026, 0.440]

More important than the cumulative number was the late-time behaviour.

Across updates 21 through 30:

transient late gain
= -0.0081 per update

95% CI
[-0.0201, 0.0039]

That passed the frozen practical-convergence criterion.

The transient trajectory did not continue accumulating positive construction at a meaningful late rate.

Once the initiating occupancy was removed, the downstream effect became small and exhausted itself.

That is the cleanest answer Chapter 23 obtained.

The Digital Crystal supports local causal amplification, but under this protocol the transient cascade is small and converges.

The thirty-update transient gain was also entirely below the descriptive reference value of one.

The experiment records that result as:

TRANSIENT_GAIN_BELOW_ONE_REFERENCE

but deliberately does not call it a branching ratio or proof of subcriticality.

That distinction stays.

Persistent State Was Different

The persistent arm told a very different story.

At thirty updates:

G_persistent(30)
= 1.164

95% CI
[0.786, 1.542]

while:

G_transient(30)
= 0.198

The cumulative difference between the two arms was:

G_persistent(30)
-
G_transient(30)

= 0.966

95% CI
[0.612, 1.333]

The continuing state difference therefore contributed much more to the thirty-step causal consequence than the free-running transient cascade.

That gives us a distinction worth keeping:

PERSISTENT STATE CHANGE
≠
TRANSIENT CAUSAL CASCADE

    flowchart LR
    subgraph Transient
    A[Forced attachment] --> B[One causal update]
    B --> C[x removed]
    C --> D[Small downstream cascade<br/>G ≈ 0.198]
    end
    subgraph Persistent
    E[Forced attachment] --> F[Remains under dynamics]
    F --> G[Persistent state difference<br/>G ≈ 1.164]
    end
    D -.-> H[Most causal consequence comes<br/>from persistent state, not cascade]
    G -.-> H
  

But V4 also killed another attractive interpretation.

There Was No Permanent Positive Growth Offset

One concern after V3 was that the persistent arm might simply grow a little faster forever.

If that were true:

gain per update
≈ constant positive c

and cumulative gain would rise roughly linearly with the chosen horizon.

At thirty updates, that was not what we found.

The frozen late-window persistent gain was:

0.0057 per update

95% CI
[-0.0159, 0.0281]

far below the predeclared positive-offset threshold and with an interval spanning zero.

The cumulative persistent trajectory instead rose early and then flattened into noisy variation:

H=1      0.156
H=5      0.539
H=10     0.839
H=17     1.008
H=22     1.190
H=30     1.164

So the persistent effect was neither:

a freely propagating cascade

nor:

a permanent constant-rate growth advantage

It was a finite consequence of keeping one state difference around long enough for it to reshape later opportunities.

The frozen verdict was therefore:

TRANSIENT CASCADE
CONVERGES

PERSISTENT POSITIVE LATE OFFSET
NOT ESTABLISHED

Opportunity Is Not One Thing

V3 had contained one more striking clue.

The causal effect varied dramatically depending on which frontier cell was forced.

V4 made those strata explicit.

The lowest-probability probe typically sat in sparse geometry.

Its mean baseline attachment probability was about:

0.372

and it had almost exactly one occupied neighbour.

Forcing it promoted approximately:

2.23

new ring-one cells into the frontier.

The total frontier changed by approximately:

+1.23

sites.

At the highest-probability probe, the geometry was almost the opposite.

Baseline attachment probability was about:

0.798

with roughly:

4.07 occupied neighbours

Forcing the site promoted almost no new frontier:

0.031

cells on average.

And total frontier changed by approximately:

-0.969

sites.

The same action therefore meant different things in different local geometry.

Sparse and Dense Interfaces

At a sparse interface:

one frontier site attaches
↓
previously unsupported empty neighbours
gain their first occupied neighbour
↓
new frontier is created

At a dense interface:

one frontier site attaches
↓
most nearby cells are already occupied
or already eligible
↓
the attached site itself disappears
from frontier
↓
little new frontier is created

    flowchart TD
    subgraph Sparse
    A1[One frontier site attaches] --> A2[Unsupported empty neighbours gain support]
    A2 --> A3[New frontier created<br/>+1.23 total frontier]
    end
    subgraph Dense
    B1[One frontier site attaches] --> B2[Most nearby cells already occupied]
    B2 --> B3[Site disappears from frontier<br/>-0.969 total frontier]
    end
  

The paired difference was enormous.

Low-probability probes created about:

2.20

more frontier sites than high-probability probes.

The 95% interval was approximately:

[1.99, 2.41]

with:

p = 0.000125

The same difference appeared in newly promoted neighbouring frontier sites.

This is now established.

Sparse and dense frontier locations respond very differently to the same one-cell attachment intervention.

But one stronger idea did not survive.

