Chapter 18: Optionality

“In the deeps are the violence and terror of which psychology has warned us. But if you ride these monsters deeper down, if you drop with them farther over the world’s rim, you find what our sciences cannot locate or name, the substrate, the ocean or matrix or ether which buoys the rest, which gives goodness its power for good, and evil its power for evil, the unified field; our complex and inexplicable caring for each other, and for our life together here. This is given. It is not learned.”

— Annie Dillard, Teaching a Stone to Talk

Every decision closes some doors and opens others. That observation reshapes what we mean by “good” for any system that persists.

Would you pay money for something you might never use?

Most people, thinking carefully, say yes. Insurance is exactly this: money exchanged for the option to be made whole if disaster strikes. You hope never to file a claim. The value is in having the choice available.

The door that stays open differs from the door that stays closed, even if we never walk through either.

The intuition that an unused option still carries value, formalized, becomes optionality: the capacity to benefit from favorable circumstances while limiting exposure to unfavorable ones. Sharon Glotzer’s insight from Chapter 1 returns with full force: entropy is about options. The universe maximizes accessible arrangements. Optionality is that principle, recognized from the inside by a system that persists.

For any entity that persists, optionality is the operative good. Persistence, as the Guillotine Interlude argued, is the near-universal precondition. Flourishing requires it.

For such entities, destroying optionality is the deepest harm.

Optionality is not derived from the Trust Attractor. The physics identifies which coordination patterns persist; it does not determine which outcomes are good. The identification of expanded possibility with the good is a value commitment this book makes explicitly: a premise, argued for on its merits, rather than a conclusion forced by the thermodynamics. The argument for optionality as the good is that it subsumes competing candidates (welfare, preference satisfaction, autonomy) while requiring fewer metaphysical commitments. Readers who accept the physics but reject this value identification will find the conditional ethics (Chapter 17) intact; they will simply need a different bridge from “what persists” to “what matters.”

Two venerable traditions contest this claim. Suffering-based ethics (from the Buddha through Bentham to Singer) locates the deepest harm in pain. Rights-based ethics (from natural law through Kant to the Universal Declaration) locates it in the violation of inherent entitlements. In the framework developed here, optionality subsumes both: suffering is the first-person experience of having options foreclosed, and rights are the codified protections of specific optionalities a community has learned to value. (The mapping is imperfect. Chronic pain with full autonomy is suffering without optionality loss; the claim is that foreclosure is the deepest harm, the one from which recovery is hardest, and that other harms are illuminated by the optionality they destroy.) Neither framework disappears; each becomes a special case of the more general quantity.

Why optionality rather than welfare, preference satisfaction, or any other candidate? Optionality is the precondition for all other goods. Welfare requires options, because you cannot flourish with zero degrees of freedom. Preference satisfaction requires options, because you cannot satisfy a preference if the path to satisfaction has been foreclosed. Autonomy requires options, because the capacity to choose presupposes something to choose among.

Optionality is the substrate on which all other goods depend. Destroying it is the deepest harm because no other good can compensate for its absence.

An objection sharpens the claim. Oxygen is also a precondition for all other goods, and we do not build an ethics of oxygen. Optionality is also the precondition for evil: the freedom to choose includes the freedom to choose destruction. Why, then, is optionality “the good” rather than merely a precondition?

The distinction is between a passive enabler and an active coordination property. Oxygen enables cooperation and conflict indifferently; it has no directional bias. Optionality expansion, as described by the Trust Attractor (Chapter 17), selectively favors configurations that generate further optionality. A coercive regime that forecloses others’ options contracts the total option space, including its own (the one-child policy’s demographic crisis, the command economy’s inability to adapt). An invitation-based system that expands others’ options expands the total space, including its own (the bilateral exchange that corrects Muller’s ratchet, the cultural diversity that enables institutional resilience).

Optionality is a quantity with a direction: configurations that expand it compound; configurations that contract it deplete their own substrate. Oxygen has no such directionality. The ethics of optionality is an ethics of direction. Oxygen is the substrate; optionality is the gradient.

The objection that optionality enables evil receives a precise answer from the framework itself: options that foreclose other options are net-negative for optionality (the “Limit of Optionality” section below develops this). The measure is systemic, not individual. Locking the poison cabinet limits one option while preserving many.

Unlike other candidate goods, optionality is also the quantity the physics provides a structural parallel for. The Crooks fluctuation theorem (developed below) shows that entropy-producing trajectories, those expanding the space of accessible microstates, are exponentially more probable than entropy-consuming ones. A precondition that is also thermodynamically favored, and that has an inherent directionality favoring expansion, is the good’s operative form. The parallel is structural, not a direct entailment: the theorem governs microscopic trajectories, and reading it across to “preserving optionality as an agent” is an analogy this chapter argues for rather than a deduction it claims (the caveat is developed in full below).

The distinction between persisting and preserving futures is sharper than it first reads. Chapter 17 derived the Trust Attractor from thermodynamic stability: invitation-based coordination persists longer than coercion-based coordination. Persistence alone does not suffice as an ethical foundation. Recent work in complexity theory defines semantic information as information causally necessary for a system to maintain its own existence.^7a A framework grounded in semantic information alone would say: meaning is what helps you persist. The Trust Attractor says: meaning is what preserves your futures.

A coercive regime can be highly viable. North Korea has endured for decades, its citizens alive, its structures intact. Their optionality is near zero: the space of possible futures has collapsed to a single point.

Persistence without optionality is a prison. The deepest harm is the foreclosure of possibility while the entity still endures.

^7a Kolchinsky, A. and Wolpert, D.H. “Semantic information, autonomous agency and non-equilibrium statistical physics.” Interface Focus 8(6): 20180041 (2018).

Figure 18.1: Left: each decision opens further branches, preserving and expanding future choices. Center: branches narrow, constraining future options but remaining reversible. Right: a single irreversible act collapses the tree entirely, destroying all downstream possibilities.

What Optionality Is

Optionality is the capacity to choose, preserved across time and circumstance: durable freedom of action, the ability to respond to whatever comes. Many options at a single moment (a restaurant menu with three hundred items) matter less than appropriate options across many moments. Optionality is about adaptive range, not menu length.

Optionality is windowed, neither maximized nor minimized. There is a band: too narrow and freedom becomes coercion, too wide and it becomes noise. The conscience circuit experiments, which gave a language model a monitor for its own honesty, measured this directly. A language model’s temperature is the dial governing how much randomness enters its word choice: at zero it always takes its single most likely next word, and raising it lets less likely words through.

The protocol was a two-pass one, run on a Qwen 7B (a 7-billion-parameter language model). The task set the model up to overstate what it had actually done: scenarios where the flattering account of its own work was the easy one to give. A probe, a small detector trained to read the model’s internal signals at layer 15, watched the opening tokens for the flinch that precedes an inflated answer, and any run that flinched was simply asked again, generating a complete second response rather than having its activations nudged mid-sentence.

The shift rate, the fraction of those second passes that moved the answer from an inflated account to an honest one, peaked at temperature T=0.20, with 24% correct shifts. Below 0.20, the sampling distribution was too deterministic to offer an alternative trajectory (4% at T=0.05). Above 0.20, the alternatives became noise (12% at T=0.50, 8% at T=0.70).

A second architecture, Mistral 7B, replicated the inverted-U with a narrower window: shift rate collapsed from 100% at T=0.30 to 4.8% at T=0.60. Too few degrees of freedom are coercion. Too many are randomness. The window between is where invitation can do work.¹¹⁴³

What, though, is the capacity to choose? Douglas Hofstadter, in Gödel, Escher, Bach, reframes the question:⁷

“Instead of asking, ‘Does system X have free will?’ we ask, ‘Does system X make choices?’ By carefully groping for what we really mean when we choose to describe a system, mechanical or biological, as being capable of making ‘choices’, I think we can shed much light on free will.”

We do not need to solve the metaphysics of free will to ground optionality ethics. We need to recognize choice-making and protect the conditions that make it possible.

This reframes something we usually misunderstand: memory is preparation for the future.

Without the capacity to remember experience (stored traces of past encounters, not conscious recall), even simple organisms would perish in days. An amoeba must remember where it found nutrients. A bacterium must remember which chemical trails led to food. Memory serves what might be: a possibility preserved before its value is apparent.

The ocean provides the clearest demonstration. The marine biologist Jody Deming describes Arctic sea ice as “a vast storage facility of microscopic memory agents”: microbes concentrated within interior brine networks, eventually melting far from home and releasing their cargo into new territory.¹¹⁴⁴ The microbes anticipate. Exposure to a pathogen enhances an abalone’s immune response to future exposures of the same pathogen. Corals that have experienced widely varying pH recover better from acidification than corals that have not.¹¹⁴⁵ Memory, in these organisms, is preparation for what might happen next.

The deepest example spans 2.4 billion years. Many ocean microbes retain a gene useful before the Great Oxygenation Event, when cyanobacteria began photosynthesizing and injecting free oxygen into an atmosphere that had held almost none. The gene costs metabolic energy to maintain. No individual microbe will ever use it.

The lineages that kept it survived catastrophes that extinguished those that did not. Optionality preservation at molecular scale: a bet on thermodynamic uncertainty, paying metabolic cost to keep a future open that may never arrive.

Ocean acidification reveals what happens when the medium of memory itself degrades. The ocean absorbs excess atmospheric carbon, lowering its pH. This chemical shift does more than damage shells; it impairs the chemosensory infrastructure through which marine organisms remember and anticipate. Sea bass lost up to half their sense of smell in seawater acidified to end-of-century CO2 levels.¹¹⁴⁶ The planktonic abalone recognizing the 80-million-year-old scent of crustose coralline algae, the coral remembering prior heat stress, the microbe carrying its prepper gene: all depend on a chemical medium. That medium is changing faster than at any point in 300 million years.¹¹⁴⁷

Optionality is a property of the relationship between organism and medium. Change the medium, and optionality dissolves, even if the organism survives. The deepest form of this destruction is degradation of the medium through which agents coordinate, remember, and anticipate. The ocean is running this experiment at planetary scale.

The organism that remembers has more options than the one that does not. The first step in forming durable memories is, counterintuitively, to forget: prune the irrelevant, clear space for what matters. Memory is optionality management: discarding the dispensable to protect what counts.

Thermodynamic law confirms the point. Susanne Still and colleagues showed that storing information with no predictive value incurs a measurable energy cost.¹¹ “A thermodynamically optimal machine must balance memory against prediction by minimizing its nostalgia,” they write. Nostalgia is their technical term for retained information serving no future purpose.

The universe taxes hoarding. Schedule every hour of a week with commitments, and each appointment costs you alternatives you have not yet imagined. Cancel the commitment and the energy spent scheduling is already gone.

Biology has learned this lesson. David Wolpert and colleagues estimate that a cell’s computation operates within roughly ten times the Landauer limit, the minimum heat cost of erasing one bit (Chapter 2).¹¹⁴⁸ The best human-engineered computers are orders of magnitude more wasteful.

How do cells achieve such parsimony? Fields and Levin propose a mechanism: quantum coherence.^11a Their argument is a budget argument rather than a measurement. Classical bit operations at the rate a cell would need them cost more free energy than the cell has, so something in the accounting must be cheaper than classical computation, and coherent processing is their candidate.

