04: The Digital Crystal

Outlier was the wrong laboratory for the question we now wanted to ask.

Not a wrong system. The last two chapters were worth every page, and nothing in them is retracted. Outlier showed us what computation can support.

But by the end of the flocking investigation, another problem had become unmistakable: it was much better at producing phenomena than at isolating them. Geometry, ancestry, distance, expansion, local environment and interaction all arrive together in that world, arising from the same 512 bits, moving together, entangled beyond separation. Every control we built was an attempt to statistically undo an entanglement we could not experimentally prevent, and one entire regime stayed unresolved because the comparison we needed simply did not exist in the data.

Matching is one way to recover a comparison when the experiment did not create one for you.

Now we want the opposite situation.

We want to construct the comparison before the world runs: hold everything we can fixed, change one mechanism deliberately, and measure what changes with it.

There is one idea worth carrying across from Outlier, and it is much smaller than an organism:

local computation ↓ repeated interaction ↓ larger-scale organization

That is all we need to import.

Not Outlier’s reproduction. Not its causal families. Not its geometry.

Only the demonstrated possibility that local rules can generate organization we did not explicitly represent.

What we want from the new system is a short list:

every rule is known
every state can be inspected
one mechanism can be changed at a time
counterfactual worlds can be rerun
the full history can be preserved

And an equally important list of what we refuse to build:

organism memory repair reproduction metabolism individual

Not because those questions are uninteresting.

Because if those concepts appear explicitly in the machinery, we lose the ability to discover whether anything resembling them arises from something simpler.

The laboratory must not contain the answer.

The result is going to be almost embarrassingly small.

One Seed

A hexagonal lattice. Every location holds 0 or 1. Every location has six immediate neighbours.

At the centre, one occupied location:

●

Everything else is empty.

The rule:

An empty location becomes occupied if at least one neighbouring location is occupied. Once occupied, it stays occupied.

That is the entire system.

One seed. One local attachment condition. Irreversible occupancy. Time.

There is no target morphology and no higher-level object directing the growth.

Hexagonal geometry is a convenience rather than a claim: six equidistant neighbours make local reasoning cleaner than a square grid’s awkward mix of edge and corner adjacency. Represent locations as axial coordinates (q, r) and the six directions are just six offsets. The world is a set of occupied coordinates, and one update collects every empty location adjacent to something occupied and fills it.

Run it and the structure grows in expanding hexagonal shells.

Nothing anywhere says make a hexagon. After t updates, every location within hexagonal graph distance t of the seed is occupied, and the global shape follows from the neighbourhood topology plus uniform local propagation. No cell holds a blueprint. No controller measures the radius. Nobody draws the six sides.

Nothing surprising has happened yet, and that is useful.

The bounded result is simply:

Under this rule, one seed produces ordered expanding geometry through repeated local attachment alone.

The hexagon is not evidence of sophistication.

It is the baseline against which later deviations will become measurable.

Growth Is Cheap

Now measure it rather than admiring it, because the population has a closed form. For a perfect hexagonal ball of radius r:

So the structure gets steadily larger and not one bit more interesting. A million-cell structure from this rule is no more conceptually complicated than a seven-cell one; the generative description is identical, and only the number changes.

larger
≠
more complex

Which is useful to establish early, because size is the cheapest possible impressive-looking result.

A second distinction appears almost for free:

continued construction ≠ reproduction

The structure can keep extending from one seed without producing a second independent copy.

That does not tell us whether reproduction matters to digital life.

It tells us only that growth and reproduction are separate computational possibilities, and therefore deserve separate experiments.

That is a substrate-first move, and it is the reason this laboratory is built around growth rather than around anything more ambitious. We get to study continued process before reproduction instead of assuming the biological ordering.

The First Temptation

Before adding anything, it is worth noting how quickly even this system tempts biological language — because it happens within one experiment.

Let the structure grow for twenty generations, then erase a region from its interior. Resume the same rule. The empty cells inside the hole touch occupied cells, so they become occupied; then their neighbours do; and soon the hole is gone.

It looks like healing. It is not.