Geometry Did Not Yet Predict Long-Run Gain Reliably

The low-probability transient probe had a larger point estimate:

G_transient(30)
≈ 0.677

while the high-probability probe was:

≈ 0.073

That looks dramatic.

But the predeclared paired low-versus-high comparison was:

difference
= 0.604

95% CI
[-0.031, 1.271]

p
= 0.0777

The persistent comparison was even less convincing:

difference
= 0.156

p
= 0.810

So the chapter must stop here.

We have established:

SPARSE GEOMETRY
≠
DENSE GEOMETRY

We have not established:

SPARSE GEOMETRY
→
RELIABLY HIGHER LONG-RUN CAUSAL GAIN

The point estimates make that relationship interesting.

They do not make it true.

Local Gain Is Not Global Gain

The finite evaluation budget also leaves a weak system-wide footprint.

Persistent local gain at thirty updates was:

1.164

but global gain excluding the intervention site was:

1.036

For the transient arm:

local
0.198

global
0.044

The implied far-field differences were negative:

persistent
-0.128

transient
-0.154

but their intervals still overlapped zero.

The candidate-selection sets nevertheless remained extremely similar between branches, typically overlapping by more than 99 percent.

So finite computation does introduce some non-local redistribution, but the intervention does not wholesale rewrite the global schedule.

Most of the measurable effect remains local.

What Survived the Hypothesis?

Chapter 23 began with a hypothesis about motion.

That hypothesis failed.

What survived was more useful.

Predeclared claim

Local process activity propagates through space-time as a reproducible distance-lag structure beyond ordinary geometry and developmental progression.

Status: FAILED

V1 did not establish a travelling process ridge.

The estimator itself was shown capable of turning a strong local signed feature plus far-field noise into apparent displacement.

Stronger interpretation

The active interface behaves as a simple source/sink field in which loss creates interface and attachment consumes it.

Status: FAILED

After zero-distance eligibility effects were separated, the neighbourhood signs were opposite to that interpretation.

Surviving causal phenomenon

Forcing one eligible Digital Crystal frontier attachment causes additional neighbouring construction during the next update.

Status: SUPPORTED

The direct result replicated across V3 and V4.

Mechanistic accounting

The immediate neighbouring causal gain is quantitatively consistent, at the tested precision, with the attachment probability changes mechanically implied by the frozen local rule.

Status: SUPPORTED

V4 measured:

g_mech_1
≈ 0.088

realized g1
≈ 0.102

discrepancy
≈ 0.013
CI includes zero

Transient causal cascade

After the initiating occupancy is removed following one causal update, the remaining construction cascade is small and practically converges over the frozen thirty-update observation window.

Status: SUPPORTED

The late transient rate was indistinguishable from zero under the frozen criterion.

Persistent-state consequence

Keeping the intervention-induced occupancy difference produces a substantially larger accumulated construction consequence than removing it after one causal update.

Status: SUPPORTED

The thirty-update persistent-minus-transient difference was:

0.966
95% CI
[0.612, 1.333]

Persistent positive late-rate offset

The persistent branch retains a scientifically meaningful positive construction-rate offset at late times.

Status: FAILED

The frozen late-window mean was too small and its interval included zero.

Frontier-geometry heterogeneity

Sparse and dense intervention sites differ strongly in how much frontier geometry one attachment creates or consumes.

Status: SUPPORTED

The paired low-versus-high frontier difference was about 2.20 sites with a narrow interval and p = 0.000125.

Geometry determines long-run gain

Sparse frontier geometry produces reliably greater thirty-update causal gain than dense frontier geometry.

Status: FAILED / NOT ESTABLISHED

The transient point estimate was larger in sparse regions, but the frozen paired comparison did not clear the significance gate.

The Causal Opportunity Principle

The Digital Crystal now suggests a refinement of the Interface Principle.

The active interface is not simply a set of empty cells adjacent to occupied cells.

Its local causal structure depends on what an event does to future opportunity.

One attachment can:

remove one current frontier opportunity

while simultaneously:

increase support for nearby candidates

or:

create entirely new frontier sites

depending on local geometry.

That means we need to distinguish:

OPPORTUNITY COUNT

from

OPPORTUNITY STRENGTH

from

REALIZED CONSTRUCTION

from

DOWNSTREAM CAUSAL GAIN

Those quantities are not interchangeable.

A site can reduce the number of frontier locations while increasing expected realized construction.

A sparse site can manufacture frontier.

A dense site can consume frontier.