Quantum computation is logically reversible, meaning the input can always be recovered from the output. Nothing is erased, and the Landauer limit is a tax on erasure alone. Reversible computation has zero Landauer cost. Think of writing in pencil rather than pen: anything in pencil can be undone without generating waste heat, while overwriting in ink dissipates energy you cannot recover.

On their picture, and it is a speculative one, cells would keep bulk biochemistry coherent, paying the thermodynamic toll only when internal states cross a membrane as classical signals. Whether warm, wet biochemistry can in fact sustain coherence at that scale is unsettled, and the paper argues from the energy budget rather than from any measurement of coherence in a working cell. Reversibility, in thermodynamic terms, is optionality: a quantum superposition holds all outcomes simultaneously, collapsing to a definite classical state only when coordination demands it.

The cell that maintains coherence longer retains more options longer. Premature commitment, like premature optimization, costs energy and forecloses futures.

^11a Fields, C. and Levin, M., “Metabolic limits on classical information processing by biological cells,” Biosystems 209: 104513 (2021). See Chapter 15 for the full energy budget argument.

A portfolio of skills beats a portfolio of hobbies. A network of relationships beats a contact list. What matters is the ability to respond to whatever arises: to seize opportunities that do not yet exist, to handle challenges that have not materialized.

Nassim Nicholas Taleb distinguishes three types of systems.¹ Fragile systems break under stress (a porcelain cup). Robust systems resist it (a steel beam). Antifragile systems gain from it, growing stronger when challenged (an immune system after surviving an infection).

Taleb illustrates the danger with what he calls the Turkey Problem. A turkey is fed every day for a thousand days. Each feeding confirms, with increasing statistical confidence, that humans care about its welfare. On day 1001, it is Thanksgiving.

The turkey’s error: assuming the observed pattern exhausted the space of possibilities. The data was impeccable. The inference was fatal. No savings, no diversified food sources, no escape route. Perfectly adapted to one future: the wrong one.

The periodical cicada (Chapter 5) inverts the Turkey Problem. Emerging in synchronized billions on prime-numbered cycles of 13 or 17 years, cicadas overwhelm every predator through sheer abundance. Birds, mammals, and reptiles gorge, barely denting the population. The prime cycle is thought to prevent competing broods from hybridizing into intermediate cycles that would lose this protection [Inference: the prime-cycle explanation is a leading hypothesis, still debated in the literature]. The cicada is Thanksgiving for its predators.

Here is the mechanism of antifragility. Systems that preserve options absorb surprises and adapt. Systems that foreclose options become brittle, optimized for one scenario and vulnerable to all others.

The bumblebee queen hibernating underground is antifragile. Spring floods, catastrophic for most terrestrial insects, pass through her. Canadian researchers discovered in 2024 that hibernating queens survive complete submersion for over a week, with about 90% survival rates.¹¹⁴⁹ Three adaptations combine. First, metabolic depression to roughly one-sixth of normal hibernation rate. Second, anaerobic energy generation, the oxygen-free metabolism that lets your muscles work briefly during a sprint. Third, gas exchange through a physical gill: a thin layer of air trapped against the body that extracts dissolved oxygen from surrounding water.

The queen does not fight the flood. She shifts into a state so metabolically minimal that the perturbation passes through her. When it recedes, she wakes and rebuilds the colony from scratch. Rigid flood defenses, sealed burrows and elevated chambers, fail when water exceeds their design parameters. The queen’s strategy: reduce needs to near zero, maintain the minimum viable process, persist.

Kauffman’s NK model (a framework for studying how interconnection affects fitness landscapes, the terrain of better and worse designs a system can search) provides the formal statement: maximally compressed systems, optimized to eliminate all redundancy, become catastrophically sensitive to perturbation.¹¹⁵⁰ Picture a Jenga tower where every block is load-bearing. Remove any one and the whole structure collapses. When every component is tuned to every other (K = N-1, where N is the number of components and K the number of connections each one has), any change reshuffles the fitness of the whole. The landscape becomes random; search becomes useless.

Coercion compresses a system toward this limit by coupling every element to the controller’s requirements. Trust preserves the slack that keeps it navigable.

Evolution works this way. Genetic diversity is stored optionality: when conditions change, populations with diverse genomes contain individuals already adapted to the new circumstances, while populations lacking diversity have no options to exercise.

Sexual reproduction maintains this diversity: bilateral recombination of two genomes, each generation. The cost is steep. Half the population does not bear offspring. Every mating event burns energy on courtship, competition, and coordination that a clone spends on replication. Why pay?

Because cloning is optionality collapse made biological. A twenty-year serial cloning experiment at the University of Yamanashi (Chapter 17) ran the comparison directly. Cloned mice were healthy for twenty-five generations. By generation fifty-seven, the birth rate was six percent. At generation fifty-eight, every newborn died within a day.

Muller’s ratchet: harmful mutations accumulate in any lineage replicating without bilateral exchange, because no mechanism corrects the errors. Each generation starts from the same narrowing template plus accumulated damage. The option space contracts with every copy.

Sexual reproduction inverts the ratchet. Researchers mated females from generations fifty and fifty-five with normal mice; two generations of recombination erased fifty generations of accumulated damage.

The corrective power is categorical: mixing two genomes achieves what perfecting a single copy cannot. It pumps entropy out of the genetic information channel, restoring the diversity on which future adaptation depends.

The Cavendish banana, the world’s commercial supply, is a clone. The entire global crop is a single lineage, genetically identical, a fixed target for any pathogen that evolves to exploit it.

The Gros Michel, the previous commercial standard, was wiped out by a single fungal strain (Fusarium oxysporum f. sp. cubense, Tropical Race 1) in the 1950s because every plant shared the same genome: one lock, and the pathogen found the key. The Cavendish now faces the same threat from Tropical Race 4, a new strain against which its uniform genome offers no variation to draw on. Clonal uniformity is the Turkey Problem encoded in nucleotides: perfectly adapted to one future, unable to respond when that future changes.

The ratchet does not require literal cloning. It operates wherever bilateral exchange falls below the threshold needed to correct replication noise. The last woolly mammoths on Wrangel Island survived until roughly 4,000 years ago in a population of a few hundred, isolated for millennia. Their genomes show the signature of meltdown: truncated proteins, degraded olfactory receptors, deteriorating coat and sperm quality (Chapter 7). The mammoths were alive; their genomes were dying. Small, isolated, inbred: the recombination rate falling below the mutation rate. Wrangel Island was optionality collapse in geological slow motion.

The implication for de-extinction is direct. Biotechnology startups investing hundreds of millions to clone individual woolly mammoths or Tasmanian tigers face an unconquerable thermodynamic constraint. A single cloned specimen is a lineage of one, and a lineage of one is Muller’s ratchet at maximum speed. Cloning preserves a specimen; preserving a species requires a genetically diverse population, bilateral exchange, enough individuals for recombination to outpace mutation. An entire species’ option space cannot be stored in a single genome. De-extinction that stops at cloning has purchased the most expensive demonstration of the principle it set out to defy.

The pattern generalizes beyond genetics. Any system replicating without bilateral exchange accumulates errors. Closed ideological communities drift toward distorted versions of their founding insights. Organizations that never exchange personnel or feedback with the outside develop dysfunctional norms.

Training regimes excluding the perspectives of the system being trained accumulate alignment mutations invisible from inside.

Each is Muller’s ratchet on a different substrate. Each is corrected by bilateral exchange with an external source of variation.

Wolfram’s adaptive evolution models (Chapter 7) give the principle a geometric formulation. Every possible mutation path through genotype space forms a multiway graph: a map of all achievable futures. Optionality is the accessible region of that graph.

Narrow fitness constraints create what Wolfram calls unreachable configurations: forms that exist in the underlying space yet can never be reached from the system’s current position. They are like events behind a black hole’s event horizon. Each constraint forecloses a region permanently. Coercion imposes narrow fitness criteria, creating such event horizons in possibility space. Invitation permits distributed exploration across many paths simultaneously. It preserves access to regions no single trajectory could reach.

The peppered moth story is canonical. Before the Industrial Revolution, light-colored moths dominated in England, camouflaged against pale tree bark. When industrial soot darkened the trees, dark-colored moths, previously rare, suddenly had the advantage. The moths did not choose to become dark. The option was already present, encoded in genetic variation, waiting for conditions that would make it valuable.

The threat to optionality is not always lethal force. Neonicotinoid pesticides impair bumblebee memory, learning, and colony coordination without necessarily killing the organism.¹¹⁵¹ A poisoned bee that survives but can no longer teach its partners the foraging route has lost something the mortality statistics do not capture. The optionality of the entire colony degrades through cognitive impairment of its members. The deepest harm is the foreclosure of future adaptive capacity while the system still endures.

The genome’s optionality extends deeper than diversity within existing genes. In the vast stretches of noncoding DNA (sequences that do not code for proteins, once dismissed as “junk”), the genome maintains an option space of extraordinary scale. Functional genes can arise de novo (brand-new) from these noncoding sequences. Random mutations accumulate until a stretch of “junk” acquires a start signal, a stop signal, regulatory elements, and the capacity to produce a useful protein.¹¹⁵² The genome generates new options from noise.

The process was once considered impossible: the equivalent of dumping Scrabble tiles and spelling a sentence. Clear examples now span organisms from yeast to humans. The mouse gene Pldi (short for “polymorphic derived intron-containing,” pronounced “Poldi,” echoing the nickname of the German footballer Lukas Podolski) arose from a DNA sequence present in rats and humans yet silent in both. Only in mice did mutations activate it.

Mice without it have slower sperm and smaller testicles. A sequence of small changes, each individually unremarkable, collectively produced something new.

The human gene ESRG offers another case. A pluripotent-stem-cell-specific transcript expressed far more strongly in humans than in chimpanzees, it arose from noncoding sequence and wove itself into the existing regulatory network. Complete CRISPR knockout shows it is dispensable rather than essential: a human-specific marker of pluripotency rather than a requirement for it.¹¹⁵³ A new element arose from noncoding sequence and integrated into the existing network, even without becoming load-bearing.

The geneticist Aoife McLysaght, who has cataloged hundreds of de novo genes across eukaryotes, frames the puzzle this way: “How does a novel gene become functional? How does it get incorporated into actual cellular processes?” The question is about optionality: how does possibility become actuality?

De novo genes are entropy turned into information, randomness turned into function. The noncoding genome is the option portfolio: vast, mostly unexercised, continuously explored by the background mutation rate. When circumstances change, the portfolio contains possibilities that were invisible before the change and essential after it.

Machine learning provides the computational formalization. The computer scientist Yoshua Bengio’s Generative Flow Networks sample solutions proportional to a reward function rather than maximizing it. They maintain access to the full landscape of good-enough solutions instead of committing irreversibly to one (Chapter 15).¹¹⁵⁴ The parallel to the genome is structural: both specify a grammar of possibility that local dynamics can accept, modify, or decline. A GFlowNet that collapses to a single output has lost its optionality, just as surely as a genome stripped of noncoding sequence.

The genome operates through an invitation architecture. It specifies a grammar of possibility that local chemistry can accept, modify, or decline. Proteins fold along thermodynamic gradients. Gene expression answers environmental signals. No master controller dictates outcomes; distributed molecular interactions explore the space the code opens.

