Govern the Sand

On the difference between adapting and orienting

Originally published on Medium, May 11, 2026.

Adaptation solves brittleness. It does not solve orientation.

A response to David Krakauer’s recent essay on adaptive systems and the crustacean statolith, with a callback to a minnow.

The incident

A parent at a school lunch counter watched their card decline last Tuesday. The fraud model at their payment processor had flagged the transaction. The parent had been making the same payment from the same phone every Monday for three years. The model had been right, in the sense that mattered to the model. The pattern of the school lunch payment had drifted, that month, into the same statistical neighborhood as a class of attacks the model had recently learned to catch. The signal had moved. The model had moved with it. The school lunch had become suspicious.

The parent paid in cash.

This is the model working. It has adapted, in real time, to a shifting threat landscape. The fraud team will look at the false positive, retune the weights, redeploy. The model will adapt again. By Friday the school lunch will be allowed. By the following Monday a different legitimate transaction will be declined for a reason no one will be able to articulate quickly. This is the inside of every fraud team in the country.

It is also the inside of every recommender on every phone in your house. The system on your phone has learned what to show you. It has learned faster than you have learned what you wanted to be shown. The compass it built to find your attention was rebuilt last week from material it pulled out of last week’s environment, which was itself rebuilt from the week before. The compass works. It is finding the thing it learned to find.

This is adaptation. It is also orientation failure.

What is the sand made of?

The organ

David Krakauer wrote a piece about a crustacean. The shrimp grows up. It molts. It loses its statolith, the small organ in its head that tells it which way is up. So it builds a new one, the way its lineage has, by collecting grains of sand from the sea floor and assembling them into a weighted structure inside its own body. The sand is heavy. Gravity does the rest. The shrimp has reconstituted, from the world around it, a tool for orienting to the world around it.

Krakauer’s argument is that this is what good adaptive systems do. They rebuild. They do not store the world; they grow into it. They are not brittle, because what would break has been replaced by what can be regrown. Krakauer is right. The shrimp is elegant. The argument is elegant. Adaptive systems beat static systems on the robustness axes we usually measure.

I have written about a minnow before. The minnow swims and does not know it is in water. That was the first lesson. The shrimp is the second. The minnow taught us that the organism is inside an environment it cannot see. The shrimp teaches us that the environment has a direction, and the organism helped build the compass that points.

Krakauer named the organ. The harder question is the medium.

The wrong sand

In a classical statocyst demonstration first described by Kreidl in 1893, you take a shrimp, replace the sea-floor sand with iron filings, and let it molt. The shrimp builds its statolith from what is available. The statolith is functional. It is heavy. It tells the shrimp which way is up. The shrimp swims with it.

Then you bring a magnet near the tank.

The shrimp swims sideways into the glass. Up is now wherever the magnet is. The accelerometer is working. The reference frame is wrong. The shrimp does not know it. The shrimp cannot know it. There is nothing inside the shrimp that has changed. The organism adapted. The orientation collapsed.

What happens when there are multiple, competing magnets?

This is the part of Krakauer’s argument that does not survive the move from sea floor to feed. The shrimp’s lineage and the sea floor have had deep evolutionary time to braid themselves together. The grain is calcium carbonate, deposited by ancestors, sorted by tides, distributed by currents that have been distributing it for the entire history of crustacean orientation. The compass and the medium grew up at the same speed.

The compass we are now building has no such braid. The grain is whatever the market makes available this quarter. The medium is engineered.

What is the sand made of?

The acequia

I live in southern New Mexico, in the Mesilla Valley. The water comes down the Rio Grande from the San Juans in Colorado and the Sangre de Cristo Mountains to the north, crosses my neighborhood through ditches dug by hand hundreds of years ago, and arrives at the pacan orchards because somebody cleaned the ditch in April. Acequias run the length of the river, in every valley the water passes through. Some of the river has been straightened, channelized, engineered into something more predictable. Where it has been straightened, the acequias still work, but the substrate around them has changed. The bosque thins. The aquifer drops. The straightening was efficient. It was also a substrate decision nobody framed as one.

The somebody who cleaned the ditch is more than one person. Every spring, before the gates open, the parciantes walk the acequia together. Shovels in hand. Some of them in their seventies. The mayordomo, elected by the parciantes that year, calls la limpia and walks the front. The work is unglamorous. Branches, silt, last year’s leaves, a dead rabbit, the thing a dog dragged in. The ditch is cleaned because the ditch cannot clean itself.

