Open any feed and watch yourself for a minute. You scroll past a hundred things and stop on one. Maybe you share it, maybe you just linger. Why that one? Almost never because it was the most accurate thing on the screen. Because it made you feel something, or it said something about your side, or it was simply easy. Now multiply that split-second choice by everyone, all day long. Chapter 5 explained how an idea gets reshaped as it travels; the prior question is which ideas get to travel at all, and it turns out to be the half of the machine with a steering wheel.
Chapter 5 leaned on the telephone game and then admitted the game cheats: it forces every player to pass the message. That rule is the whole of the parlor game, and it is the rule real networks do not have. In a real network each receiver decides whether to pass a thing on, to whom, and in which of its circulating forms. That decision is not a re-encoding; it is a gate with criteria. Selection is the sum of those decisions. The selection claim: what crosses a network is not what transport left intact; it is what selection chose; and because selection chooses by criteria, selection, unlike transport, can be aimed. The second clause is load-bearing. A mechanism that works by criteria is a mechanism with a steering wheel, and most of what follows is about where that wheel is, whose hands are on it, and what it would take to grip it differently.
Three operations, not two
selection note and Chapter 5 both describe a pipeline stage as doing two things: transport re-encodes, selection picks. That picture needs a correction, because selection picks from a set, and ranking is not generating. Selection never invents the options it chooses among, so where does the set come from? Transport takes one message and scrambles it, and run enough hands across enough copies and what you get is not one output but a cloud of variants drifting every way at once. *Information Evolution in Social Networks*, tracing over 460 million Facebook meme instances, found that: variants accumulate by a neutral Yule process, “most meme variants are neutral with respect to the parent’s meme,” edit distance compounding as copy is made from copy. That directionless cloud is the raw material. The blindness is not a quirk of memes: Luria and Delbrück pinned it down for biology in 1943, their fluctuation test showing bacterial mutations arise before the selective pressure that later rewards them. Variation proposes in the dark, and only afterward does anything choose.
But not every variant can even be expressed at the next stage. A 280-character medium cannot carry the long, qualified variant; an image medium cannot carry the one whose whole content is a subordinate clause. Something stands between the cloud and the winner, fixing which variants are available to be chosen, and that something is the medium. So there are three operations, present at every stage. Transport generates: lossy re-encoding scatters the idea into a cloud of variants. The medium bounds: it fixes the option space, the set of variants expressible here at all. Selection picks: it ranks the viable set against its criteria, and one variant travels on. Generate, bound, pick. A rumor gets retold a thousand ways (generate); only the versions short enough to text get sent (bound); the juiciest of those is what people forward (pick).
The medium is frozen selection
The middle operation turns out to be the most important, so take it first. The medium note established that the medium sets a gate’s criteria, that the medium is the selection criteria. Drawing the boundary of the option space, ruling that the long careful variant cannot be expressed and the short alarming one can, is itself a criterial act: it keeps some forms and forbids others. So the medium is not a third kind of thing alongside transport and selection. The medium is selection, performed once, in advance, and frozen into the substrate until it stops looking like a choice.
A 280-character limit was somebody’s decision, made in a meeting, on a date. To everyone who has ever posted inside it, it is simply the shape of the world. That is what “frozen” buys. A runtime gate is contestable; you can watch it rank things and argue about its weights. A medium does not look like a gate at all, it looks like physics, which makes it the most powerful selection surface in the picture and the least visible: a runtime gate has to actively suppress the variant it does not want, and a medium never has to. A variant the option space cannot hold never has to be suppressed, because it never forms. The most effective censorship is the kind where nothing is censored, because the disfavored thought was never sayable.
So the medium and the runtime gate are one continuum of criterial constraint at two speeds. At the slow, frozen end, the medium, set once, bounding what can be said. At the fast, live end, the recommendation algorithm, a gate with adjustable weights re-tuned continuously; *Psychology of Virality* gives the worked case, an audit of the Twitter/X algorithm that “amplified political content that did not align with people’s stated preferences.” Both are selection. One was just set so long ago and so far upstream that we forgot a choice was ever made. And “aim” is a spectrum hiding inside one verb: a criterion is aimable to the degree it sits in some actor’s causal reach, and reach varies wildly. The runtime gate’s weights are re-tunable in an afternoon by whoever owns the server; the medium is far harder to reach and far higher-leverage, redrawn once by the few positioned to do it; the evolved human gate is reachable only slowly, by training across thousands of turns of the want-loop. “You can aim a choice” was never yes-or-no; it says a choice has somewhere to put an aim, and how hard that somewhere is to reach is the whole distance from nudging a slider to capturing a platform.