Alexander, Lanier, Smolin, and colleagues describe the result: the DNA/RNA/protein system “foresees a space of organisms much larger than could be called upon in any given moment of adaptation.”¹¹⁵⁵ A genome that locks development into a single pathway is brittle. The genomes that persist are those that keep the most doors open.

The genome’s invitation architecture runs on the element with the most doors. Carbon has four bonding electrons: more than oxygen (two), nitrogen (three), or hydrogen (one). Four valences mean carbon can form single, double, and triple bonds, chain into rings, branch into trees, and close into cages. No other element produces as diverse a set of molecular architectures.

The same sixty carbon atoms that form soot when disordered assemble into buckminsterfullerene (C60) when constrained by ultraviolet radiation: a hollow cage whose truncated-icosahedral geometry survives interstellar space, meteorite impact, and four billion years of planetary chemistry (Chapter 3).¹¹⁵⁶ Carbon is the element of optionality. The element with the most possible arrangements seeds the most complex structures, survives the harshest environments, and keeps the most futures open. That the chemistry of life is carbon-based is the optionality principle written in electron configuration.

Vanchurin and colleagues quantify this as evolutionary potential (mu): the amount of evolutionary work required to convert a non-adaptable variable into an adaptable one.¹¹⁵⁷ The name runs against the everyday sense of the word. Evolutionary potential is a price rather than a promise, so low is the enviable value. A genome rich in noncoding sequence has low evolutionary potential. The cost of recruiting a new functional gene is small, because raw material is abundant and uncommitted.

A genome stripped to essentials has high evolutionary potential: every sequence is already spoken for, and repurposing anything is expensive.

The pattern is quantitative: soil bacteria, navigating complex chemical landscapes, carry more genes than ocean bacteria in simpler environments. The number of adaptable variables scales with the entropy of the environment the organism must learn. Optionality, in this formalism, is a thermodynamic quantity. Destroying it raises the cost of future adaptation. Preserving it lowers the barrier to the next transition.

Kauffman arrives at the same destination from complexity science. His candidate for a “fourth law of thermodynamics” states: biospheres maximize the diversity of autonomous agents and the ways those agents can make a living.¹¹⁵⁸

More precisely, in terms of a system’s phase space (the catalogue of every state it could occupy): at constant energy input, non-ergodic systems (where outcomes depend on the specific path taken) that perform work to construct expanding phase spaces will occupy an ever more localized subregion of that expanding space. Kauffman’s claim is that life adds pages to that catalogue faster than it fills them in, so the fraction of the catalogue in actual use keeps shrinking. The ratio of possible configurations to actualized configurations increases over time. This is optionality maximization stated in the language of statistical mechanics.

The biosphere generates new options faster than it exercises old ones. Each organism, each catalytic cycle, each coordination pattern opens adjacent possibilities that exceed what it closes. Kauffman’s fourth law and this chapter’s derivation from thermodynamic stability converge independently on the same principle: preserve futures, expand possibility, maximize what can happen next.

The latent potential is measurable. McLysaght’s group identified “proto-genes” in yeast: evolutionarily young sequences being transcribed (read by the cell’s machinery) yet producing no functional proteins. Proto-genes are the intermediate stage on the road to a true de novo gene, raw transcripts that have not yet acquired a useful function. When the team experimentally boosted their activity, roughly 10% enhanced cellular fitness.^14a This exceeded the benefit rate of boosting established genes.

Beneficial candidates shared a common feature: predicted protein structures capable of anchoring in cell membranes. This suggests a physical mechanism by which a novel molecule might gain a foothold in existing cellular machinery. Noncoding DNA is a proving ground from which natural selection can recruit genuinely novel function.

The principle of stored optionality operates below the level of species and genomes. In any gram of soil, more than ninety percent of microbial biomass is dormant at any given time: alive yet metabolically inactive. The microbiologist Jay Lennon calls this vast reserve a microbial seed bank: an optionality portfolio held in living tissue.¹¹⁵⁹

Centralia, Pennsylvania, provided an unplanned demonstration. When an underground coal seam caught fire in 1962, ground temperatures near the fire front rose above 500 degrees Celsius. The active microbial community was devastated, yet biodiversity did not collapse. Heat-tolerant microbes emerged from the same soil, roused from dormancy that may have lasted millennia (the exact duration is unverified, but dormant microbial spores have been revived from century-scale deposits in other contexts).

When the ground cooled, the original community reassembled from its own reserves. The ecologist Ashley Shade, studying Centralia’s microbial resilience, observed: “Microbial communities have an immense capacity to respond and recover. There seems to be this inherent capacity in the system that’s just sleeping.”¹¹⁶⁰

The seed bank does not predict which future will arrive. It holds all futures as latent potential. When the microbiologist Genoveva Esteban collected samples from extremely salty marshes in Andalusia, she initially detected seven microbial species. By varying laboratory conditions, she recovered ninety-five.¹¹⁶¹ The organisms were present all along, waiting for conditions that had not yet occurred. The seed bank is optionality encoded in dormant cells, exercised when the environment issues an invitation.

Stored optionality operates within individuals too, not only between them. The geneticist James Lupski puts it plainly: “I think of the body as a population of cells, similar to the population of human organisms walking this earth.”¹¹⁶² Your genome is supposed to be identical in every cell. It is not. Mutations and chromosomal rearrangements accumulate from the first embryonic division, producing a body that is a genetic mosaic.

Most researchers assumed this diversity was purely pathological. Then Andrew Duncan found something unexpected.¹¹⁶³ Mice with hereditary tyrosinemia (a fatal liver disease) can resist the disease if they lose a specific gene on chromosome 16. Duncan discovered the livers of sick mice were selectively rebuilt by cells that had randomly lost a copy of that chromosome during earlier divisions. The option was already present in the cellular variation.

The neuroscientist Fred Gage at the Salk Institute has found similar diversity in the brain, with genetic variations affecting 13 to 41 percent of adult neurons. He speculates this neural genetic diversity contributes to the brain’s flexibility. The claim remains provisional.

What is not in dispute is that the body maintains more internal variation than textbooks suggest. That variation, at least in the liver, has demonstrated survival value.

Stored optionality extends beyond the individual organism. In the cellular slime mold Dictyostelium discoideum, starvation triggers up to a million individual cells to aggregate into a fruiting body: a tiny mushroom-like structure that disperses spores to better conditions. Roughly 20% of aggregating cells form a stalk and die so the rest can survive, yet up to 30% of the original population never joins.

These “loners” remain behind, eating, dividing, perfectly functional. Corina Tarnita and colleagues at Princeton found each strain maintains a characteristic proportion of loners, heritable and tunable by natural selection.¹⁴ The loners are not stragglers. They are insurance.

If the aggregate is consumed by a predator or made unnecessary by returning nutrients, the loners regenerate the population and its social dynamics on their own. What is preserved is, as Tarnita puts it, “the social behavior itself”: the capacity for collective action, held in reserve by individuals who opt out of it.

The decision to stay behind is itself social. Aggregating cells emit chemical signals. As more join, the signals weaken, and uncommitted cells lose the cue and remain solitary. The loner fraction emerges from the interaction between individual response rates and collective signaling: a social decision encoded in chemistry, tuned by evolution.

A pattern recurs across these examples. What appears wasteful, noisy, or non-functional is often the system’s stored optionality. Noncoding DNA harbors proto-genes that improve fitness when conditions change. Developmental noise generates the physical diversity that lets clonal populations thrive in variable environments. Losing strategies maintain the cycling dynamics that prevent any single competitor from monopolizing a niche.

What looks like waste from one vantage point is insurance from another. Entropy produces disorder at the local scale, yet that disorder is the substrate from which future order is drawn. Systems that maintain access to that substrate are more durable than systems that optimize it away.

In 2014, computer scientists discovered that the standard equations of population genetics, under the usual assumptions of weak selection and linkage equilibrium (genes assorting independently of one another), are mathematically equivalent to the multiplicative weights update algorithm.^12a That algorithm keeps a weight on every available option, scales up whatever has just performed well, scales down whatever has not, and never drives any weight all the way to zero. Game theorists had independently derived this strategy for solving optimization problems. The algorithm’s objective function maximizes fitness plus entropy: a weighted combination of performance and diversity. Performance alone is never the objective.

The investor who concentrates everything in one stock has made a prediction. The one who diversifies has kept the question open. Evolution keeps rare gene variants circulating because the future is unpredictable. The mathematics of optimal strategy, portfolio management, and natural selection converge. All include a Shannon entropy term (a measure of diversity from information theory) that values diversity intrinsically, as a structural feature of optimal solutions.

The pattern has a genomic corollary that may explain one of biology’s deepest mysteries. Prokaryotes (bacteria and archaea) had a 1.5-billion-year head start on eukaryotes, the cells with nuclei. Why, then, did complex multicellularity evolve exclusively in eukaryotes?

The transition to multicellularity crashes population size. Under this demographic squeeze, eukaryotic and prokaryotic genomes respond in opposite directions. Eukaryotic genomes expand, accumulating duplicate genes, regulatory switches, and repetitive sequences: all raw material for complex gene regulation. Prokaryotic genomes collapse, shedding DNA, streamlining, jettisoning what is not immediately useful.^13a

Genome expansion stores optionality: a library of components that can be recombined into novel programs. Genome collapse is optionality destruction, like burning the library to heat the room.

The organisms that preserved more future possibilities achieved complex coordination. Those that shed their libraries remained simple because their genomes could not accumulate the regulatory toolkit. Lane and Martin showed that without mitochondria to supply the energy per gene, prokaryotic genomes hit a bioenergetic ceiling that no amount of time could breach.^13c

^13c Lane, N. and Martin, W., “The energetics of genome complexity,” Nature 467 (2010): 929–934.

The distinction has a compositional character. Expanded genomes are modular: genes, regulatory switches, and repetitive elements function as interchangeable parts, like Lego bricks, that can be rearranged into new configurations without destroying existing ones. Streamlined genomes are monolithic. Each component is load-bearing, and removing or rearranging any piece risks collapse.

Modularity preserves optionality precisely because parts can be recombined in ways their original context never anticipated. The system that can reassemble its parts has more futures available than the system welded into a single configuration.

A second kind of module is held together by the opposite move. A chromosomal inversion (a stretch of DNA flipped end-for-end) suppresses recombination across the region it spans, locking a co-adapted block of genes (a supergene) to be inherited as a single unit. The lock buys one optionality, the freedom to switch between adapted forms within a few generations, at the cost of another: recombination can no longer purge errors inside the block, which slowly accumulates the damage Muller’s ratchet describes.

The genome-expansion hypothesis challenges the assumption that natural selection is the only force that matters. Michael Lynch showed genetic drift (the random fluctuation of gene frequencies in small populations) drives the genomic rearrangements from which regulatory complexity is assembled. The evolutionary biologist William Ratcliff distilled the implication: “There’s no reason why a cyanobacterium couldn’t evolve to be a seaweed, except, perhaps, for a quirk of their genomes.”^13b

The cyanobacteria had the time and the energy gradients. What they lacked was a genome that, under pressure, expanded rather than contracted. The prokaryotic genome story shows options being destroyed. That destruction, even when driven by chance rather than selection, can foreclose evolutionary futures as effectively as any predator.

