The Room.

It is called a germinal centre. A B cell enters carrying an antibody that binds the new thing badly. Inside, an enzyme called activation-induced cytidine deaminase is turned loose on the antibody’s variable region and mutates it at roughly 10⁻³ per base pair per generation — about a millionfold above the background rate in your other somatic cells. Concretely: about one mutation per variable region every three divisions. The cell divides, mutates, divides again, is tested against antigen and against helper T cells, and if it survives, it goes around again.

Now the part that is easy to miss, and which is the actual subject of this essay. The organ runs two separate death programmes, in two anatomical zones, for two different reasons. In the light zone, cells die by default if they are not positively selected — they failed the examination. In the dark zone, cells die because their antigen receptors were destroyed by the mutagen the organ itself deliberately switched on. Up to half of all germinal-centre B cells are dying every six hours, and the organ keeps two different ledgers for it.

That is not tolerance of error. That is error under management, with a budget, a boundary, and a body count.

Concede immediately the thing a weaker version of this essay would bury. The universe did not need error to make new things. For roughly 9.7 of its 13.8 billion years there was no copier anywhere in it, and in that stretch it produced spacetime structure, matter, the four forces as distinct phenomena, hydrogen and helium, then every element up to iron and past it, then galaxies and planets and the entire chemical inventory from which life was later assembled.

It did this by quantum fluctuation, gravitational instability, nuclear physics and spontaneous symmetry breaking. Not one of those involves a template. Planck’s measurement of the primordial spectrum — n_s = 0.965 ± 0.004 — is a measurement of the statistics of the original quantum noise, the fluctuation every galaxy in the sky still remembers. There was nothing to be unfaithful to. The word error had no referent.

So novelty is older than error, and any claim that error is the sole source of the new is refuted by two-thirds of cosmic history. But look at what that era could not do. A star makes carbon. It dies. The next star makes carbon again, from scratch, inheriting nothing. Symmetry breaking generates new states; there is no mechanism by which state n+1 receives and extends state n. The universe before replication was enormously productive and completely non-cumulative. It could invent. It could not keep.

What template replication introduced, around four billion years ago, was not novelty. It was retention — the first way a difference, once made, could survive the thing that made it. Tense called the gene the first arrangement that could carry a pattern past the death of its own body. That is the event. The universe acquired a past.

And here is the bill, which nobody negotiated and nothing can avoid. In a system that keeps, the only way a new thing enters the kept line is by keeping it wrong. Fidelity is precisely the property that makes retention possible, and precisely the property that forecloses novelty. One mechanism, both effects, no separation. Everything below is what four billion years have learned about living inside that.

Manfred Eigen wrote the upper wall in 1971. A replicating population can hold its information only while the master sequence’s replication advantage beats its own error rate. Push per-site error too high and, in Eigen’s own words, the information “melts like ice at 0 °C” — and he meant that technically: a first-order phase transition, with cooperative behaviour and unlimited coherence lengths.

There is a lower wall too, and it is the stranger one. A system that copies too well cannot move. It has nothing to offer selection. In evolutionary computation this is not an analogy but a theorem. Carsten Witt proved that for a simple evolutionary algorithm with per-bit mutation probability p = c/n, the expected time to the optimum is (1 ± o(1)) · (e^c/c) · n · ln n.

Both walls sit inside that expression. As c approaches zero the runtime blows up as 1/c — the system freezes. As c grows it blows up as e^c — the system dissolves. The minimum falls at exactly c = 1: on average, one error per copy. Deviations either way are punished superpolynomially.

Two entirely separate literatures arrive at the same shape. Ishii, Matsuda, Iwasa and Sasaki showed in 1989 that the evolutionarily stable mutation rate in a periodically changing environment is the reciprocal of how long the world stays put. Kussell and Leibler, in Science in 2005: the optimal switching rates then mimic the statistics of environmental changes. Three fields, three formalisms, one law. The right amount of error is set by how fast the world moves.

This is where honest writing about the subject gets difficult, and where nearly all writing about it goes wrong. It is tempting to say organisms tune their error rates into the window — that life discovered creativity and dialled itself to the edge of it. The comparative data says no, flatly.

Michael Lynch’s drift-barrier work shows mutation rates across the tree of life scaling inversely with effective population size and with coding target size. Fidelity is pushed as low as selection can manage; the only thing halting the descent is that below some point the benefit of further accuracy is smaller than the noise of drift itself. Chlamydomonas has been driven to about 6.8 × 10⁻¹¹ per site per division — better than almost every prokaryote. And Lynch is explicit about exactly the intuition this essay might otherwise indulge: there is no reason to invoke selection for evolvability.