What the gate rewards
A selection gate is not mysterious. It is a receiver (a person, an editor, an algorithm) deciding whether a variant is worth passing on, by criteria local to them. Stack those decisions across a network and a fitness landscape falls out: what gets passed is whatever keeps winning local decisions. Call the criteria memetic fitness, with the warning that nothing in “fitness” implies the criteria are good, coordinated, or pointed at truth. (A caveat on “meme”: the word carries a contested theory of discrete copyable culture-units, and *Memetics — a critique* attacks that. The critique largely lands, and nothing here needs the strong claim. It needs only that receivers re-transmit some variants more than others by local criteria, which is observable whatever the underlying ontology.)
What are the criteria? The literature is blunt. Psychology of Virality lists what predicts spread: “out-group animosity, incivility, negativity, moral outrage, anger, misinformation, indignant disagreement, and high-arousal information.” Information Evolution found the textual fragments that confer measurable replicative advantage, and they are not subtle: “allusions to competition (‘see how many people’), persistence cues (‘status for at least’), or conditions that are easy to match or identify with (‘if you love your’, ‘paste if you agree’, ‘proud to be a’),” and the self-flattering challenge, “95% of you won’t copy this, but the 5% who [have a positive attribute] will.” Identity, belonging, a small action to perform, a reason to feel like one of the good ones. None of them is truth, and none is even aimed near it. A friend, in conversation, put the floor bluntly: “good memetics reduces down to evoking raw strong emotions,” which Chapter 7 takes up as a stable floor. The point here is narrower: the gate has criteria, the criteria are nameable, and nobody who set them was asked whether you end up with a true picture of the world.
Transport is a cost; selection is a choice
Here is the asymmetry everything downstream rests on. For a long time I wanted to say selection is primary, that transport is a substrate and selection the real dynamic. selection note worked through why that is wrong: they are co-equal and parallel, neither downstream of the other. But co-equal is not identical, and the way they differ is the point. Transport and selection differ in modality: transport is an un-steerable cost, and selection is a criterial choice.
Transport scrambles. The word I had to fight to is un-steerable, not directionless: the most directional process in physics is a cost. Entropy only ever increases, and you still cannot aim it at a target you picked. Transport is un-steerable because it is blind to content, scrambling a variant the same way no matter what it says. You can make it cheaper or dearer, with a share button or a character limit, but you cannot make it favor one idea over another, because favoring means reading what the idea says, and a blind cost does not read. Selection is different in kind: it has a what-it-rewards, and a criterion is the kind of thing that can be set, tuned, captured, and contested. You cannot aim a cost. You can aim a choice. (One honest correction, since the tidy form is false: a choice is a cost too. Maxwell’s demon looked like order for free until Landauer and Bennett showed it pays to erase the bits it forgets; natural selection buys its direction with differential death. The real split is not cost against no-cost but blind cost against criterial cost, and aiming is never free, which is exactly why capture always reaches for the cheap criterion and never the bulky pipe.)
None of this is new math, which is the reassuring part. Population genetics has carried the same split since George Price’s 1970 equation: the change in any trait decomposes into a selection term, the covariance between the trait and fitness, and a transmission term, everything imperfect copying does on the way. The aim lives entirely in the first term; if the trait does not covary with fitness, nothing is aimed. The second term, transmission, is the blind cost in formal clothes, and a biased transmission sits right there too, a cost that scatters unevenly without ever consulting fitness. You cannot aim the transmission; you can aim the covariance. That is why the book centers selection, and the reason is exact. Selection is not central because it comes first or does more work. It is central because it is the mechanism with a steering wheel, and every question the later chapters care about (manipulation, political economy, repair) is a question about that wheel. You cannot manipulate transport; there is nothing to grip. Transport is the terrain; selection is the steering; a book about a vehicle that has gone somewhere bad is going to be mostly about the steering. And the historical claim slots in here instead of competing: transport cost has collapsed, the moment a “share” button replaced manual copy-and-paste. When transport was expensive it was also a brake, and collapsing the cost did not delete a mechanism, it removed the brake and handed selection the whole wheel it always held.
