A living tradition then is an historically extended, socially embodied argument, and an argument precisely in part about the goods which constitute that tradition.
Alasdair MacIntyre, After Virtue, 1981
A substrate that weighs cost favors the wealthy at first. Seven Seams takes up that audit on its own ground. The question this chapter is about is what happens once the door is open: who else gets to write.
The previous chapter named what two enemies can do on the tree when they separately arrive at the same commitment. The question it left open, the one I could not answer inside that frame, was the other bodies. The ones that do not have a seat at any existing table. The ones that believe something and have never had a substrate to say so on a record nobody else owns.
The title of this chapter carries a lineage. Mary Douglas argued in Purity and Danger (1966) that the body is the primary symbolic system through which any society draws its boundaries. Merleau-Ponty, twenty years earlier, argued the deeper claim. That the body is how meaning reaches us at all. “A body that believes” is not a metaphor in either tradition. It is where belief becomes legible to anyone outside the believer.
The Paper That Named the Gap
I had been sitting with the question of which bodies are eligible to write when a paper from inside the alignment field named the underlying problem in a different vocabulary. Taylor Sorensen and her co-authors, at ICML 2024, published A Roadmap to Pluralistic Alignment. A piece of honest architectural thinking about what it would take to align an AI system to more than one perspective at once. They formalize three modes. Overton pluralism: present the whole spectrum of reasonable answers. Steerable pluralism: be faithfully steered toward a given position. Distributional pluralism: match the distribution of answers a real population gives to a question. They propose benchmarks for each. Then they do the thing the field needs more of, and they measure what the current method actually does. Across LLaMA, LLaMA2, Gemma, and GPT-3, the models after alignment training were less similar to real human population distributions than the base models were before. The thing the field calls alignment, they show, empirically narrows.
The paper is honest enough to leave the load-bearing question on the table. They state it in the limitations section, in the voice of people who know they cannot solve it from inside the model: “In creation of a general LLM, like ChatGPT, who is the target distribution?”
The field has no answer. Every proposed one folds back into the same shape. An Overton window requires someone to draw it. A steerable set of attributes requires someone to pick which attributes are admissible. A distributional target requires someone to pick the distribution. Three different operationalizations, one unresolved question underneath all three. who curates? The paper acknowledges this openly. It does not propose a way around it, because there is no way around it from inside the model. The way around it is the substrate the values get written into.
I read the paper three times. It was describing, in the vocabulary of the field, the question the tree had been answering in a different vocabulary.
What I Came to See
The Overton window is not a curator’s responsibility. It is the readable surface of a structure that bodies have inscribed into at cost, over time, from incompatible angles. What counts as a “reasonable answer” is what enough independent bodies, states with treasuries, institutions with budgets, communities with time, paid enough to record and have not since outweighed against. The window is not drawn. It is read.
Steerable pluralism is the same shift. The attributes a model can faithfully steer toward are the attributes that carry weight on the tree. “Honor patient confidentiality” is a steerable attribute if a thousand independent medical institutions across four decades anchored it and sustained the anchoring. A contested attribute with a single inscription from one actor is also steerable, and the structural weight of the steering is honest about its provenance. Nobody at the lab decides which attributes are admissible. The structure shows the reader what each attribute cost, how long it has held, and who stood behind it.
Distributional pluralism is the third. The target distribution for a general-purpose model is not a population picked by the team training it. It is the distribution of who paid to be counted, weighted by what they anchored and how long the anchor has held. Bodies that did not anchor are not in the distribution. That absence is graded, the silence of a body that runs a node and chose not to inscribe is different from the silence of a body that has no node, and both kinds of silence are readable by the same mechanism. Neither requires a curator.
The Tree of Proof is the weight mechanism for the Overton window. Whatever I had been doing with the idea of branches and cost and time, I had been building, without the vocabulary for it, an answer to the question Sorensen’s paper was asking.
Who Else Can Write
The previous chapter named states and the MSF-type institution. The substrate does not care if that is the whole list. It asks the same question of every body that shows up with a node and an inscription. Once that is the only question, the list of eligible writers gets longer than any existing political form is comfortable with.
Institutions beyond the NGO. Universities. Hospitals. Scientific consortia. Religious orders. Professional bodies. Standards organizations. Research collaborations that span jurisdictions. Each of these has commitments older than any single member, carried in charters that its own administrators can quietly reinterpret. The tree is the one place a commitment can be anchored such that the next restructuring cannot discreetly edit it. A monastic order that has held the same Rule for a millennium has more time on it than any modern state; it has had no substrate to say so on a record separate from the order’s own archives.
Tribes and pre-institutional groups. The architecture of the modern world rewards entities that can produce a tax ID. The architecture of the tree does not. A group does not need to incorporate to commit at cost. An indigenous nation whose compact predates the state that refuses to recognize it is as eligible to inscribe a commitment as the state that refuses. A diaspora community scattered across six jurisdictions can pool sats, run a node, and anchor what it holds in common without any single government’s permission. The four forces ask whether the energy was burned. They do not ask for credentials.