And a locally strong immediate causal effect can still fail to sustain a free-running cascade.

    flowchart TD
    A[Local attachment] --> B{Local geometry?}
    B -- Sparse --> C[Creates frontier<br/>increases opportunity count]
    B -- Dense --> D[Consumes frontier<br/>reduces opportunity count]
    C --> E[Immediate causal gain may be high]
    D --> F[Immediate causal gain may be low]
    E -.-> G[Long-run gain not reliably predictable<br/>from geometry alone]
    F -.-> G
  

This is a much richer object than the static boundary we went looking for in Chapter 22.

The Crystal Does Not Carry the Event Forward

The most important negative result may be the simplest.

The event itself does not seem to travel.

After the initiating state change is removed, the causal consequence fades.

There is local influence.

There is downstream consequence.

There is finite amplification.

But we did not find a self-sustaining propagating process.

So the progression is now:

LOCAL EVENT
↓
LOCAL RULE CHANGES FUTURE PROBABILITIES
↓
SHORT-RANGE CAUSAL GAIN
↓
SOME DOWNSTREAM CONSEQUENCE
↓
ATTENUATION

not:

LOCAL EVENT
↓
TRAVELLING PROCESS
↓
SELF-SUSTAINING CAUSAL FIELD

That distinction matters enormously for the larger project.

The Digital Crystal can remember consequences in its material.

It can create and destroy construction opportunity.

One local change can alter later local construction.

But causal influence does not automatically organize itself into a persistent process.

Something else would be required.

Maybe the Important State Is Opportunity

The crystal began this book as occupied cells.

Then morphology became insufficient.

Checkpoint state mattered.

History became lossy.

Frontier accessibility mattered.

Loss became construction opportunity.

Finite computation rationed opportunity.

Now intervention tells us that the same material change has different immediate geometric consequences depending on where it occurs.

That suggests a new candidate experimental object:

LOCAL CAUSAL OPPORTUNITY FIELD

Not a biological field.

Not an energy field.

Not a hidden substance.

Just an observer-defined description of where a state change would have high or low causal leverage on future construction.

For a location x at time t, we might eventually want something like:

C(x,t)
=
{
    occupied-neighbour count,
    frontier creation potential,
    attachment probability,
    crowding,
    evaluation probability,
    transient causal gain
}

The question would no longer be:

Where is the crystal?

Or:

Where is the wave?

It would be:

Where can one local change meaningfully alter what the process does next?

That is a different kind of map.

And unlike the boundaries we drew first, this one can be derived from intervention.

What This Does Not Establish

Chapter 23 does not establish:

a wave
an excitable medium
a Hawkes process
a branching process
a critical point
a phase transition
directed percolation
a self
an individual
autonomy
homeostasis
an organism
life

The experiment explicitly records those as forbidden overclaims.

It also does not establish that sparse geometry causes larger long-run gain.

That remains a candidate relationship.

And it does not establish that a thirty-update transient gain below one is a universal property of the substrate.

The result is conditional on this model, these parameters, this intervention, this horizon and this definition of causal gain.

What We Have Earned

We have earned something narrower and more useful.

A local Digital Crystal attachment has measurable causal leverage on subsequent construction. Its immediate effect is consistent with the frozen local attachment mechanics. When the initiating occupancy is removed after one causal update, the remaining cascade is small and converges over the tested horizon. Keeping the occupancy difference produces a substantially larger accumulated consequence, but no persistent positive late-rate offset was established. Sparse and dense frontier locations differ strongly in how an attachment changes future construction opportunity, although a corresponding difference in long-run causal gain remains unconfirmed.

That is not life.

It is not individuality.

It is not even a coherent process yet.

But it is something we did not have before Chapter 23.

A measured causal leverage of local state.

Next: Where Is Causal Gain Created?

The obvious next experiment is no longer a search for a boundary.

Nor should we immediately sweep a global parameter and announce a critical point.

The substrate has already shown us that local geometry matters.

So Chapter 24 should begin there.

Map the conditions under which an intervention has high or low transient causal gain.

Then ask whether high-gain regions are:

rare
common

isolated
clustered

momentary
persistent

disconnected
connected through space-time

Only after that would it make sense to ask whether a change in global neighbour coupling or computational budget moves the system into a new collective regime.

The research path has changed again:

    flowchart TD
    A[MATERIAL] --> B[INTERFACE]
    B --> C[OPPORTUNITY]
    C --> D[LOCAL CAUSAL GAIN]
    D --> E[HIGH-GAIN REGIONS?]
    E --> F[PERSISTENT HIGH-GAIN REGIONS?]
    F --> G[CONNECTED CAUSAL ORGANIZATION?]
    G --> H[ONLY THEN:<br/>A NATURAL INDIVIDUAL?]
  

Chapter 23 began by asking whether the process moved.

It did not find a travelling process.

It found something more basic.

A local event can change the future.

But whether that influence becomes organization depends on where the event happens, how long its state persists, and whether the consequences can survive after the original cause is gone.

That is the next frontier.