Hammer of the Witches, properly this time
Chapter 5 told this story and deferred the real reading to here. The two books are from the early printing press: The Hammer of the Witches (1486), a short, emotionally loaded manual for finding and prosecuting witches, which ran through edition after edition and burned its image of the witch so deep into European culture you can still feel the outline; and Copernicus’s On the Revolutions of the Heavenly Bodies (1543), which reached a specialist few (Arthur Koestler called it “an all-time worst seller,” though Owen Gingerich later showed the first run sold out and was closely annotated). Two generations apart, but the pure-transport reading, that the Hammer was compressed and Copernicus complex and the compressed form travels, cannot explain the thing that matters: both were printed on the same press. Their transport costs were identical. Run the three operations instead. Transport generated copies of both, cheaply and equally. The medium, a printed book, could carry either. The entire difference is the third operation, selection. Against the criteria of sixteenth-century Europe the two had wildly different fitness: the Hammer offered fear, urgency, a crisp in-group and out-group, an action to perform, and a narrative with the reader as hero; Copernicus offered a redrawn cosmology with no action, no payload, no enemy. The compressed-versus-complex distinction is real but downstream: the Hammer is compressed because it was selected for. Transport did not pick it. A continent of gates did, the same gates over and over, and transport merely carried what they kept choosing.
The manipulation surface is the gate itself
Chapter 5 gave the transport version of the manipulation surface: bigger network, lower complexity ceiling, ideas closer to raw emotional payload, easier to weaponize. Real, but the shallow version. If selection runs by criteria and criteria can be set, the manipulation surface is the criteria themselves. You do not have to craft a perfect viral object and inject it; you tune the gate, and the network does the rest. Information Evolution names the disanalogy with biology precisely: “evolution is a blind process in biology, but in social media, there can be a conscious effort to create mutations that will intentionally spread a meme’s presence.” Memetic selection has an operator’s console. And the frozen-selection point makes it worse: the highest-leverage move is not tuning the live gate, which someone might audit, but shaping the medium so the variant you do not want is never well-formed. That manipulation leaves no fingerprints, because it produces no suppressed post and no demoted link, just a world where the inconvenient idea is hard to say. This is also where manufactured content meets selection: a gate scores a variant on fitness, not provenance, with no input for was this ever measured against reality. A manufactured claim and a genuine finding compressed into something equally punchy arrive looking identical, and the gate promotes them with equal enthusiasm. Who tunes the gates, and whether anything in the criteria points at truth, is the whole game, and it is Chapter 10’s.
”Complexity” was doing two jobs
Chapter 5 defined complexity as precondition count, then kept catching itself using the word for other things and worried the real gating variable might be handle-ability: how many mental models an idea plugs into, how much emotional or identity content it carries, whether it tells the receiver what to do. The two-mechanism structure resolves it: “complexity” was one word for two variables, one per engine. Precondition count is transport-complexity, how much outer message must survive the hop, the property that sets the complexity ceiling. Handle-ability is selection-complexity, how the idea scores at a gate. They come apart in the wild: “inflation eased two-tenths of a point last quarter” is trivial to carry (low transport-complexity) and goes nowhere (low handle-ability), while a high-precondition idea can spread if its compressed form scores well, which is the Hammer dynamic exactly. Two separate filters, and a theory with one word for both could never say so cleanly. (Provisional: I am not sure handle-ability is itself one variable rather than a small bundle. But the split along the transport/selection seam I am confident in.)
The strongest objection: people are not that gullible
The honest pressure comes from Hugo Mercier’s *Not Born Yesterday* and, from another angle, Cailin O’Connor and James Owen Weatherall’s *The Misinformation Age*. Mercier argues humans are not the soft targets the easy-manipulation story needs: we come with epistemic vigilance, evolved machinery that screens claims for plausibility and sources for reliability, and people “do not passively consume information, but interpret, transform, or ignore” it. I think he is right, and I think he sharpens the case rather than breaking it, once you notice that epistemic vigilance is itself a selection gate. The receiver is not a passive endpoint; the receiver is a gate, and vigilance is one of its criteria. So Mercier’s correction is not “there is no selection” but “the receiver’s own gate is more truth-correlated than the cynical story assumes,” which is what Part IV’s hope rests on. He relabels the term open vigilance to mark that the system is paired, open and vigilant at once. But it does not save us, for two reasons. Vigilance is one criterion competing against identity, threat, and belonging, and the fitness data shows those routinely winning; a person with a working immune system still passes the meme that flatters their tribe. And the receiver-gate’s criteria are not fixed, they are trained by the medium over thousands of turns of the want-loop, which can slowly select for receivers who pass things on without checking. Mercier describes the immune system we evolved, not one that cannot be worn down. Chapter 7 works through the full engagement and the justification-market reframe it surfaces.