The rule says any empty location adjacent to occupied structure becomes occupied, and that rule operates identically outside the structure and inside a hole. The system does not distinguish damage from ordinary frontier — there is no target morphology anywhere, nothing that could represent what the structure is supposed to look like. The hole closes because continued growth plus available empty space closes holes.

But nothing in the rule distinguishes damage from ordinary empty space.

The same attachment rule that advances the exterior frontier also fills an interior hole.

So the stronger interpretation disappears:

The structure refills missing space through ordinary growth.

That is not yet repair.

We will return to material loss later with a system and controls designed specifically for that question.

Less exciting, more informative, and a good reminder that a laboratory built specifically to avoid biological assumptions will still generate biological-sounding descriptions within about five minutes of being switched on.

The prototype generated other questions too: obstacles could leave traces, multiple seeds could merge, and finite worlds eventually imposed limits.

We are going to resist following those branches here.

Each becomes a much better experiment later.

For now, the prototype has done its job: it gives us a transparent process that grows, can be perturbed, and contains almost nothing we did not deliberately put there.

Can the Environment Leave a Mark?

The prototype is almost too predictable.

Good.

A laboratory should begin with a baseline we understand. What we want next is genuinely different:

Can changing external conditions influence growth strongly enough to leave a persistent, measurable signature in the finished structure?

That question is not askable of the prototype. Its rule is deterministic and saturating — every available location fills, every time. There is no room for an external signal to change which locally available events actually happen, because all of them happen.

So the model has to change, and this needs stating plainly rather than slipped in:

PROTOTYPE

binary occupancy + hexagonal neighbourhood
+ irreversible growth + deterministic local attachment

DIGITAL CRYSTAL v1

binary occupancy + hexagonal neighbourhood
+ irreversible growth + stochastic attachment
+ environmental forcing + fixed lattice anisotropy
+ crowding penalty

Each addition earns its place through the experiment we want to run. Stochastic attachment gives the forcing somewhere to act — if attachment is probabilistic, the environment can shift the odds. Anisotropy gives the lattice persistent directional structure. The crowding penalty stops attachment probability rising without bound simply because a location has many occupied neighbours.

These mechanisms are introduced for this experiment. They are not discoveries carried over from the prototype, and that has a consequence worth being strict about:

Results from the prototype do not automatically transfer to Digital Crystal v1.

The hole-filling result above tells us nothing about how this stochastic model responds to damage. Different model, different claims, and everything from here has to be earned again.

This is where the name Digital Crystal becomes useful.

Not because the structure resembles quartz, and not because it sits on a hexagonal lattice.

The analogy is mechanistic: local interactions during formation accumulate into persistent larger-scale structure.

So we can state a hypothesis rather than a definition:

Can a local computational growth process turn characteristics of an external input into persistent, measurable morphology?

If the answer is no, the name has earned nothing.

If the answer is yes, we can decide what the name is worth afterwards. morphology.**

Deliberately narrow. It says nothing about life, memory, learning, adaptation, reproduction, intelligence or agency. Its only virtue is that it can be tested, and might be false.

Influence, Not Instruction

The prototype was:

if signal == "sine":
    draw_sine_shape()

That would be a strange plotting library with extra steps. The same growth mechanism operates under every source; the signal only changes the conditions under which individual local attachment events occur.

For a candidate frontier cell, attachment probability takes a form like:

The experimental constraint that makes the whole thing work:

The growth mechanism remains fixed while the source changes.

Otherwise recovering the source would tell us nothing except that we had changed the machine.

The contrast with a conventional visualization is the cleanest way to see what is going on. A graph does:

value → coordinate → pixel

The Digital Crystal does:

value → local attachment conditions → many stochastic interactions → persistent morphology

The picture is not drawn from the signal.

It is grown under its influence.

No local attachment event receives the source label, the future sequence or the desired final form. It receives only the current local state and the environmental value available at that step.

Whatever source information survives has to survive through the growth process.

First, the least interesting possible check: does the generalized model still grow? Give it a constant signal, E(t) = 0, and run it. The baseline reached approximately 5,924 occupied cells at a maximum hex radius of 44, with 552 boundary edges.

That proves nothing except that the machine runs. Which is all it needs to prove.