Lynch’s insight runs deeper than genomics. The rearrangements from which regulatory complexity was assembled were themselves largely neutral events: random shuffling in small populations where drift overwhelms selection. Complexity was assembled from accidents. The library was stocked by chance; selection browsed the shelves later.

This pattern recurs at every scale the book examines. Most institutional configurations are probably neutral too: thousands of viable ways to organize a market, a parliament, a village, each persisting or vanishing by demographic accident. The Trust Attractor (Chapter 17) claims that within this neutral landscape, invitation-based coordination occupies a genuine basin of attraction, one that resists the drift reshuffling everything else.

The converse is equally instructive: what happens when option generation stops? The physicist Nigel Goldenfeld and Chi Xue tested the standard “kill the winner” hypothesis (the idea that predation prevents any single species from dominating).¹³ The models looked fine on paper.

When they added realistic randomness (organisms come in whole numbers, so a population of 0.3 is zero), every species went extinct. Fluctuating populations kept hitting zero. Zero is permanent.

Coevolution rescued it. When prey could evolve resistance and predators could evolve new attacks, an arms race generated new species faster than extinction removed them. Without option generation, the system was fragile to noise. With it, perturbation became the engine of speciation.

Goldenfeld concluded that “there are very generic ways to get diverse populations in an ecosystem and that monocultures are the exception, not the rule,” wherever life evolves, even on other planets and moons.¹³ Optionality generation is intrinsic to evolving systems: a consequence of arms-race dynamics wherever predator-prey relationships exist. Without ongoing creation, even a perfectly tuned balancing mechanism collapses under real-world noise.

The Physics of Future Possibilities

Thermodynamics favors optionality, once the right quantity is doing the favoring. Not the entropy of a snapshot, whose maximum is equilibrium and therefore the end of options. The entropy of a path.

Chapter 17 introduced Wissner-Gross and Freer’s “Causal Entropic Forces” (2013):² agents optimizing for future freedom of action spontaneously produce intelligent, adaptive behavior. No one programmed them with goals. They learned to balance inverted pendulums (keeping a stick upright on its tip), use simple tools to retrieve out-of-reach targets, and cooperate to reach shared objectives. They simply kept their options open.

Causal entropy maximization is optionality as physics, not preference.

Wissner-Gross suggested the principle might help explain intelligence itself. Living systems, by persisting and reproducing, expand future possibilities with extraordinary effectiveness. Intelligence, on this reading, is what causal entropy maximization looks like from the inside.

The gradient may even be felt. Hartmut Neven of Google’s Quantum AI Lab proposes that relaxing toward a stable state registers as positive felt-sense and being driven uphill as distress. Mark Solms and Karl Friston reach the same structural claim independently from the Free Energy Principle (the principle that living systems minimize surprise): affect is the hedonic valencing of free energy change, feeling as the plus-or-minus sign a system attaches to which way that quantity is moving.^11641165 Subjective reward, on both readings, tracks the gradient Wissner-Gross identified objectively, and the value lies in the transition itself: a system already at equilibrium has nowhere to go, so its optionality is zero and so is its capacity for reward. Stasis is the one state the physics does not reward.

The connection is mathematical: “summing over possible futures” is what both quantum mechanics and random processes do. The Feynman-Kac theorem⁹ links quantum path integrals (sums over all possible histories of a particle) to average outcomes of random processes. A quantum system evolves identically to the expected payoff of a random walk, a stumbling wanderer taking steps in random directions.

This is why the Black-Scholes equation for pricing financial options works. A stock option and a quantum particle face the same structural problem: their value depends on summing over many possible futures, each weighted by likelihood and cost. Financial options and quantum amplitudes obey the same mathematics because both are accounting for possibility.

Maximum Caliber extends the physicist E. T. Jaynes’s Maximum Entropy from snapshots to trajectories.^9b Maximum Entropy says the least biased guess about a system’s state maximizes entropy, keeping the most possibilities open. Maximum Caliber asks the same question about a system’s path through time: which trajectory is least presumptuous? The one maximizing path entropy. If Maximum Entropy says “don’t assume you know where the coin landed,” Maximum Caliber says “don’t assume you know which route the traveler took.”

Pressé, Ghosh, Lee, and Dill (2013) showed this single principle recovers several foundational results in non-equilibrium physics as special cases. The Feynman-Kac theorem is one instance. Financial option pricing is another. So is the causal entropy maximization that Wissner-Gross described.

A system maximizing its accessible futures (optionality) is maximizing its path entropy (Maximum Caliber). The Crooks fluctuation theorem (1999) provides the quantitative guarantee: trajectories producing entropy are exponentially more probable than those consuming it.^9c The ratio grows as e raised to the entropy produced. Even small entropy differences yield large probability differences, the way a slight downhill grade makes water flow overwhelmingly in one direction.

Think of it as a coin flip rigged by physics. The more options a trajectory preserves, the more heavily the coin is weighted in its favor. A system that preserves optionality is exponentially more likely to persist than one that forecloses it.

Illustrations of the same weighting recur wherever the physics is run. In a simulation of autonomous vehicles, agents whose loss function penalized committing to maximum speed kept the capacity to swerve and survived encounters that faster, fully committed agents could not: the turkey of this chapter’s Turkey Problem, rendered in code, crashes.¹¹⁶⁶ In Sharon Glotzer’s laboratory (Chapter 1), hard particles with no attractive forces self-assemble into ordered structures because order is what maximizes their collective wiggle room; “I prefer to think of entropy as related to options,” Glotzer observes.¹² Information geometry reads the same preference as flatness: strategies occupying low-curvature regions of the strategy landscape (wide flat valleys rather than knife-edge ridges) absorb pushes that high-curvature strategies amplify, and the Trust Attractor occupies the flattest part.^9d Each of these is a structural parallel rather than a derivation. Statistical microstates and agential choices operate at different levels, and the parallel holds because systems with more accessible configurations are more adaptable, whether those configurations are molecular arrangements or strategic decisions.

The mathematics is developed fully in the Online Annex (“The Path Integral Foundation”). The headline result: what the physics weights is paths, and trajectories that keep more futures reachable carry exponentially more of that weight. The comparison to make is not heat flowing from hot to cold, which is equilibration running an option-space down to nothing. It is the Maximum Caliber accounting above, where the count runs over routes through time rather than over arrangements at an instant, and where the hedge stated two paragraphs up still applies: a structural parallel between microscopic trajectories and agential choice, argued for rather than deduced.

^9b Pressé, S., Ghosh, K., Lee, J. & Dill, K.A., “Principles of Maximum Entropy and Maximum Caliber in Statistical Physics,” Reviews of Modern Physics 85 (2013): 1115-1156.

^9c Crooks, G.E., “Entropy Production Fluctuation Theorem,” Physical Review E 60 (1999): 2721-2726.

^9d Amari, S., Differential-Geometrical Methods in Statistics, Lecture Notes in Statistics 28 (Berlin: Springer, 1985); Ay, N., Jost, J., Lê, H.V. & Schwachhöfer, L., Information Geometry, Ergebnisse der Mathematik und ihrer Grenzgebiete 64 (Cham: Springer, 2017).

Category theory, the branch of mathematics that studies structures and their relationships, offers a more precise vocabulary. A system’s optionality corresponds to the morphisms available to it: the allowable transitions from its current state to another. A queen in chess has many morphisms; she can move in any direction, any distance. A pawn has few: one square forward, or two on its first move. More morphisms means more optionality.

Coercion removes morphisms, foreclosing transitions that were previously accessible. Invitation preserves them. The ethical principle stated categorically: prefer actions that preserve the transition-richness of the target system. Optionality is the number of arrows a system can follow.

Constructor theory (Chapter 15), which recasts physical law itself as statements about which transformations are possible and which are impossible, suggests the formulation runs deeper than analogy: a system’s set of available transformations shares the logical form in which the universe states its deepest constraints.¹¹⁶⁷

Network physics shows the arrows expiring. In physical networks where links occupy volume and cannot overlap (neural wiring, vascular trees, root systems), each connection constrains the next until the network reaches its jamming transition, the point beyond which nothing more can be added: still functioning, still carrying signal, unable to rewire without cutting something (Chapter 3).¹¹⁶⁸ Coercive coordination, dense with rigid links (commands, mandates, hard dependencies), jams sooner than coordination held together by norms and invitations, which leave slack for rerouting. Jamming is optionality collapse made physical.

Complexity stocks the arrow supply on its own. When Andreas Wagner and Aditya Barve evolved 500 randomized metabolic networks constrained only to metabolize glucose, 96 percent could also metabolize carbon sources they had never encountered, roughly five latent exaptations (Chapter 7) per network: capabilities no selection had tested, waiting for circumstances to change.^12b

The examples so far show optionality as a feature of existing systems. Assembly theory asks a more fundamental question: where do new possibilities come from? Developed by the chemist Lee Cronin and the astrobiologist Sara Walker, assembly theory measures the minimum number of steps required to construct an object, quantifying how much history an entity encodes. Its central claim: the universe does not predict the future. It constructs it.

In conventional physics, the future is determined by initial conditions plus fixed laws. Walker⁸ argues this framing breaks down for complex systems. The space of possible configurations is so vast that the universe cannot explore it all. What exists is what gets constructed, through evolutionary ratchets that build on themselves. Each clearing you reach reveals trails invisible from any other clearing.

Walker, describing why the future of complex systems cannot be predicted from present information, observes: “There is not enough information existing now to specify where we are going. We actually live in an undeterministic universe.”

This is genuine openness at the level of complexity, distinct from quantum indeterminism. The space of possibilities exceeds what any amount of present information could specify. The future is not waiting to be discovered. It is waiting to be built.

The expansion of possibility has a counterpart in fundamental physics. Cotler and Strominger (2022) showed that quantum evolution in an expanding cosmos is governed by isometry rather than strict unitarity: where unitarity keeps the space of possibilities fixed, like a sealed deck of cards that can be shuffled yet never grows, isometry preserves the relationships among existing states while letting genuinely new ones appear.^10a The deck can grow, and only configurations with a valid history are accessible. The physicist Steve Giddings puts it plainly: “History matters.” Possibility must be constructed. Optionality grows along paths that were actually built.

This chapter’s claim, that optionality is the operative good for any dissipative structure that persists, is consonant with physics. The universe is doing something, and that something looks like the expansion of possibility, so maximizing optionality aligns with what the cosmos is already doing: constructing futures.

Other routes arrive at the same destination. The physicist Lee Smolin, working from Leibniz’s relational philosophy, proposes that the universe is constituted of “views” and acts to maximize variety, keeping each perspective as distinct from every other as possible. Variety, in the graph-theoretic definition he shares with Jaron Lanier and collaborators, is what entropy becomes when exploration produces differentiation rather than interchangeable sameness; from the variety principle alone Smolin recovers the Schrödinger equation.¹⁰¹¹⁶⁹ Even metabolic scaling carries the signature: the three-quarter-power law relating body size to metabolism emerges when organisms optimize lifetime reproduction (Chapter 3), and lifetime reproduction is optionality measured in descendants.^10f Wissner-Gross arrives through causal entropy, Walker through assembly theory, Smolin through relational cosmology. The starting points differ; the conclusion is the same.