O’Connor and Weatherall point at the other half: their network models show that what corrupts collective belief is often not lossy transmission but selection on social criteria (conformity, the credibility we extend to people like us) operating even among agents who each reason perfectly. That is not a counter to the argument; it is the argument, arriving from formal modeling. The sharpest single piece of evidence is the paradox of virality: across the spectrum people report they “do not want” high-arousal negative content to go viral, and it goes viral anyway. Read naively that breaks any selection story, but it doesn’t; it is the frozen-selection point caught in the act. What spreads diverges from what receivers want because the gate that decided was not the receiver’s gate. It was the medium’s, tuned by someone whose interest is not the receiver’s. The paradox is not evidence against selection. It is evidence of whose selection is running.
What selection does to truth
Selection picks the variant that travels. What travelling does to that variant’s truth is Chapter 5c’s, and the handoff is one sentence: transport, scrambling without direction, produces variants that are merely off; selection, choosing by criteria, gives the error a direction, and whether a compressed claim ends up preserved, inverted, or emptied into a tribal flag is set by which variant the gate reaches for. One more sentence the seam needs, now that Chapter 2 made the book say which reality is in transit. Everything above is the objective-content case, a gate filtering reports about facts it cannot touch. When the cargo is intersubjective (currencies, nations, creeds) the same gates are doing something else: not filtering descriptions of a fact but weighting votes in what becomes the fact, because there, distribution participates in constitution. The book’s oldest case carries the distinction unmarked: the Hammer’s claims about its victims were objective and false, but the judicial category it fed, the witch, prosecutable and burnable, was intersubjective and lethally real. The gates that kept choosing the Hammer were not just spreading errors about the world. They were voting a new thing into it.
Where I land
Selection is the other engine, and the one that matters most. Transport explains the shape of what moves; selection explains what moves; the two run in parallel at every stage, neither downstream of the other, but they are not symmetric. Transport is an un-steerable cost and selection is a criterial choice, and that single asymmetry is why every later chapter is a selection chapter. The diagnosis, as a selection claim: the breakage the book keeps circling is not that networks select (they always have, they must), it is that the dominant medium’s gates have been tuned, by people whose interest is engagement, to reward identity and threat and arousal over anything correlated with truth, and because the medium is selection frozen into the substrate, most of that tuning no longer even looks like a choice. That sets the target precisely. There is no fix that routes around selection, because a network with no gates is not a network. The fix, if there is one, is gates tuned differently and a medium whose option space does not quietly forbid the complex form: a harder target than “stop the compression,” and a more honest one. That brief propagates: Chapter 7 inherits its floor from here (raw emotion wins because a gate was tuned for engagement, so a different gate has a different floor), Chapter 10 becomes unavoidable (who owns the medium sets every criterion in the network), Chapter 11 gets a sharp question (an LLM is a gate and a medium at once, so what are its criteria and who froze them), and Chapter 12 gets its spec.
Where I’m still uncertain
- Handle-ability might not be one variable. I split “complexity” into precondition count and handle-ability and am confident in the split. I am not confident handle-ability is a single axis rather than a bundle: emotional load, identity fit, and action affordance may not move together, and if they don’t, selection-complexity has to become a short list, not a word.
- Emergent versus engineered tuning. I have said gates “are tuned” and let the passive voice carry it. A gate that drifts toward engagement under an ad-revenue incentive and a gate someone deliberately set to invert a finding are different things, and “tuned” hides the difference. Telling them apart is Chapter 10’s job; until then the villain is a little blurry.
- The receiver-gate deserves its own treatment. I introduced vigilance as “the receiver is a gate too” and used it to absorb Mercier in a paragraph. That is a lot of weight on one move; within-receiver selection, what a single mind attends to, retains, and re-transmits, may need its own section.