Six Environments

Now hand exactly the same growth rule six different kinds of forcing:

constant
sine
square
sawtooth
white noise
random walk

Within each family the individual instances vary — periods, phases, noise realizations, random-walk trajectories. At each growth step the structure receives only the scalar value currently presented to it. Not the family name, not the history, not the shape of the signal. One number.

Then grow them.

They look different.

So what?

This is the same trap that has followed us since Chapter 1, and it is no less inviting for being familiar. Different signals generate different crystals would make an attractive demonstration and establish nothing. Perhaps one random seed happened to produce a larger structure. Perhaps square waves simply produce a higher mean attachment probability, and we are looking at area. Perhaps our eyes are busily categorizing noise, which they are extremely good at.

We need populations, not specimens.

Six Hundred Crystals

One hundred each of constant, sine, square, sawtooth, white noise and random walk. Six hundred structures, all grown with the same rule, the same experimental horizon and the same measurement system, differing only in source instance and stochastic realization.

For each finished structure, 42 morphological measurements: area, perimeter, maximum radius, compactness, boundary roughness, bounding-box aspect, centroid displacement, radial and angular distributions, six-fold angular structure, boundary-radius variation, and so on.

The source signal is not among them. The measurement describes the finished object and nothing else.

Then hide the labels.

The question becomes concrete and slightly unnerving: hand over one finished crystal, say nothing about what grew it, and ask whether the forcing family can be recovered from morphology alone. Six classes, so random guessing succeeds 16.7% of the time. Train on part of the population, test on structures never seen before.

    flowchart TD
    S1["Constant ×100"] --> G["Same fixed growth rule"]
    S2["Sine ×100"] --> G
    S3["Square ×100"] --> G
    S4["Sawtooth ×100"] --> G
    S5["White noise ×100"] --> G
    S6["Random walk ×100"] --> G
    G --> M["Morphology: 42 measured features"]
    M --> H["Hide source labels"]
    H --> C["Train on one subset"]
    C --> T["Test on unseen crystals"]
  

The held-out result:

chance                 16.7%

random forest           52.2%
logistic regression     53.9%

Substantially above chance on held-out structures.

The environment has left a readable morphological signature.

Two different classifier families produce similar held-out accuracy, which makes the result less dependent on one particular decision boundary.

That is reassuring, but not magical. The important result is simpler: unseen crystals contain enough morphological information for source family to be recovered substantially above the six-way chance baseline.

The confusion matrix also shows that the information is uneven. Some forcing families leave much more distinctive signatures than others.

The claim to keep is smaller than the excitement it produces:

The final morphology contains information that makes forcing-process family recoverable substantially above chance.

Not the crystal remembers its history. Not the crystal understands the environment. Something happened during formation, and the finished structure still carries enough of it to be read.

Attack the Boring Explanation

The obvious deflation is that the classifier is not detecting anything about the process at all — merely some trivial aggregate. Square waves might spend more time at high values, raise mean attachment probability, and produce larger structures. Then all we would have discovered is that bigger signals make bigger crystals.

So remove the simplest aggregate explanation among the varying signals: normalize them to approximately the same mean and ask whether their resulting morphology still differs.

The constant-zero condition remains a separate baseline.

Standardized distances from that baseline were:

sine          4.45
square       13.82
saw           2.44
white noise   2.40
random walk   1.40

The variable-source populations remain morphologically distinct after their means are approximately aligned.

So differences in mean forcing are not sufficient to explain the separation among those source families.

Something about the distribution or temporal structure of the forcing is also contributing.

Which one?

That is now the interesting question.

One more obvious worry: perhaps recovery works only at one carefully chosen forcing strength, and we happened to find it. Varying the forcing strength while leaving the rest of the growth mechanism untouched, held-out random-forest accuracy was:

forcing         accuracy

0.75             34.1%
0.85             43.2%
0.95             50.0%
1.00             52.3%
1.05             52.3%
1.15             63.6%
1.25             43.2%

against a 16.7% chance baseline throughout.

Discipline here, because this sweep is noisy and it would be easy to over-read. There is no clean monotonic trend; 1.15 outperforms 1.25 substantially, and a seven-point sweep cannot tell us the shape of a response curve. We are not entitled to say that increasing forcing predictably increases recoverability, and we do not know the optimum.