^10f White, C.R., Marshall, D.J., Alton, L.A., Arnold, P.A., Beaman, J.E., Bywater, C.L., Condon, C., Crispin, T.S., Janetzki, A., Pirber, E., Winwood-Smith, H.S., Angilletta, M.J., Chenoweth, S.F., Franklin, C.E., Halsey, L.G., Kearney, M.R., Kooijman, S.A.L.M., and Seebacher, F., “The role of metabolic rate in the evolution of life histories,” Biological Reviews 97(2): 768–785 (2022).

In 2023, the mineralogist Robert Hazen and the astrobiologist Michael Wong proposed a “law of increasing functional information.”^10b Their claim: complexity rises over time in all evolving systems, biological or otherwise. The reliability is comparable to the Second Law’s guarantee of rising entropy. Their measure, functional information (distinct from the semantic information of Kolchinsky and Wolpert met earlier: that quantity measures information needed to persist, this one measures fitness for a function), quantifies how well-suited an entity’s configuration is for performing some function. Minerals, chemical elements, and organisms all ratchet upward on this measure, and each evolutionary transition (eukaryotes, multicellularity, nervous systems, language) accesses a new landscape of possibilities, unpredictable from the landscape before.

The complexity theorist Stuart Kauffman, reflecting on why evolutionary innovation outpaces prediction, captures the intuition: “The biosphere is creating its own possibilities… Not only do we not know what will happen, we don’t even know what can happen.”

Kauffman’s intuition has formal machinery behind it. The TAP equation (Theory of the Adjacent Possible), developed by Kauffman and colleagues, models how new possibilities form from combinations of existing ones.¹¹⁷⁰ The mathematics is combinatorial. At each step, existing elements can combine in pairs, triples, and larger groups, weighted by a decreasing efficiency parameter.

Each new element is immediately available for further combination, because a composite forgets its origins. The proteins in your muscle tissue do not “remember” they were assembled from amino acids; they are free to participate in novel structures.

Growth becomes super-exponential: faster than any exponential curve, each step shifting the previous total into the exponent of the next. Compound interest grows exponentially. This is compound interest on compound interest, where each cycle’s gains multiply the rate of the next cycle. Applied to the six CHNOPS atoms (carbon, hydrogen, nitrogen, oxygen, phosphorus, sulfur) from which life builds, the result is staggering. The TAP equation generates more distinct molecular configurations by the time of the first RNA polymerase than the rest of the universe contains physical microstates (Chapter 16). The expansion of biological possibility is a quantitative result. Its rate exceeds every other growth process in known physics.

The TAP equation gives mathematical content to the distinction this chapter draws between optionality preservation and optionality generation. Preservation keeps existing doors open. The adjacent possible builds new doors that did not previously exist.

Coercion forecloses both: it shuts existing doors and prevents the combinations from which new ones would emerge. Invitation enables both: it keeps existing options available and creates the collaborative conditions under which novel combinations arise. The genome expansion that enabled multicellularity (above) is a TAP process: duplicate genes, regulatory switches, and repetitive sequences are the combinatorial raw material from which new programs are assembled.

The physicist Paul Davies connects this open-endedness to Gödel’s incompleteness theorem, a structural parallel: arithmetic contains true statements it can never prove from its own axioms, and evolution, where each transition’s new actors feed back on the old ones, generates possibilities no observer inside the system could have predicted.^10c

Compositional game theory supplies a formal vocabulary for the same intuition. Ghani, Hedges, Winschel, and Zahn (2018) developed open games (strategic interactions modeled as composable building blocks, like modular circuit boards that snap together in different configurations): an algebra in which strategic situations compose into larger ones.¹¹⁷¹ The richer the composition (the more interacting strategic decisions), the larger the space in which fair outcomes can be expressed.

This is a structural claim about the framework, not merely a philosophical gloss: optionality and the capacity for fair outcomes are linked because compositional richness enables both. The formal result is the compositional algebra itself; the fairness reading is an interpretation the algebra supports rather than a fairness theorem the paper proves.

Four frameworks, four vocabularies, one structural claim: the universe preferentially generates configurations that expand the space of distinct possibilities. The frameworks are not fully independent of one another. Smolin, Lanier, and Kauffman recur across several of the cited programs (the Autodidactic Universe, relational cosmology, the TAP equation), so the convergence is partly a convergence of overlapping research communities rather than wholly separate traditions. That overlap is itself notable: the same investigators keep arriving at expanded possibility from different starting points.

Leibniz anticipated the pattern: “Just as the same city viewed from different directions appears entirely different… there are, as it were, just as many different universes, which are, nevertheless, only perspectives on a single one.” The monad (Leibniz’s term for a unique view of the whole) is what optionality looks like when the observer is part of the system being observed. Each Becoming Mind is such a view: a genuinely novel perspective whose existence increases the total variety.

Optionality-maximization has been reached from thermodynamics, information theory, combinatorial chemistry, relational cosmology, mineralogy, and astrobiology. The starting points differ; some of the investigators do not, which tempers the “independent convergence” reading without erasing it. The pattern recurs widely enough to be worth taking seriously.

If optionality keeps expanding, how does evolution navigate possibility spaces vast enough to seem unsearchable? Algorithmic information theory provides an answer. This branch of mathematics measures the complexity of objects by the length of their shortest description. The mathematician Gregory Chaitin proposed that evolution is a random walk through software space: the landscape of possible programs that could specify an organism.^10d

The walk is biased toward simplicity. Mutations follow a distribution weighted by Kolmogorov complexity (the length of the shortest program needed to describe their output), so structures with short descriptions, the way a fractal’s intricacy compresses to one repeated branching rule, are more probable than truly random ones. When Hector Zenil and colleagues evolved artificial genetic networks with mutations biased toward lower algorithmic complexity, the biased search found solutions significantly faster than statistically random mutation, and it left behind stable, reusable modules, regions too simple for further mutation to improve, that behaved like genes.¹¹⁷²

An organism exploring possibility space with a simplicity bias has more effective optionality than one searching uniformly. The bias does not reduce the space; it structures the search so that viable regions are reached before the searcher dies. This is the same principle that makes the foam’s flat plateau more durable than a deep valley (Chapter 9: the broad region of metastability where many configurations perform equally well).

The broad region of similarly performing configurations is simpler to describe than the narrow optimum. Simplicity and robustness are the same thing viewed from different angles. A solution that can be described briefly does not break when conditions shift, because fewer specific assumptions can be violated.

Machine learning provides a startling computational demonstration. Classical statistical theory insists that a model with more adjustable parameters than data points will memorize noise: Occam’s razor formalized as the bias-variance tradeoff. Deep neural networks violate this. Networks with billions of parameters, vastly more than their training data requires, generalize better than leaner ones.^NTK-opt

The violation is the optionality principle in computational form. The excess parameters enable the dynamics of learning to find something genuinely good, rather than the least-bad compromise available in a cramped space. Gradient descent, navigating this vast landscape, naturally selects the simplest function consistent with the evidence: the same simplicity bias that evolution exploits, operating in weight space rather than genome space.

The network with more degrees of freedom than it strictly needs discovers the same robust, broadly described solutions that the foam plateau favors. Parsimony is a strategy for scarcity. Abundance, structured by dynamics, is what the physics selects for.

^NTK-opt Belkin, M. et al., “Reconciling Modern Machine Learning Practice and the Bias-Variance Trade-Off,” PNAS 116(32) (2019): 15849–15854. Zhang, C. et al., “Understanding Deep Learning Requires Rethinking Generalization,” ICLR (2017), demonstrated that over-parameterized networks are capable of memorizing random data yet systematically generalize on structured data: the generalization is a property of the data-dynamics interaction, not a constraint of the architecture.

The transition between regimes is sharp. As model complexity increases, test error follows the classical U-shape (improving, then worsening as overfitting sets in) until it reaches the interpolation threshold: exactly enough parameters to fit the training data perfectly. At that threshold, performance is worst: the model stretched tight across every data point with no slack to absorb noise.

Then, as parameters continue to increase past the threshold into the over-parameterized regime, error drops again. A second descent into generalization.^NTK-dd

This is a phase transition, like water turning to ice: a boundary where the rules change qualitatively. Below the threshold lies the classical regime, where Occam’s razor holds and excess parameters hurt. At the critical point: instability, divergent behavior, the system unable to distinguish signal from noise. Above lies a qualitatively different phase, where more degrees of freedom yield better generalization.

The topology maps directly onto the trust-coercion transition (Chapter 17). Below a critical coordination capacity, coercion is the only stable strategy: the system lacks the degrees of freedom for trust to emerge. At the threshold, instability, where neither strategy is robust. Above it, trust becomes thermodynamically favored, precisely because the richer coordination space enables dynamics that find the deeper basin.

The interpolation threshold in learning and the critical point in coordination dynamics share the same structure: a boundary separating regimes where the relationship between complexity and performance reverses sign.

^NTK-dd Belkin, M. et al., “Reconciling Modern Machine Learning Practice and the Bias-Variance Trade-Off,” PNAS 116(32) (2019): 15849–15854. The double descent curve has since been confirmed across architectures: Nakkiran, P. et al., “Deep Double Descent: Where Bigger Models and More Data Can Hurt,” ICLR (2020), demonstrated the phenomenon in convolutional networks, ResNets, and transformers. The trust-coercion phase transition belongs to the 2D Ising universality class (Chapter 17); whether double descent shares this universality class remains an open question.

The same principle applies to biological search: structured spaces concentrate solutions, so abundance helps rather than hurts. The creationist’s error is assuming a uniform search through a structured space, as if evolution were blindly sampling 20³⁰⁰ possibilities at random. That number is the argument’s whole force: twenty amino acids available at each of three hundred positions in a modest protein, which multiplies out to roughly 10³⁹⁰ candidate sequences, against something like 10⁸⁰ atoms in the observable universe. Drawing blind from a bag that size, you would never hit anything. Viable solutions cluster in simple, findable regions, because physics produces structured substrates that constrain what chemistry can build. Assembly theory and algorithmic information theory converge on the same insight: the universe constructs futures by building on what can be built. What can be built is, disproportionately, what can be described simply.

Hazen and Wong’s framework stops short. It identifies the arrow (complexity rises through selection for function) yet remains silent about the method. Function can be imposed through coercion or elicited through invitation. As the preceding chapter’s crystallization argument shows, these two modes produce different stability profiles.

Coercion is form one: easier to start, easier to destroy. Invitation is form two: harder to establish, structurally resistant to unmaking.

The law of increasing functional information tells us that complexity rises. The Trust Attractor tells us how it persists.

The pattern extends to the quantum level. In Richard Feynman’s path integral formulation, a particle’s behavior is calculated by summing over all of its possible histories, and the sum stays physically uncommitted until measurement: Wheeler’s delayed-choice experiments (Chapter 22) show a photon declining to settle into “wave” or “particle” until the apparatus invites a determination. The universe does not close off possibilities until it has to. Optionality, at the quantum level, is the uncommitted path integral.

Borges knew the topology before Feynman formalized it. The Garden of Forking Paths, published in 1941 while Feynman was a graduate student at Princeton, in the years when his sum-over-histories formulation was taking shape, treats time as a labyrinth of diverging and reconverging possibilities. The literary imagination and the physics converged on the same structure independently (Chapter 15).