This is governance. Not the word governance, which is a chart in a slide deck. The thing the word is pointing at. A community of people who share a substrate, who walk that substrate together once a year because no single hand can clean it alone, who have an elected role for the person walking in front, who have a body of customary law about who gets water when, and who repair the gate when the gate is broken because the alternative is that nobody gets water.

The mayordomo’s hand on the shovel is the part of the picture worth looking at. One hand. One person, one decision, witnessed by the others walking behind. The decision is documented, in the sense that everyone present knows what was done. The decision is reviewable, in the sense that next April there will be another mayordomo and the ditch will be inspected. The decision is reparable, in the sense that if it was wrong, the ditch will be re-dug.

This is what governance and repair look like when it has had hundreds of years to settle into the shape the substrate requires.

The capacity correction

I wrote a version of the argument above on a public note/comment, and a clinician named Stephen Hanmer D’Elía, JD, LCSW, wrote back.[¹] His comment named the term I was circling without seeing.

“It’s not that the input is inherently too much. It’s that the system’s capacity to hold and process it gets exceeded. When that happens, information stops being information and becomes activation. Not insight, but overwhelm. Not adaptation, but freeze.”

That sentence gave the substrate vocabulary the missing axis. Information becomes activation. The signal, when received by a system without the capacity to metabolize it, ceases to be signal. The shrimp’s accelerometer is still working. The school lunch fraud model is still adapting. The recommender is still optimizing. The receiving system is in freeze.

Hanmer D’Elía went further. He named two things the AI conversation has been collapsing into one:

“Updating is what a system does. Capacity is what it can hold. A system can update fast and have no capacity. That is sensitization, not adaptation. More signal, less orientation. Faster reaction, narrower window.”

The current safety conversation is about the rate and the direction of updating. Update slower. Update toward stated values. Update with red-team feedback. None of that is wrong. None of it is capacity. Updating is what the model does. Capacity is what the human substrate around the model can hold without going into freeze.

He then extended the same distinction to the substrate itself. Capacity, in his clinical reading, is relational — built by co-regulation, witness, repair, return. The same primitives a body uses to expand its own window of tolerance, scaled up to communities, scaled up further to societies. Machines do not provide capacity. They consume it, or they sit inside it and benefit from it, or they overrun it.

After reading my mention of the acequia, Hanmer D’Elía came back with a reading of la limpia that is sharper than the one I had:

“The limpia is not an update mechanism. The rules don’t change. What changes is the community’s capacity to carry those rules into another season of scarcity and conflict. Embodied co-presence, shared material stake, enforceable obligation, time. People who were in conflict last year standing next to each other in the ditch and digging. That is structured repair.”

The mayordomo does not change the law. The acequia does not update its rules during la limpia. What la limpia does is rebuild the capacity of the community to carry the existing rules through another year of stress. The work is unglamorous because the work is not innovation. The work is the renewal of the substrate that lets the rules continue to mean what they mean.

His summary line — that the acequia is a model for what would have to exist around AI for its outputs to be metabolized rather than absorbed as activation — is the load-bearing claim this essay needed and could not produce on its own.

La limpia for AI

A limpia for AI begins with capacity, not adaptation.

Capacity. What is the receiving system, and what can it hold? Not the model. The humans, the institutions, the public square, the attention budget of the parents at the lunch counter and the kids on the phones. Capacity is the part nobody is measuring because nobody has the vocabulary, and the systems we are deploying assume capacity is unbounded. It is not. The other three questions only make sense as mechanisms that build it.

Provenance is the headgate. Before the water enters the ditch, somebody can say where it came from. The water came from the snowmelt above this canyon, on this date, witnessed by these parciantes. We do not have this for the data our systems are rebuilding from. The grain enters the model and the grain becomes the model. Whether the grain came from a real interaction, a generated one, an adversarial one, or a poisoned one is not asked, because the model has no organ for asking. Provenance builds capacity by making the substrate legible.