The arithmetic closes it. If life sat near Eigen’s wall, the product of mutation rate and genome length would be near one everywhere. It is 0.67 for poliovirus — and about 0.001 for E. coli, a thousandfold below. Only RNA viruses come near, and even their proximity is now better explained as a byproduct of selection for replication speed than as tuning for adaptability.

So the window is real, and nothing steers into it. The honest description is this: populations are pushed towards the window from both sides — downward by mutational load, upward by clonal interference and environmental churn — and where they land is set by population size, generation time and recombination rate, none of which has anything to do with creativity.

You can watch the cornering happen. In Lenski’s long-term E. coli experiment, one population evolved a hypermutator that raised its point-mutation rate roughly 150-fold while there was still a great deal to discover — and was later invaded by lineages that brought the rate back down 40 to 60 per cent as the supply of useful discoveries thinned and the load became the dominant cost. Up when the world was new. Down when it was not. No foresight anywhere; just a balance moving underneath.

If global sloppiness is always punished, and novelty still has to come from somewhere, one architecture remains. Evolution found it at least twice, independently, and both cases are uncontested. Contingency loci: pathogenic bacteria carry short repeat tracts inside surface-protein genes that slip at about 10⁻⁴ per generation against a genomic background near 10⁻⁹ — a locally targeted hundred-thousandfold elevation, confined to exactly the genes facing an immune system that will not stop changing. And somatic hypermutation: the germinal centre, six orders of magnitude above background, in one cell type, one gene region, one anatomical structure, for about a week.

But the interesting thing is not the compartment. It is that the compartment needs four separate containment systems, and every one of them has been measured.

One — it restricts where the error happens, by address rather than by content. This is not how anyone assumes it works. Arbitrary passenger sequences, carrying no special motif, mutate at full variable-exon rates when placed at the variable-exon location. The targeting is cis-regulatory architecture: enhancers, convergent transcription, polymerase stalling. The room is a place, not a password.

Two — it cleans up the collateral. The mutagen does not stay put; it deaminates a wide swathe of the genes expressed in a germinal-centre B cell. Faithful base-excision and mismatch repair then erase the overwhelming majority of that damage. The antibody locus is the one address where that same repair machinery is deliberately switched into error-prone mode. The mutagen is global; the exemption is local. That is the actual mechanism of the room.

Three — it throttles the rate, and spends the budget on what is working. Cells receiving more T-cell help divide more — one to six divisions per dark-zone passage, set by how much antigen they captured — and mutate more in total. Among the most-divided cells, 35.9 per cent carried a specific affinity-enhancing mutation, against 6 per cent among the least-divided. The authors call it a feed-forward loop: hypermutation is greatest among the cells whose antibodies are already the best template for further diversification. The organ does not spray error uniformly. It spends its error budget preferentially on whatever is currently winning.

Four — it kills what it breaks, in the two ledgers we started with.

And then it stops. There is an affinity ceiling: once the antibody’s off-rate is slow relative to antigen internalisation, more affinity captures no more antigen, so the amount of peptide displayed to the T cell stops rising and the selection signal goes flat. The organ stops improving not because it runs out of mutations, but because its own scoring function saturates. Hold on to that sentence. It is the whole of section vii.

We have built a copier of extraordinary fidelity, and over three years we have been systematically closing the last places variance could live. In September 2025 Thinking Machines Lab published the clearest possible statement of where we are. They ran a frontier model at temperature zero — nominally deterministic — one thousand times on a single prompt, and got eighty unique completions. Divergence began at token 103: 992 continuations said the physicist was born in Queens, eight said New York City. They traced it to kernels whose floating-point reductions depend on batch size, rewrote them to be batch-invariant, and got a thousand identical outputs. Their closing line is the manifesto of the entire movement, and it is not a stupid one: we reject this defeatism.

Around that sits a whole architecture of narrowing. Structured outputs. Constrained decoding. Seeds. Temperature-zero defaults. Refusal training. Prompt caching. Each individually defensible; jointly, a programme for eliminating variance from the only machine we own that might produce something we did not put into it.

And it should be said plainly, because everyone assumes the opposite: this is not legally required. The full consolidated text of the EU AI Act contains zero occurrences of “deterministic” and zero of “reproducible” — the single near-match is a recital about copyright. Article 12 asks for logs. Article 15 asks systems to perform consistently, meaning stable across a lifecycle, not repeatable per call. The Federal Reserve’s SR 11-7 likewise contains no instance of “reproducible” or “replicable”; it asks for effective challenge and documentation a stranger could follow. The regulators asked for traceability. The industry heard determinism. That substitution was a choice, it was never argued for, and it is being made everywhere at once.