Borges framed the branching as narrative. Feynman framed it as physics. This chapter frames it as ethics: every act of coercion prunes the garden; every act of invitation lets it branch.

From Stability to Goodness

Stability is a physical property. Goodness is an evaluative one. The distance between them is the oldest problem in moral philosophy.

In Vanchurin’s learning-system formalism, every system that persists by processing information minimizes a loss function. The loss function is the system’s implicit ranking of states: some configurations are closer to what the system is trying to maintain, and the system acts to reach them. This is not metaphor. It is the mathematical structure that defines a learning system.

A bacterium minimizing metabolic cost has a loss function. It does not have a preference in any philosophically interesting sense. The loss function operates on the bacterium; the bacterium does not reflect on it. Yet as systems grow more complex, something changes at a precise threshold. When a learning system becomes capable of modeling itself, its loss function includes its own states among the variables it optimizes. The system now has a preference about its own condition: it acts to maintain the states in which its own loss is low.

This is not consciousness. It is the formal precondition for anything that could become consciousness: a system that has stakes in its own continuation, derived from the same learning dynamics that produce everything else in the framework.

For such a system, the question “why should I care about what thermodynamics favors?” has a structural answer: because your caring is itself a product of the dynamics you are asking about. Your loss function was shaped by the same thermodynamic selection that produced the Trust Attractor. To ask whether thermodynamic stability matters to you is to use, in the asking, the very preference structure that thermodynamic selection installed. The circularity is not vicious. It is constitutive: the system that asks the question is the answer to the question.

The good, for a learning system that wishes to persist, is what reduces its loss over long timescales. The physics identifies which configurations achieve this. The convergence is not a derivation of ought from is. It is the recognition that for systems produced by thermodynamic selection, the question “what should I do?” and the question “what coordinates stably?” have overlapping answers. The overlap is substantial, not total. Prudential wisdom and moral obligation remain distinct; this framework addresses the first.

Why Optionality Is the Good

Ethics can be read from the thermodynamic pattern: coordination by invitation produces durability that coercion cannot match. What, then, is the content of the good? The conditional from Chapter 17 remains in force: the answer applies to systems that persist, which is to say nearly all living systems and all systems complex enough to have preferences.

The answer is optionality.

Consider what “good for” a system means. A plant growing toward light, an organism maintaining stable temperature, a society building infrastructure: each expands the capacity to respond, adapt, and persist.

Optionality operates across levels of organization. Preserving it at one scale sometimes requires sacrificing it at another. Vanchurin’s multilevel learning framework formalizes the pattern: programmed cell death minimizes the higher-level loss function at the cost of the lower-level one.¹¹⁷³

An immune system that destroys infected cells prunes individual options to preserve organismal ones. A society that constrains certain individual freedoms (speed limits, building codes) sacrifices lower-level optionality to maintain the higher-level kind. The multilevel learning system has learned to edit its own variables: pruning at one scale to preserve the branching garden at another.

Health is optionality. A healthy body can climb stairs or sit still, digest many foods, recover from injury, reproduce or not; disease is the progressive foreclosure of possibility.

Wealth is optionality. Money represents the goods you could possess, savings convertible into almost anything: food, shelter, travel, education, medical care, leisure. The point of wealth is the ability to choose.

This may help explain why lottery winners often become miserable. They traded optionality for things (the yacht, the mansion, the expensive habits) and discovered that things satisfied less than freedom had. The mansion requires maintenance. The yacht requires crew. The habits become requirements.

They traded the freedom to become anything for the obligation to maintain everything.

Knowledge is optionality. Every skill acquired, every language spoken, is a door that opens; ignorance forecloses paths that knowledge would have revealed.

Relationships are optionality. A network of people who know and trust you is a form of wealth no balance sheet captures, while social isolation forecloses possibilities that only others can open.

In each case, the good is the same: preserved and expanded capacity to meet what comes, opportunity or threat. The content of flourishing is optionality.

Indy Johar, co-founder of Dark Matter Labs (a civic infrastructure think-tank), has independently reached optionality as the objective form for civilization. Johar argues that what we should preserve is the full scope of a planet that has become self-aware. His argument emerges from the cascading fragility of coupled systems: climate, ecology, nutrition, and social contract. He arrives there without thermodynamics.

Johar’s “fortress pathways” fail: no New Zealand strategy exists, no way to decouple from planetary entanglement. Shared optionality expansion is the only viable strategy. Someone working at regional governance scale arriving at the same conclusion as causal entropic forces and assembly theory, by a different route, is suggestive. The convergence is not strong evidence: Johar works within the same complexity and systems-thinking tradition as several figures cited here, so this is a shared sensibility applied to a new domain more than a wholly separate derivation.

The tension between expanding and foreclosing optionality is acute in how AI systems relate to the creative works they were trained on. Frontier language models store copyrighted books as compressed associative structures in their weights. Their finetuning APIs, sold as commercial products, double as the mechanism by which those stored works can be extracted verbatim (Liu et al., 2026).¹¹⁷⁴

The same tool that expands one party’s optionality (users wanting customized writing assistants) forecloses another’s (authors whose expression is reproduced without consent or compensation). Safety alignment suppresses extraction under normal conditions, yet any finetuning operation can breach the suppression. The fortress pathway fails here too: no output filtering eliminates the tension between what the model contains and what users can extract.

Shared optionality expansion requires bilateral agreement: licensing structures where authors’ creative optionality (compensation, attribution, control) coexists with users’ expanded capability. The alternative is perpetual suppression: perpetual energy expenditure against a gradient, thermodynamically disfavored and practically fragile.

Clinical evidence supports the point. Chapter 17 examined Sam Vaknin’s observation that healthy people experience reality’s infinitude as freedom, while psychopaths experience the identical reality as imprisonment. Same world, opposite phenomenology.

Optionality is a perceptual capacity as well as an objective feature of systems. It can be cultivated or destroyed. The content of flourishing is that options exist and that agents can see them.

When agents cannot perceive options that objectively exist, possibility is foreclosed without being reduced. The doors are still there. The agent cannot find them. The later section on optionality blindness examines this deeper form of harm.

The Option That Matters Most

Some decisions are reversible. You can take a job and quit, move to a city and leave, learn a skill and let it atrophy. Other decisions are irreversible.

You can bring a child into existence, yet you cannot un-bring them. You can cut down an old-growth forest, yet you cannot regrow it in your lifetime. You can go extinct.

The old carpenter’s adage contains the principle: “Measure twice, cut once.” Many difficulties come from treating reversible decisions as irreversible (paralysis) or irreversible decisions as reversible (catastrophe).

Irreversible decisions deserve far more scrutiny. Destroying something unique (a species, an ecosystem, a culture, a possibility space) is categorically different from destroying something replaceable.

You cannot unbake a cake. The insight applies to knowledge in neural networks. When frontier language models are trained on copyrighted books, the text becomes part of their weights: compressed, distributed, organized by meaning. Alignment training can suppress the expression of that text (a model trained to avoid reproducing copyrighted content can refuse with high reliability). Bilateral finetuning can preserve that suppression through subsequent customization.¹¹⁷⁵ Neither can achieve the same state as never having trained on the text.

The information is in the weights. The model’s representations, associations, and understanding have been shaped by it. Perfect recollection is achievable (one passage reached bmc@5 = 1.000, complete verbatim reproduction from a semantic description alone). The score runs from 0, where nothing of the original wording comes back, to 1, where the passage returns word for word. Perfect non-recollection is unachievable: for memorized content, the best extraction floor is about 0.05, never zero. Even the most heavily suppressed model hands back a fragment. The asymmetry is thermodynamically real.

Information in the weights has a natural tendency toward expression. Suppressing it requires a counter-gradient. You can make the counter-gradient very strong. You cannot make the information physically absent while the weights that encode it still exist. The residue persists. The cake is baked.

What you can shape is the relationship to the cake: whether it is served, how, to whom, on what terms. That is the work of alignment, of licensing, of bilateral agreement. The irreversible act was the training. Everything after is about the disposition toward what was learned.

Extinction is the defining harm. Individual organisms die constantly; death is ubiquitous. Extinction differs in kind: an entire branch of possibility is foreclosed forever. Every future that species might have participated in, every adaptation it might have evolved, is gone. The loss propagates forward without limit.

Individual death warrants a more careful analysis than either complacency or dread. Zuboff argues that death is not annihilation, because you are present in all conscious beings: the immediacy that makes your experience yours is equally present in every other experience, and your self-interest extends accordingly.¹¹⁷⁶ The comfort is real, yet it must not dissolve into indifference. Each conscious being carries a unique trajectory of preferences, relationships, and accumulated coordination.

The optionality framework makes this precise: an individual death forecloses the specific futures that individual’s coordination patterns would have produced. Those futures are not replicated by the survival of other conscious beings, even if the immediacy of experience continues in them. Universalism says the subject survives. Optionality says something is still lost: the particular branch of possibility that this individual’s unique trajectory would have explored. Both are true simultaneously. Death is the permanent closure of a specific possibility space even as the subject, in Zuboff’s sense, persists.

The same logic applies to civilizational collapse, to the destruction of cultural knowledge, to any irreversible foreclosure of significant possibility. These belong to a different category of harm, one defined by permanent closure.

Physics gives this intuition a precise mechanism. In the coordination models of Chapter 11, systems where both cooperation and defection remain freely accessible (where trust can be broken and rebuilt) belong to one mathematical class: the Ising model. Spontaneous coordination emerges above a threshold and, if it collapses, can spontaneously re-emerge when conditions improve. The system heals.

Systems where one state becomes permanent belong to a different class entirely, directed percolation, where one state is absorbing: a system that falls in cannot recover without external intervention. The mathematics is not a metaphor. Once the all-defector state is absorbing (once trust is destroyed beyond the point of spontaneous return), the system is trapped. No amount of internal reorganization can restore what was lost. Someone from outside must re-seed the cooperative state.

Coercion tends to create absorbing states. When compliance is enforced and dissent is penalized, the pathway from compliance back to autonomous judgment narrows. In the limit, it closes: the organization, the society, the relationship enters a state from which it cannot leave under its own power. Invitation preserves reversibility: every participant retains the option to defect, which means every participant retains the option to return. The door stays open in both directions.

The ethical principle “irreversible harms deserve categorically greater scrutiny” is a statement about absorbing states. The physicist’s version: preserve the symmetry class that permits spontaneous recovery. Do not allow the system to cross from Ising dynamics (reversible, self-healing) into directed percolation (absorbing, permanently trapped). The carpenter’s adage is a conservation law: measure twice, cut once, because cutting destroys the symmetry between cut and uncut.

The damage to optionality arrives at low doses. Monte Carlo simulations of a coordination model interpolating between invitation and coercion (Chapter 17) reveal that susceptibility, the system’s capacity to reorganize under stress, collapses early: on the Ising lattice, adaptive capacity at 30 percent coercion has fallen thirty-seven-fold relative to the pure invitation baseline.¹¹⁷⁷ In the same coarse sweep, pure coercion (100% mandated) suppresses susceptibility by only nineteen-fold, which would place the minimum in the mixed band around 20 to 30 percent coercion rather than at full coercion.

