A Universe in a Strand of Cells

Part 1 · Four letters

The alphabet was the last constant

The story so far, in one breath. Page one: a row of cells, each cell black or white, each new generation decided by three parents above — eight parent-combinations, so a rule is an answer sheet of eight yes-or-no answers, 256 rules in all. Page two changed the world the rules run in; page three changed the rule mid-run. But every plate on every page kept one thing sacred without ever saying so: every cell was one of exactly two things. Ink or paper. On or off.

Biology got there first, by a few billion years, and made a different choice. The strand of DNA in every living cell is a row of letters from a four-letter alphabet — A, T, G, C — and the letters come paired: across the double helix, A always faces T, and G always faces C. The pairs are not equal in strength. An A–T pair is held by two hydrogen bonds; a G–C pair by three. Chemists bake this into their standard shorthand: G and C are the strong letters, A and T the weak ones. And there is a second, older division: A and G are purines (two-ring molecules), C and T pyrimidines (one-ring).

Two independent two-way splits of four letters. Stop and look at that, because it is the door this whole page walks through: a DNA letter is exactly two bits. One bit says purine or not; one bit says strong or not. Answer both questions and you have named the letter. Writing each base as two bits this way is a published convention — Nemzer set it out in 2017 and showed that the mutations we will meet later come out as bit operations — so we are borrowing, not inventing. (There is even a third classic split, amino versus keto — and it is the exclusive-or of the other two. The chemistry hands you a tidy little algebra and refuses to explain why.)

Which means our two-letter universes were never the wrong size. They were half the right size. A four-letter cell is a pair of page-one cells — and everything this series has built suddenly applies twice over.

How this page draws four letters

Two honest flags before we build anything. First: there is a research field called DNA computing that builds actual computing machines out of actual DNA molecules. This page is not that. Here the DNA is the metaphor, not the machine. Second: using the four bases as the four states of a cellular automaton is itself not new — Sirakoulis and colleagues did it in 2003, treating A, C, T and G as digits of a base-4 number. What we could not find published anywhere is what happens next: the two-bit anatomy and the automata, joined.

Part 2 · A base pair of universes

Two rules on two bit-planes

Here is the construction, and it is smaller than it sounds. Take a row of four-letter cells. Since each letter is two bits, the row is secretly two rows of page-one cells riding in the same paper: a purine plane (all the left halves) and a strong plane (all the right halves). Now give each plane its own elementary rule — one of the ordinary 256 for the purine bit, one for the strong bit. No new machinery, no bigger answer sheets. A four-letter universe, built from two 8-entry answer sheets.

Why not just invent rules for four letters directly? Because of the arithmetic. A four-letter rule must answer for every combination of three four-letter parents — a 64-entry answer sheet, and 4⁶⁴ ≈ 3.4 × 10³⁸ possible rules. The 256 you could meet personally; that number you could not meet a meaningful fraction of if you enumerated for the age of the universe. (Wolfram hits the same wall with three colors and tames it by restricting to totalistic rules, which only count their parents' total — that cuts three-color rules from 3²⁷ down to 2,187. As far as we could find, his book never features a four-color version at all.) The two-plane construction is a different kind of taming: it keeps every rule browsable, and it earns the name this page wanted anyway — a base pair of universes.

If the two planes never speak, you get exactly that: two transparent worlds overlaid, each obeying its own law, sharing paper and nothing else. The interesting move is coupling them — let one bit's rule read one of its three parents from the other plane. Then the purine bit feels the strong bit and the letters become a genuine four-letter dance:

And a subtler discovery, found by planting the wrong seed on purpose. Seed the world with a single A — purine bit lit, strong bit dark — and whether the strong plane ever wakes up depends not on the coupling but on the rule sitting there. Couple Rule 90 into a dark plane and nothing happens, forever: 90 answers "white" to both of the only questions the dark plane ever asks it. Couple Rule 30 in its place and the plane ignites on the very first step:

The lesson of the whole section, compressed: on page two we learned a rule's personality is a partnership between the rule and its geometry. Here the partnership acquires a third member. A rule's reach — which letters can even exist — is a fact about the rule table, not about the wiring. The alphabet is not a given. It is an achievement.

Part 3 · The photocopier

One of the 256 is a copying machine

Before this page can talk about mutations, it needs a world where copying is the whole law. It turns out we have had one all along. Ask: is there an elementary rule whose answer, for every one of the eight parent-combinations, is simply "whatever the middle parent was"? Check all 256, and exactly one survives: Rule 204, the photocopier. Every generation is a perfect copy of the one before. On page one it would have been the most boring plate imaginable — vertical stripes, forever. It took three pages of context for 204 to become interesting, because 204 is not a picture. It is a promise: nothing will change. It is also the rule DNA cares about, because a dividing cell is, before it is anything else, a photocopier of its strand.

And real copying makes mistakes. Run the photocopier on a four-letter strand — the same rule on both bit-planes — and let each copy occasionally miswrite a letter. To keep faith with the series' promise that everything here is reproducible, our accidents are seeded: the errors fall where a fixed random sequence says they fall, so every replay of this plate makes exactly the same mistakes. Even our accidents are reproducible.