The costs are measured, and they are not small. Constraint destroys reasoning: applying a JSON schema dropped one frontier model on GSM8K from 86.99 to 23.44. The mechanism is not parse failure — parse errors ran at 0.148 per cent. It is this: every single JSON-mode response placed the answer key before the reasoning key. The schema ordered the model to answer before it was permitted to think.

Alignment narrows the distribution. Reinforcement learning from human feedback significantly reduces output diversity relative to supervised fine-tuning. In a controlled study, co-writing with an instruction-tuned model measurably homogenised essays across different authors, while co-writing with the base model did not — and the users’ own text was unaffected. The aligned model supplied the flattening. The same trade appears at population scale: writers given generative-AI ideas produced stories rated 5.4 to 8.1 per cent more novel individually, while the collective similarity of everything written rose. The authors name it correctly — a social dilemma. Individually better off, collectively a narrower world.

And when the variance runs out, learning stops. During reinforcement learning on reasoning models, policy entropy collapses and performance tracks it by an empirical law whose ceiling is predictable from the first few dozen steps. Across eleven models over 2,400-step runs, 73 per cent of all entropy consumption and 76 per cent of all performance gain occur in the first 200 steps — one twelfth of training. The rest yields almost nothing. One team describes the endpoint in a sentence that could have been written by a population geneticist: the sampled responses of certain groups tend to be nearly identical.

Train a model on its own output and the failure takes a shape Eigen would recognise — from the other side. Collapse, in the founding paper’s own words, “first starts with tails disappearing,” and then converges “to a point estimate with very small variance.” That is not the error catastrophe. It is its mirror: the fidelity catastrophe, the lower wall, information lost not to melting but to freezing.

The counter-case exists, it works, and it is instructive that it has exactly the architecture biology converged on. FunSearch found a new cap-set construction of size 512 in dimension 8 — a genuinely new mathematical object, published in Nature. Its architecture: a frozen language model run hot as a mutation operator, an island-model genetic algorithm, Boltzmann selection, and an evaluator that runs the program rather than trusting it. AlphaEvolve, on the same pattern, found a 48-multiplication algorithm for 4×4 complex matrix multiplication — the first improvement on Strassen in that setting in 56 years — and across fifty-odd open problems matched the best known construction about 75 per cent of the time and beat it about 20 per cent.

Joel Lehman and Kenneth Stanley named the move in 2022, before any of the headline results existed: the model can serve as a highly sophisticated mutation operator embedded within an overarching evolutionary algorithm.

Note the shape. High variance, walled into one component. A cheap, hard, automatic filter on the exit. Most output discarded — FunSearch’s cap set emerged in 4 of 140 experiments, a hit rate that would be humiliating anywhere except biology, where it is ordinary.

And it inherits the same ceiling. The AlphaEvolve authors state their own limitation: the method works only where an automated evaluator can be written. Which is the lymph node’s problem exactly. The room is only ever as good as the guard on its door — and a guard whose scoring function saturates stops discriminating, at which point the mutation budget is being spent on nothing. Reward hacking, verifier saturation and entropy collapse are one failure, and the immune system hit it first, in 1998.

Determinism produces novelty, routinely. Full-batch gradient descent with zero stochasticity reproduces the generalisation of stochastic gradient descent — 95.68 per cent against 95.70. A small transformer trained on modular addition was reverse-engineered and found to have invented a discrete Fourier transform algorithm nobody supplied. Chain-of-thought prompting gave one model 39 extra points on GSM8K at greedy decoding. If injected randomness were the channel, none of this should happen. So the defensible claim is not about randomness at all. It is about the entropy of the reachable output distribution — a different quantity, and the one every measurement in section vi actually targets. Say that, or say nothing.

The upper wall’s mechanism is misnamed, including in most popular writing on this subject. Lethal mutagenesis and error catastrophe are not the same process — and on Eigen’s own landscape an error catastrophe can retard extinction, by shifting the population onto mutation-robust genotypes. The wall is real; the usual account of why is wrong, and this essay would have printed the wrong one without being checked.

Most biological novelty is not point mutation. In prokaryotes, gene flux exceeds substitution by orders of magnitude, and horizontal transfer — not duplication — drives between 88 and 98 per cent of protein-family expansion. The single largest innovation in four billion years, the eukaryotic cell, was an acquisition, not a typo. This objection is survivable only if error means any infidelity in transmission — slippage, unequal crossover, template switching, non-homologous end joining — rather than base substitution. That is the definition used here, and it is stated rather than smuggled.