What survives is narrow and sufficient:

Source-family information remains recoverable above chance at every tested forcing strength in this sweep.

The phenomenon is not an artefact of one lucky parameter.

Maybe It Recorded What Happened

Here is where it becomes tempting.

We have a fixed local process, an external environment, and a finished structure from which the character of that environment can be recovered well above chance. The mean does not explain it. The result holds across forcing strengths. It survives on structures the classifier has never seen.

The tempting sentence is:

the crystal has recorded its environmental history.

And this time the temptation is not merely visual.

Something about the conditions during formation really is recoverable from the final structure.

But information about past conditions and a recoverable history are not the same claim.

Notice how little distance there is between the result we have and the claim we want. The result says information about the kind of environment survives. The claim says the history survives. Those feel like the same statement, and the whole book so far suggests they are not.

There is also a specific reason for suspicion. Consider sine and square forcing. They differ in temporal ordering — but they also differ in value distribution, time spent near extrema, autocorrelation and transition structure. If the classifier is picking up broad statistical properties of the values experienced during growth, that is still a real result. It is a different result.

So we test the promotion rather than accepting it.

Destroy Time

Take a source sequence and shuffle it.

0.7   0.2  -0.3   0.9  -0.8  ...

becomes

-0.3   0.9   0.7  -0.8   0.2  ...

The shuffled version preserves the same values, mean, variance, minimum, maximum and histogram. Only the temporal ordering is destroyed.

If the finished morphology retains recoverable information about chronology, crystals grown from ordered and shuffled sequences should be distinguishable. Binary chance is 50%.

chance                 50.0%

random forest           51.3%
logistic regression     51.7%

The result sits essentially at chance under both tested classifiers.

Whatever allowed us to recover source family did not provide a usable ordered-versus-shuffled signal under this morphology representation and protocol.

Be precise about what this does and does not establish. It does not prove that no conceivable measurement could ever recover temporal information from this model. It establishes:

Our morphology representation and classifiers do not recover ordered-versus-shuffled history above chance under this protocol.

Which is already enough to kill the claim we were drifting toward.

But the control can be stronger, and should be. A sine wave and a square wave do not merely differ in ordering — their distributions differ too, so the successful source-family classifier may be leaning primarily on distributional structure. We can remove even that.

Construct one fixed multiset of 72 values. Every condition receives exactly the same values — not approximately the same distribution, not merely matched mean and variance, the identical multiset. Only the temporal arrangement changes:

RANDOM        random permutation
BLOCK         low values grouped, then high values grouped
ALTERNATING   low, high, low, high...
SMOOTH        neighbouring values change gradually
BURST         quiet periods interrupted by concentrated excursions
PERIODIC      values arranged into a repeating temporal motif

This is the comparison Outlier could never have given us. Same values, same growth rule, same everything — one variable changed deliberately, and the counterfactual world constructed rather than found.

Then grow them, with the model frozen. No parameter tuning, no attempts to rescue the hypothesis.

One safeguard matters enough to state. All temporal arrangements built from a single value set stay together during the train/test split, so the classifier cannot train on one ordering of a value set and then be tested on a different ordering of that same set. Held-out value sets are genuinely unseen. Without that, a subtle leakage path would let the model recognize the values rather than the ordering.

With the value multiset held exactly constant, the tested morphology representation and classifiers still do not recover temporal organization above chance.

The stronger interpretation fails again.

The values mattered.

Their exact ordering, under this test, did not remain readable.

A Crystal Is Not a Tape Recorder

This is the point where the result becomes better than the idea we started with.

We had imagined:

environmental history
↓
crystal
↓
history written into morphology

What the experiments support is narrower:

environmental statistics
↓
local dynamics
↓
persistent morphology

Under the measurements and classifiers we tested, exact temporal organization is not recoverable.

Broader characteristics of the forcing are.

That is a more precise result than saying the sequence was simply “lost,” because unmeasured information may still exist in the state.

Which is, when you look at it, strangely appropriate. Inspect a physical crystal and its structure may reveal a great deal about the conditions under which it formed — temperature regime, rate, impurities, pressure. It does not contain a frame-by-frame movie of formation. Nobody expects to read Tuesday off a quartz sample.