How bad is real copying? The enzyme that copies your DNA, left alone, writes a wrong letter roughly once every 10,000 to 100,000 letters. That sounds impressive until you multiply: a human genome is about three billion letters — an unassisted copy would carry tens of thousands of new errors, every cell division. The actual measured rate, in the finished product, is about one error per 10⁸ to 10¹⁰ letters — roughly a hundred-thousand-fold better than the enzyme alone. That factor is not chemistry getting lucky. It is machinery: a proofreading step that catches most errors as they are written, and a second inspection that patrols behind the copying fork. We will meet both in Part 5 — but first, we owe the errors themselves a fair portrait. What exactly does one mistake cost?

Part 4 · The cone of consequence

What one changed cell costs

Here is the experiment this whole page has been circling. Run two copies of the same world side by side, in perfect lockstep, from the same seed — then reach into one of them, once, and flip a single cell as it is copied. From that moment on, paint every cell where the two worlds disagree. The result is called a difference pattern, and drawing it is a fine old tradition, not our invention: Wolfram draws exactly this picture in his book to tell his four classes of rules apart, physicists have studied it since the 1980s under the name damage spreading, and the widening wedge of disagreement has a name of its own — the defect cone.

The pair of plates above is the page's central contrast, so let it be said plainly: a mutation has no cost of its own. Its cost is set by the law it lands in. Under the photocopier, an error is a scar — permanent, local, legible. Under Rule 30, the same error is an avalanche. The damage-spreading literature sorts every rule by exactly this behavior: in Wolfram's classes 1 and 2 the cone stays shut, in class 3 it opens at full width, and in class 4 it flickers between the two. And the fine print holds a mercy worth quoting: Wolfram's own notes record that in some rules a flipped cell, landing among the wrong neighbors, changes nothing at all after a step. Not every mutation matters. Some land silently.

Mutations with a chemistry

Now bring the four-letter alphabet back, because biology does not flip abstract bits — it swaps letters, and the swaps have kinds. A transition swaps within a chemical family: A↔G (purine to purine), or C↔T (pyrimidine to pyrimidine). A transversion crosses the family line. Count the possibilities and transversions should dominate — every letter has two transversion partners and only one transition partner. Measure real genomes and the opposite holds: in human de novo mutations, transitions outnumber transversions about two to one. (The best-understood reason is a chemical booby trap: one commonly modified form of the letter C, when it degrades, turns directly into a T — a transition the repair machinery is poorly placed to notice.)

Here is where the two-bit anatomy pays for itself. Look at what the mutation kinds do to the bits: a transition flips only the strong bit. A↔G and C↔T are exactly the swaps that keep the purine half and toggle the strong half. A transversion is any swap that flips the purine bit. The biology and the bit-planes agree perfectly — which means on an uncoupled pair of universes, a transition is invisible to an entire plane:

Part 5 · The other strand knows

Mismatch, and the machinery of repair

Part 3 promised an explanation for that hundred-thousand-fold improvement, and the explanation begins with the double helix's deepest trick. DNA is not one strand — it is two, and each strand determines the other completely: A across from T, G across from C. In our two-bit language, complementation is a one-bit operation, and it is not the bit you might guess. Pairing always sets a purine across from a pyrimidine, so the complement flips the purine bit — but an A–T pair is weak on both sides and a G–C pair strong on both sides, so the strong bit is shared. Strength belongs to the pair, not the letter. That is the whole reason the ladder's rungs hold.

This redundancy sounds wasteful — the second strand adds no information — until the first mistake happens. The classic 1958 experiment of Meselson and Stahl showed that when the helix is copied, it splits, and each old strand templates a fresh new partner. So every copying error is born into a very particular situation: a brand-new letter, sitting across from an old letter it fails to pair with. That failure has a name — a mismatch — and it is physically detectable, a lump in the ladder. An error under pure copying (Part 3) was invisible; an error in a paired strand announces itself. And announced errors can be caught. Proofreading catches most as they are written; a second system, mismatch repair, patrols afterward and — crucially — fixes the mismatch by trusting the old strand and rewriting the new one. The template wins.

Notice what repair cannot do. It cannot restore a letter both strands have lost — once a mutation survives a round of copying, the wrong letter is the template, the pair rhymes again, and the mismatch machinery sees nothing to fix. Repair is not memory. It is a brief window of second chances, held open by redundancy, and it closes at the next division. We looked for prior art here and report the negative honestly: we could find no published cellular automaton that runs two strands with complementary pairing and a repair step between them. If it exists, it is well hidden.

Part 6 · When the law itself mutates

One bit of the answer sheet

Everything so far mutated the material — a letter here, a letter there, the law serenely constant. But this series has already broken that symmetry once: page three swapped whole rule tables mid-run, planned and purposeful, like an amendment. A mutation is the unplanned version. And the smallest possible mutation of a law is precise: an elementary rule is eight yes-or-no answers, so flip one answer. Rule 110 and Rule 111 differ by a single bit. What happens to a world whose law drifts, one seeded random bit at a time, while it runs?