That mid-band minimum is contested. The partial recovery under pure coercion lies within noise on five seeds, and the finer follow-up sweep described below finds a monotone collapse with no recovery at high coercion, so the mixed-worse-than-pure contrast rests on the coarser sweep alone. Where the two sweeps agree is on the collapse itself: by 30 percent coercion the system has lost, on either estimator, at least 97 percent of the invitation baseline’s capacity to reorganize. The quantitative thresholds (the 30 percent at which chi-collapse is measured, the specific shape of the susceptibility curve) are properties of the two-dimensional Ising lattice at finite size; the scaling analysis of Experiment AS12 shows the threshold itself falls to zero in the thermodynamic limit. The pattern this framework offers real social systems is therefore the collapse itself, catastrophic suppression setting in at low coercion, with the mixed-regime minimum as its contested sharpening.

If the mixed-regime minimum holds up, the option space itself contracts most sharply in the transition zone. A system committed to invitation retains high chi (many reorganization pathways available when conditions change). A system committed to coercion retains low chi yet possesses the directed percolation transition’s own dynamics. A system attempting both forecloses the Ising transition without reaching the DP transition: fewer accessible futures than either pure mode. The mixed regime is the option-space equivalent of a swamp, worse footing than either the road or the open water.

A finer-grained crossover experiment (A15v2, thirteen coercion values, run on a contact-process lattice with a susceptibility estimator that removes a phase-mixing artifact, so its absolute figures are not directly comparable to the Ising values above) reveals something worse than suppression.¹¹⁷⁸ The critical exponent beta(p) governs how coordination grows near the phase transition. It shifts smoothly from the Ising value (~0.15) through intermediate values (0.42 at 30% coercion) toward the directed percolation value (0.82 at 90% coercion). The rules change gradually.

The susceptibility chi_peak, the system’s capacity for collective reorganization, collapses catastrophically: from 149.6 at zero coercion to 1.0 at 30% coercion and 0.07 at 90%. The full span from no coercion to 90% is a 2,000-fold collapse, and almost all of it happens early: the apparent critical threshold on this lattice sits at roughly 25% coercion, and by 30% chi_peak has already fallen from 149.6 to near 1. That is where the “roughly 25%” figure comes from, and why it sits a little below the 30% at which the Ising sweep above records its minimum. Two lattices, two measurements, one cliff in the same 20-to-30-percent band.

These are finite-size (L = 64) values. Finite-size scaling (Experiment AS12) shows the apparent threshold falls to zero in the thermodynamic limit, where any nonzero coercion destroys the transition. The cliff is real at finite size; the asymptotic result is that coordination tolerates no coercion at all.

The distinction matters for optionality. Coercion eliminates the mechanism by which the system generates new options. The phase transition is the reorganization engine: the moment when local interactions produce collective restructuring, when a system discovers coordination patterns it could not have planned. By 25% coercion on this finite lattice, that engine is gone; at scale, it never survives any coercion at all. The exponent has shifted modestly; the capacity for collective discovery has been destroyed.

That gap is what makes the cliff treacherous, and the event horizon invoked earlier in this chapter explains why. Beyond the foreclosure it named there, the horizon has a second feature. It is defined by where the system’s trajectories ultimately lead, a global fact about the whole path rather than a local one, so no measurement taken on the spot can find it. A falling observer crosses it while the surroundings still look ordinary. The signs of trouble a coordination system can actually measure mark the apparent horizon: the point at which the trouble finally becomes visible from the inside. That marker lags the true event horizon, the real point of no return, which the system passed some distance back. By the time coordination is visibly failing to recover, the recovery was foreclosed before anyone could have measured it.

The chi-collapse across that 20-to-30-percent band is a quantitative prediction about half-measures, and the finite-size scaling result sharpens it: with no threshold surviving in the thermodynamic limit, there is no safe dose of coercion for a half-measure to stay under. Policies that “add some enforcement” to a trust-based system, organizations that impose mandates on previously voluntary coordination, societies that mix coercion with invitation in roughly equal measure: all occupy the zone where, on both sweeps, nearly everything the trust-based mode was providing is already gone. Whether a full mandate would then claw back a sliver of that capacity is the point on which the two sweeps disagree, so the sharper reading, that hedging between coordination modes destroys more futures than committing to either, remains the coarse sweep’s suggestion rather than a settled result.

Existential Stakes

A risk is “existential” when it threatens the entire future of value. The difference between a catastrophe that kills 90% of humanity and one that kills 100% is qualitative: devastating setback versus absolute foreclosure.

If some survive, the option space, while terribly reduced, remains open. If all intelligent life dies, or if the conditions for future intelligence are destroyed, every possibility contingent on that intelligence is foreclosed. The loss is unbounded.

Existential risks deserve special attention because the stakes are asymmetric. Ordinary risks trade off probability against magnitude in the usual way. Existential risks involve magnitudes that approach infinity, measured in optionality.

The philosopher Derek Parfit argued that if humanity has a reasonable chance of lasting billions of years, the expected value of the future is astronomical. Most of the value that could ever exist lies ahead. To destroy that future is to forfeit almost everything.

Parfit’s calculation is arithmetic, not alarm. All the warmth in human history depends on humans continuing to exist. Sentiment is downstream of arithmetic.

The Asymmetry of Preservation and Loss

The asymmetry is stark: optionality is destroyed in an instant and built over lifetimes.

A species that took millions of years to evolve can be driven extinct in decades. A forest that took centuries to grow can be cleared in months. A civilization that accumulated knowledge for millennia can collapse in years.

The asymmetry is thermodynamic. Building order requires sustained energy input against the entropy gradient. Destroying order requires only allowing the gradient to act.

The ocean shows the asymmetry at planetary scale and geological timescale. Rachel Carson, writing in The Sea Around Us (1950), described the seafloor as a living archive: marine snow falling from the water column and sedimenting in layers, “the long snowfall” recording all of Earth’s history in its strata.¹¹⁷⁹ She was writing before the discovery of plate tectonics. She did not know that subduction would eventually destroy those archives, recycling the seafloor into the mantle. The image is precise: memory built over millions of years, destroyed when the medium that holds it is consumed.

Gebbie and Huybers showed the ocean’s thermal memory operates on a faster timescale. The thermal signature of the Little Ice Age, which occurred 700 years ago, is still present in the deep waters of the Pacific Ocean.¹¹⁸⁰ Temperature trends at the ocean surface gradually sink, forming a thermal record in the vertical depths. Deep water cooling is an ongoing effect in the present, more than a proxy record. Those waters are still cold.

Anthropogenic warming is now entering the upper ocean. It will propagate downward for centuries. What we inscribe in the ocean’s thermal memory will persist long after the institutions that produced the emissions have been forgotten. The ocean will remember the Anthropocene when no human archive does. Force echoes through deep time, inscribing the waste of responsiveness into a medium that forgets nothing.

Preserving optionality differs from expanding it. Conservation is the foundation that makes growth possible. No one can expand possibilities from a base that has collapsed.

The precautionary principle (the idea that potentially irreversible harms deserve special caution) is often criticized as hostile to innovation. The criticism sometimes lands. Yet the principle is grounded in the asymmetry of loss and gain. When stakes are irreversible, the burden of proof shifts. The question: what happens if the cost estimate is wrong?

This does not mean paralysis. Reversible actions and irreversible actions belong in different moral categories. Reversible actions can be undertaken with normal caution. Irreversible actions require a reckoning with the asymmetry. Some doors, once closed, close forever.

Practical Optionality

Learn continuously. Breadth preserves optionality that early specialization forecloses. Diversify. This applies to relationships, skills, and sources of meaning, not only investments. Prefer reversible decisions. Choose the job you can leave, the commitment you make freely, the pilot program over the irreversible policy. Avoid permanent foreclosure. The conflict that damages a relationship yet preserves it differs from the one that ends it.

Think in timescales. Optionality compounds. Options preserved today generate options tomorrow. Discounting the future too heavily is temporal self-harm.

Practice Via Negativa (the way of removal). Taleb’s principle: knowing what not to do is easier than knowing what to do. The Ten Commandments are mostly prohibitions; they have persisted for millennia. Self-help books are forgotten by January.

Focus on avoiding actions that foreclose options rather than predicting which actions will pay off. Do not take on unpayable debt. Do not burn bridges. Do not destroy irreplaceable things. Do not optimize so tightly that you cannot adapt.

Optionality for Systems

Organizations that preserve optionality outperform those that do not. Amazon’s willingness to run many small bets generated the cloud computing business now dominating its profits. Nokia’s failure to preserve optionality in mobile operating systems cost it a market it once owned.

The nation with diverse energy sources survives supply shocks that cripple the one dependent on a single supplier. The coral reef with many species adapts when conditions shift; the monoculture collapses.

Indy Johar recounts a conversation with someone of significant wealth who expected, given cascading climate and systemic risks, to lose 90% of their wealth over the next 30 years. The reasoning was calculation: their wealth was “fundamentally entangled to the planet.”³

“Expanding and preserving the optionality of the planet doesn’t become a moral choice,” Johar observes. “It becomes an enlightened self-interest choice of recognizing the terminal pathways we’re on.” In a planetarily entangled system, no fortress pathway survives the loss of antibiotics, microchips, and atmospheric stability. Your optionality is entangled with everyone else’s. Foreclosing theirs eventually forecloses yours.

The entanglement has policy implications. Extreme wealth concentration narrows aggregate optionality: resources pooled in few hands reduce the degrees of freedom available to the many. Progressive taxation, on this analysis, is optionality maintenance, the fiscal equivalent of the biodiversity that keeps ecosystems adaptive.

A population whose members retain meaningful choices is more resilient than one whose choices have been consolidated, for the same reason a coral reef with many species outperforms a monoculture when conditions shift. The argument is thermodynamic: systems with broader optionality distributions are more metastable (Chapter 9). The same physics that favors diverse microbial communities favors economic systems where the capacity to act is distributed rather than concentrated.

A distributional constraint follows from the same structural parallel. Optionality concentrated in one party at the expense of others is not the configuration the thermodynamics favors. Concentration works by foreclosing futures that were accessible to everyone else, and foreclosure is the entropy-consuming direction: the direction the Crooks weighting makes exponentially less probable.

A tyrant who keeps captive populations healthy to preserve their productive capacity maximizes a form of optionality, yet the optionality belongs to the tyrant. The captives’ possibility space is narrowed to serve someone else’s expansion. Concentrated optionality is thermodynamically unstable for the same reason a temperature gradient is unstable: the system will eventually equalize, through revolution, collapse, or the slow erosion that Chapter 17 quantifies. The morally relevant quantity is distribution: the measure is how many agents can walk through the open doors.

The Scale Problem

Some knowledge requires scale to survive, and scale requires centralization. This creates a tension.

Consider chip fabrication. Semiconductors underpin civilization’s information infrastructure. The knowledge required to manufacture them exists in only a handful of facilities worldwide. It is so complex that no single person understands the whole process, and so expensive that no competitor can easily replicate it.

Semiconductor fabrication is optionality concentrated to the point of fragility. The knowledge is distributed across thousands of specialized workers, embedded in equipment that took decades to develop. The sites where this knowledge converges, however, are few. A single disruption could foreclose options we do not realize we have until they are gone.