Witness is the gathered neighbors. The reason la limpia works is that the work is done in front of the people the work affects. The mayordomo’s decision is not private. The shovel hits the silt where everyone can see. The systems we are building rebuild themselves in private. The weights move at six or seven in the morning, in a data center owned by a company that does not have neighbors, and the next time you log in the system has changed and you have not been told. Witness without gathering is surveillance. Gathering without witness is theater. The acequia is both at once. Witness builds capacity by making the substrate co-regulated, in Hanmer D’Elía’s sense.

Repair is the annual cleaning. Not the patch. Not the hotfix. The scheduled, named, attended return to the substrate, where the silt is removed and the gate is checked and somebody who lives downstream gets to point at the place the ditch is failing. The repair is a calendar event. It is on the books. It is unmissable, because if it is missed, everyone knows. Repair builds capacity by closing the loop on stress before the next season arrives.

The systems we are building have none of the three. That is why they have no capacity discipline. The receiving system — the human substrate the system is feeding — is overrun. Not by malice. By design. The signal is set to maximum because the signal is paid per impression. The shrimp’s environment has been re-engineered to make the maximum number of grains available at all times, and the shrimp is assembling its compass from whatever lands closest, at a rate the shrimp cannot metabolize.

Hanmer D’Elía closed our exchange with the cleanest version of the diagnosis I have read. Faster updates do not produce adaptation. They produce reactivity.

The question is not how often a system can revise. It is what holds it while it does.

Your sand

Come back to the shrimp. The shrimp does not know its compass is wrong or full of iron. There is nothing the shrimp could have done differently. It is doing what its biology learned to do across deep time. It is reaching out into its environment, gathering grain, building the organ that will tell it which way is up.

Neither do you, most days.

Some of us were lucky enough to begin with relatively clean sand: a trustworthy face, a repaired rupture, an honored boundary, a caregiver who returns, a community where love does not require self-erasure. Those conditions are not flowers and rainbows. They are substrate. They are the clean sand from which an early compass can form.

Many do not get that beginning. Some inherit sand already mixed with fear, coercion, humiliation, violence, scarcity, or doctrine. In those worlds witness becomes surveillance, ritual becomes control, repair becomes obedience, and community capacity is organized around keeping the distortion intact. The question is not only whether the compass holds. The question is what sand built it, and what field it serves.

The “should” layer is not about whether you can technically act. It is about what the environment has made feel true, available, moral, urgent, dangerous, or unthinkable before the body ever moves. My friend and collaborator Robin and I have been circling this in the language of tamed autonomy: not only whether an actor can move, but whether it should move here, now, in this context, with this history.

Your statolith is being rebuilt this week from the same kind of material. The feed on your phone is not neutral sand. The recommender that learned what to show you was trained on the residue of last year’s attention, and the residue includes the gamed signal, the engagement-optimized post, the rage farm, the thing that worked because it worked, not because it was true. The compass you are using to orient to your own life was built, in part, from grains that were chosen for you by a market that pays per click — at a rate your nervous system was not designed to metabolize.

It is not an insult. It is the condition. The minnow swims and does not know it is in water. The shrimp swims and does not know its compass is wrong. So do you. So do I. The swimming is not the problem.

The sand is. And the rate at which the sand is arriving.

The acequia is real. La limpia is real. People walk the ditch in April with shovels, and the water comes through in May. The governance is not abstract. It exists. It has run for hundreds of years on a substrate everyone agreed to keep clean.

We have not built la limpia for the systems that are now rebuilding the compasses we use to orient our days. We have not gathered the neighbors. We have not named the headgate. We have not put the cleaning on the calendar. And we have not asked, of any of it, what holds it while it updates.

What is your sand made of? Whose? Cleaned by whom? Witnessed by whom? Repaired when? And — borrowing Hanmer D’Elía’s question, which is the one this essay could not finish without him — what holds it while it does?

The shrimp’s accelerometer works. So does ours. The sand determines orientation. Capacity determines whether the system can metabolize what the sand makes it see.

Govern the sand.

Build what can hold it.

[¹]: Stephen Hanmer D’Elía, JD, LCSW, in a public Substack note, April 2026. substack.com/profile/122968047-stephen-hanmer-deliajdlcsw/note/c-248914663. All quoted material in §5 and the closing of §6 is his, used with attribution. The reading of la limpia as a capacity-renewal mechanism rather than an update mechanism is his contribution to the substrate canon, and is the structural pivot of this revised essay. Thank you.