Is a mutating law just a metaphor? Wolfram himself makes the connection in his book, and it is worth quoting, because it is the bridge between this page's two halves:

In the pages that follow, he explores exactly our move — programs subjected to random mutations — but with one structural difference: his picture is a sequence of programs, each one a mutation of the one before, each run from the start. The change happens between runs, not inside one. We could find no place in the book where a rule is changed partway through a single run. He did come back to something adjacent a year later — a 2003 note on "cellular automata with global control," where two rules alternate according to whether the world's total count of black cells is even or odd. That is a law that switches every step by a rule of its own; a world that runs cleanly and is interrupted once, at random, seems to be a picture he never drew. It is the picture this page cares about, because it is the biological one: the genome your cells copy today is not re-run from the start. The mutation lands midstream, and everything downstream inherits it.

Part 7 · The garden of descent

Every bud inherits, imperfectly

Now assemble the finale from parts this series already owns. Page three built the recursive garden: spiral worlds that bud new spiral worlds, every head growing by the golden angle, nothing ever rewritten, children leaning away from parents by pure skipping. It had one aristocratic assumption left: the rules were assigned — by us, from outside. Replace assignment with inheritance. When a head plants a bud, the bud receives its parent's rule — copied with, sometimes, one flipped bit. The mutation of Part 6, made heritable. The copying errors of Part 3, applied to the law itself.

Say what this object is, carefully. Each head's rule came from its parent's rule by copying, with occasional one-bit error. Lineages accumulate their private histories of flips; cousins diverge; the family of laws has a family tree, and the tree is drawn in space, because children grow away from parents by the skipping rule. There is an old and honorable name for descent, with variation, down a branching tree — and one thing must be said as carefully as the rest: nothing here is evolution in the full sense, because nothing is selected. No head competes, none dies for its law, no mutation is rewarded. This garden is what Wolfram's biology chapter argues so much of the living world's raw diversity is: strings of random mutations that happened not to be fatal, accumulating. Variation without a judge. The judge is a different page's problem.

And notice — quietly, because the series has been building this image for three pages — that the garden keeps every promise ever made on these pages at once. Nothing rewritten: the append-only law, from page two. The law can change: from page three. The change is copied imperfectly down generations: this page. A single seed, one strict global order of cell-laying, no erasures, and out of it: a phylogeny. The adding is still enough.

Part 8 · What the strand taught us

The error was never the enemy

Page one: complexity is cheap — eight answers can fill a world. Page two: the rule is half the universe; geometry is the other half. Page three: even the law can change, and changing it buys power. This page's lesson completes the set, and it is the one biology has been shouting all along. A universe that copies perfectly is inert — the photocopier's stripes, the collapsed diagonal, the founder disk of dead 204. Everything on this page that grew, surprised, diverged, or invented did so downstream of an error: a flipped letter's cone, a bud's misinherited law, a relic column surviving its world. The machinery of life is not built to prevent copying errors absolutely — at its measured error rates it could go a million generations without one, and nothing would ever have changed. It is built to make errors rare, detectable, and survivable. Which is a different engineering goal entirely.

Where does this page sit in the literature? We report what we found, and what we did not. The four-letter alphabet as CA states: published, 2003, Sirakoulis and colleagues, who treat the bases as digits of a base-4 number. The two-bit anatomy of the bases, with mutations as bit operations: published, 2017, Nemzer. The difference cone: Wolfram's own book, and a whole damage-spreading literature since the 1980s. Automata that stay reliable under constant noise: a celebrated and genuinely heavyweight corner of mathematics — Gács proved a one-dimensional automaton can remember a bit forever even when every cell errs at every step, a result that took him some two hundred pages; Toom's lopsided majority rule does it in two dimensions with three votes. (Our rare, seeded, one-at-a-time accidents are a much gentler regime than that literature's everywhere-noise — a distinction worth keeping.) What we could not find published anywhere: the two-bit decomposition joined to CA dynamics as coupled bit-planes; a two-strand automaton with pairing and a repair step; laws inherited with mutation by budding growth centers. Perhaps we searched badly. Perhaps this strand, too, is unphotographed country.

One last honest signpost, for readers who know some biology and have been waiting for it: there is another famous table in the living cell — sixty-four three-letter words mapping to twenty amino acids and three stops. That table belongs to translation, the reading of genes into proteins, and this page has deliberately never touched it: everything here is about copying, the other machine. Different machine, same four letters. It would make a fine page of its own.

Part 9 · Your turn

Grow your own strand

All five worlds from this page. Two rules to rewrite at will — by number, or by flipping their answers directly — a coupling to choose, mutation kinds to aim, repair machinery to switch on and off, and a garden that inherits whatever law you give it. The accidents are seeded: keep the seed and every run reproduces exactly; reroll it and history runs again, differently. A question we left open, in case you want somewhere to start: which rules make the best ancestors — the founders whose mutated descendants stay interesting longest?