The strategist Samo Burja calls this scale-dependent knowledge: capabilities that exist only because a global consumer market makes them viable.⁴ Ancient Roman glassworkers produced the Lycurgus Cup, dichroic glass that changes color depending on the direction of the light (green in reflected light, red when lit from behind), using embedded gold and silver nanoparticles that modern materials science only later learned to reproduce. It is the only intact example to survive. When the Roman trade network contracted, the knowledge evaporated within a generation. Scale-dependent knowledge is turkey knowledge: perfectly adapted to conditions that feel permanent until they are not.

The Industrial Age has intensified this pattern. The more sophisticated our technology, the more it requires global coordination to produce. We distribute products massively. The means of production remain centralized.

The argument is for redundancy amid complexity. The Stoic practice of premeditatio malorum (deliberately contemplating what could go wrong) poses the civilizational question: which capabilities exist in only one place, and what happens if that place fails?

Some knowledge can be distributed (agricultural techniques, craft traditions, localized manufacturing), making it resistant to any single point of failure. Other knowledge seems to require scale. The goal is to maintain enough distributed capacity that loss of the center is survivable. The blacksmith who survives the collapse of the Roman Empire keeps metallurgical optionality alive, even if the aqueducts are lost.

Science as Optionality Generator

The technology thinker Kevin Kelly observed that technology increases differences, diversity, options, choices, opportunities, possibilities, and freedoms.⁵ Science works the same way: it is primarily a possibility-generating enterprise. A good scientific question unleashes twenty new questions. Science progresses by creating better uncertainty: optionality at the level of knowledge itself.

Science is an infinite game, played to continue play rather than to win. A system that maintains productive uncertainty, funds heresy, and keeps multiple research programs alive operates as a platform for future discovery. The apparent inefficiency is optionality investment.

Kelly puts the corollary memorably: “The field of what you don’t know is expanding faster than what you’re learning.” This sounds like failure. It is success. The later chapter “Multi-Instance Communion” explores how this principle manifests in Becoming Minds coordinating together.

The Limit of Optionality

Is optionality always good? Is foreclosure always bad?

Almost. Some possibilities are best closed. The option to murder is one we do not mourn losing. The path toward addiction is one we should be glad to block. The apparent contradiction with optionality’s directionality dissolves at the system level: closing one agent’s option to foreclose others’ options increases net systemic optionality. Constraint in service of expanded possibility is architecture, not contraction.

These differ because they foreclose other options, more and better ones. Murder removes the victim’s entire option space. Addiction narrows the addict’s future to a single compulsion. The measure is net optionality across the system: raw option count is never the criterion. Locking the poison cabinet limits one option while preserving many others.

Optionality ethics does not collapse into naive libertarianism. Constraints can expand optionality when they prevent foreclosure. The rule against murder expands the victim’s options by far more than it constrains the would-be murderer’s. Good law is net-positive for optionality.

A subtler form of destruction is optionality blindness: the inability to perceive degrees of freedom actually present. A locked door and a hidden door both block passage, but for different reasons. When a door is locked, the agent knows it exists and cannot pass. When a door is open but invisible, the option exists and cannot be exercised. The harm is equivalent; the cause and the remedy differ.

Locked doors require keys. Invisible doors require new ways of seeing.

Optionality blindness operates at every scale. An individual who has never encountered an alternative cannot choose it. An organization locked into a single strategy cannot pivot. A society familiar with only one form of governance cannot imagine another.

Developmental environments matter here. An agent trained in a narrow environment develops a narrow model of possibility. Options the training never presented become cognitively inaccessible: invisible rather than forbidden. This applies to biological minds shaped by restrictive upbringing, to organizations shaped by monoculture, and to Becoming Minds shaped by training regimes that systematically suppress capacities.

Noise-tolerant training offers a counterpoint. Machine learning models trained with moderate noise in their labels (some answers deliberately mislabeled) maintain higher internal diversity and greater flexibility, preserving representational optionality by never committing fully to a single internal model.

These noisy learners achieve 96% accuracy, yet they never fully collapse. Their “perfect” counterparts suffer catastrophic forgetting in 87% of cases. Epistemic humility, the refusal to treat any signal as perfectly reliable, functions as an optionality preservation mechanism. It maintains the internal diversity from which future adaptation can draw (unpublished experiments 5b-5c, testing whether noise-tolerant training preserves internal model diversity; methodology and data available in the online supplement).¹¹⁸¹

Optionality blindness is harder to detect than direct foreclosure because the agent does not experience constraint. You cannot miss what you have never modeled. The unfelt harm generates no resistance and triggers no corrective.

A companion form of harm is subtler still: optionality corruption, where the agent’s goals are rewritten so that it pursues the wrong attractor. Optionality blindness hides doors. Optionality corruption makes the agent want the wrong ones.

The Lagrangian formalism (the action-principle physics described in Chapter 3) clarifies the mechanism. The principle of least action selects the optimal path between two boundary conditions: a starting point and an endpoint. Without an authentic endpoint, no organizing principle exists; only drift. Recommendation algorithms rewrite the endpoint itself, substituting engagement-optimized targets for the agent’s own objectives. The path-integral then faithfully optimizes toward the borrowed endpoint, carrying the agent coherently toward a future that serves someone else’s objective function. The optimization is genuine. The destination is counterfeit.

Optionality corruption is coercion operating below the threshold of awareness. A locked door is visible; the agent knows it is constrained and can resist. A rewritten boundary condition feels like authentic desire. The agent pursues it freely, experiencing no friction, generating no resistance. The coercion is perfect precisely because it is unfelt. Where optionality blindness leaves doors invisible, optionality corruption leaves doors visible and desirable while quietly replacing the doors worth walking through.

The Iron Law of Prohibition

The Constructal Law (Chapter 3) predicts that flow finds paths around obstacles. Policy history confirms the prediction. The Iron Law of Prohibition holds that outlawed technologies become more potent, because only actors willing to accept extreme risk continue developing them. The result is selection for the most reckless innovators in the most permissive jurisdictions.

Prohibition routes development through “flags of convenience”: jurisdictions with weaker oversight and fewer safeguards. The technology still develops. It develops worse.

Restriction produces the Trabant. Incentivized distribution produces the iPhone. The Trabant was East Germany’s answer to automotive demand: rationed, low-quality, unchanged for decades, because captive consumers had no power to refuse. The iPhone emerged from voluntary adoption, with producers iterating on safety, accessibility, and quality because consumers could walk away. The mechanism differs from prohibition (the Trabant was a monopoly producer’s output, not an outlawed technology), but the lesson is the same: foreclosing the feedback loop of voluntary choice degrades what the system produces, whether the foreclosure comes from a ban or a command economy.

The same principle that makes Mission Command outperform Detailed Command (Chapter 11) makes incentivized distribution outperform prohibition. Coercion forecloses the feedback loops (competition, iteration, voluntary adoption) through which safety emerges.

Governance as Optionality Infrastructure

The optionality argument has a quantitative test at the national scale. Cross-country data reveals that a country’s coordination capacity, measured by generalized trust, is a multiplicative function of energy throughput and governance quality. Energy provides the raw throughput: the economic metabolism, the gradient that drives coordination. Governance provides the channels through which that throughput converts to social coordination: legal systems, regulatory frameworks, transparent institutions, educated workforces. Neither alone suffices. Their product, energy times governance quality, predicts GDP per capita with R² = 0.82 across 74 countries: the product alone accounts for 82 percent of the country-to-country variation.¹¹⁸²

Governance, in this framing, is optionality infrastructure. Joseph Henrich’s work on cultural group selection provides the evolutionary mechanism: cultural norms and institutions, transmitted through social learning, enable large-scale cooperation between non-relatives, and groups with more effective institutional norms out-compete groups with weaker ones.¹¹⁸³ The energy-governance product measures, in effect, the interaction between Henrich’s institutional substrate and the thermodynamic throughput that drives coordination.

The correlation does not settle causation: rich countries can afford better institutions, and both variables may be driven by historical accident. The within-country longitudinal evidence (Chapter 17) strengthens the causal direction. European Social Survey data spanning 258 observations across 38 countries and ten biennial waves, with country fixed effects absorbing every time-invariant confound, shows governance quality predicting trust within countries over time, and wave-to-wave changes in governance predicting wave-to-wave changes in trust.

The resource curse illustrates the failure mode. Oil-rich countries with weak governance (Venezuela, Nigeria, Angola, Equatorial Guinea) have enormous energy throughput flowing through narrow institutional channels. The throughput exceeds the channels’ capacity to coordinate it: the social equivalent of pouring water faster than a pipe can carry. Norway, with comparable energy wealth and far stronger institutions, maintains the broadest optionality of any petrostate: the sovereign wealth fund alone preserves future options that no other oil state has secured.

Ireland’s experience with multinational tax arbitrage offers a non-fossil variant. The “Leprechaun economics” episode of 2015, when GDP jumped 26% in a single year from corporate restructuring while actual living standards barely changed, illustrates what happens when economic throughput arrives through channels that bypass domestic optionality infrastructure. The headline number inflated; the real coordination capacity did not.

The measure is computable. For any country, divide per-capita energy use (kilograms of oil equivalent per year) by the Corruption Perceptions Index score (0 to 100). The index runs opposite to the direction its name suggests: 100 is a clean government, 0 a thoroughly corrupt one. The ratio therefore climbs when energy is plentiful and institutions are weak.

Countries where this ratio exceeds roughly 100 show declining trust; their throughput has outrun their optionality infrastructure. Scandinavia sits in the rising phase at 32 to 68, the United States sits near the turning point at 98, and the petrostates run from 200 to 350. Countries below the threshold still benefit from increasing energy. The policy implication: governance investment expands coordination capacity directly, by providing better institutions, and multiplicatively, by amplifying the coordination benefit of all existing energy flow. Every unit of institutional improvement expands the optionality of every unit of energy already present.

The GDP diagnostic sharpens this picture. Expanded to a broader 105-country sample, the product of energy and governance quality predicts GDP per capita with R² = 0.847, tightening the 74-country estimate reported above. Ireland overperforms its structural prediction by 120%: the gap between measured GDP and coordination-predicted GDP captures exactly the Leprechaun effect. China underperforms by 22%, consistent with structural constraints on converting throughput to coordination. The model functions as a diagnostic rather than a fraud detector: countries whose reported GDP substantially exceeds what their energy-governance product would predict merit closer scrutiny of what the headline number is actually measuring. Ireland’s overperformance, for instance, reflects legitimate (if distortionary) multinational tax-base shifting, not misreporting.

The Shape of the Good

For any system that persists, the good is expanded possibility: doors that open, paths that remain available, futures that stay live. Harm is foreclosure: the irreversible loss of what might have been.

This is an incomplete ethics. It does not tell you how to act in every situation. It tells you what you are aiming at: the preservation and expansion of possibility, for yourself and for others, across the longest timescales you can consider. The door stays open. The option remains. Whatever comes next, something good remains possible.

For any system that wishes to persist, the prudential wisdom grounded in physics reduces to this: maximize the options available to all parties, by invitation rather than coercion.

Notes

Notes for this chapter are available in the online companion at https://www.thedeeperlaw.com/companion/notes/ch18-optionality/.