Our Digital Crystal turns out to be closer to that than to a recording device. The morphology behaves less like a tape recording and more like a compressed consequence of formation.

Which means the failure has told us something the success could not: what kind of information this substrate preferentially preserves.

State Is Not History

The crystal has state.

Its present configuration is a consequence of earlier attachment events. A cell added at step 10 can still be present at step 70, and changing earlier events can change the final structure.

But:

the past affecting the present is not the same as the past remaining recoverable from the present.

That is the distinction this experiment has finally forced us to make.

Consider two sequences, A B C D and D B A C. Both influence the process. Both alter the final state. But if the final state contains no recoverable information distinguishing which ordering occurred, then the process has accumulated consequences without retaining chronology.

That is where Digital Crystal v1 stands, and it separates two things our vocabulary bundles together:

past contributed to present
        SUPPORTED

source family recoverable
        SUPPORTED

mean alone explains source effect
        FAILED

exact temporal order recoverable
        NOT SUPPORTED

complete chronology retained
        NOT ESTABLISHED

So:

PAST-DEPENDENT
≠
PAST-READABLE
≠
RECOVERABLE HISTORY

A process can be thoroughly shaped by its past without being a record of it.

A useful working description is lossy integration:

external forcing
↓
many irreversible local attachment events
↓
aggregate structural bias
↓
persistent morphology

The process integrates consequences of earlier forcing into persistent form, while our measurements recover broad source characteristics more readily than temporal arrangement.

That suggests a hypothesis worth carrying forward:

Irreversible growth may preserve coarse information about formation conditions while making temporal order difficult to recover.

Not a law.

A result from this substrate, waiting to be attacked by a better experiment.

One boundary statement, once, and then we trust the reader with it: nothing here establishes memory in the strong sense, learning, adaptation, interpretation of the environment, a stored event log, reproduction, individuality, or life. What it establishes is that a fixed local growth process can turn differences in environment into persistent, measurable, recoverable differences in form.

That is a real result. It does not need help.

The Digital Crystal

The name was a label at the start of the experiment. It has now earned a little weight, and the definition it earned is the modest one:

A Digital Crystal is a local computational growth process in which characteristics of an external input become expressed as persistent, measurable morphology.

    flowchart TD
    A["ENVIRONMENTAL INPUT"] --> B["FIXED LOCAL DYNAMICS"]
    B --> C["DIGITAL CRYSTAL"]
    C --> D["PERSISTENT MORPHOLOGY"]
    D --> E["SOURCE-FAMILY INFORMATION RECOVERABLE"]
    D --> F["TEMPORAL ORDER NOT RECOVERED"]
  

The definition is small because the stronger interpretation failed.

We established that formation conditions can leave recoverable morphological information.

We did not establish that the final morphology preserves a recoverable chronology of those conditions.

Keeping those statements separate is the result.

It is worth noticing what just happened procedurally, because it is the thing Outlier could not offer. We wanted to know whether temporal order mattered. So we built two worlds identical in every respect except temporal order, and looked at the difference. No matching. No searching a complicated world for naturally occurring comparisons. No unresolved regime where the data happened not to contain what we needed.

That is what the laboratory is for, and it worked on its first serious question.

Give It a Past

We asked the crystal what happened to it.

The morphology could answer something like:

I can tell you something about the conditions under which I formed.

Then we asked a harder question:

Can you tell us in what order those conditions occurred?

Under the tests we ran, the morphology could not.

That is not a disappointing end to the Digital Crystal. It identifies the next missing capability with unusual precision. The crystal has a present, and that present carries information about the conditions that produced it.

What we have not yet given it is a recoverable past: enough preserved state that two different histories can remain distinguishable even when their current visible morphology is otherwise comparable.

So the next step is not to make it intelligent. Not to give it goals. Not reproduction. Not learning.

Something much smaller.

Give the process a way to keep what happened.

Not to interpret the past. Not to learn from it. Not to call it memory.

Just preserve enough internal consequence that two different histories remain different later.

Because before the past can change the future, some distinction from that past has to survive.

That is the next experiment.