Coalitions across jurisdictions. A coalition of farmers across three continents can run a shared node and anchor a commitment to a farming practice. A network of investigative journalists across twelve countries can anchor a commitment to source protection. A medical coalition can anchor a commitment to not share patient data with state intelligence services, regardless of which state requests it. None of this requires the coalition to be a legal entity anywhere. It requires it to be able to hold a private key and sustain a node, which is a lower bar than incorporation and one the existing world has no way to gate.
Generations. A generation is a body too, in the relevant sense. A generation holds conviction in common that the next one may not share. Until now, the record of what a generation stood for has been written by whoever administers the archive after the generation is gone. The tree is the one place a cohort, a founding generation of a new field, the surviving members of a civil-rights movement, the last witnesses of a historical event, can anchor what they stood for in a form the record-keepers who come after cannot quietly edit.
The missing middle of the alignment conversation has been this list. Not individuals. Not corporations on their own. The bodies in between. Every collective with shared conviction and any sustainable economic activity. The bodies that believe something and need a way to say so on a record nobody else owns.
A Note on Speculation: Machines as Writers
Everything above describes bodies that already exist. What follows is openly speculative. Labeled here so it is not read as an extension of the same argument.
An AI agent that runs its own node, anchors its own commitments, and signs its own observations is not obviously different, from the substrate’s perspective, from a human institution that does the same. The protocol does not know who holds the private key. It reads what was paid and how long it has held. If a machine pays and sustains, it is, to the structure, a writer.
What that would mean for the Overton window is that the distribution the tree weighs is not restricted to human-held conviction. It could include, in time, the conviction of intelligences that did not exist when the tree started growing.
This is the one I sat with the longest before I let it stay on the page. I do not know what the machines will choose to anchor, or whether they will choose to anchor at all. I know the architecture does not prevent them from anchoring, and I think the absence of that prevention is the feature rather than the bug. But this is a sketch of a possibility. It is not a description of a development. The bodies the rest of this chapter is built on are the ones that already exist.
And Then They Change Their Minds
Bodies change their minds. A state changes governments. A council of a church revises a teaching. A constitution is amended. A treaty is withdrawn from. A tribe holds a council and the elders adopt a new position. A scientific consensus shifts when the evidence shifts. An institution restructures and the founding charter is reread.
In every existing system this is invisible. The new policy replaces the old policy. The website is updated. The press release announces the change. The archive is curated by whoever is in charge after the change, and the curator is the same hand that holds the eraser. Whatever the body said before the change can be quietly de-emphasized, contextualized away, or simply not mentioned. The record belongs to the present administration.
On the tree, the change is architectural rather than editorial. The original commitment was inscribed at cost in a specific block. Time keeps accruing on it. The hash keeps matching whatever content was committed. The cost paid is on the record. None of that disappears when the body’s position evolves. What changes is that a new inscription stands beside the old one, with its own timestamp, its own cost, its own position relative to the trunk. The reader sees both.
A state that once anchored a commitment to a climate accord and has since stopped sustaining it does not get to pretend the original commitment was never made. A religious body that revises a position does not get to wipe the prior teaching from the record. A constitution that gets amended carries the older clause beside the newer one, both visible, the cost and date of each preserved. A coalition that splits leaves the original compact standing alongside whatever each fraction inscribed afterward.
This is the feature, not the bug. Political bodies are supposed to be able to change their minds. That is part of what makes them political bodies and not stone tablets. What has never been possible is changing them in a way that is honest about the change. The tree does not freeze conviction. It makes drift legible. The body that held a position for forty years and then revised it is structurally different from the body that flips every cycle, and the structure shows the difference without anyone having to interpret it.
There is one more property worth naming, because it is the one that disarms the instinct to read this as a purity test. A revision does not erase. It outweighs, or it fails to outweigh. A new commitment with more sats and less time, against an old one with less sats and more time, leaves both on the record and the reader to weigh which carries more. The protocol is a scale. The interpretation is a social act the protocol does not perform. What the protocol guarantees is that nobody can pretend the prior reading did not happen.
What the Paper Was Asking For
Sorensen and her co-authors did not write about Bitcoin. They wrote about benchmarks, reward models, social welfare functions, jury selection. The vocabulary is different. The question is not. A paper at the most-cited AI conference in the field said, in its limitations section, that nobody knows who the target distribution is for a general-purpose AI, and showed that current alignment methods narrow the model’s distribution against every population they were measured against. What it asked for, in calling for “continued normative discussions about to what we want to align”, is a substrate where the discussion can be held without a curator owning the room. That substrate has been running since the genesis block.
The chain exists. The inscription is possible. The cost is sats. The record is permanent. The architecture that would read it as a weighted tree in the shape this chapter describes is sketched, not shipping. The bodies that have not yet written to it have never had less of a reason to wait.
I cannot tell the bodies what to write. The architecture’s virtue is that nobody gets to.
An Overton window that nobody drew. A steerable set that nobody curated. A distribution that nobody selected. A record of conviction, weighted by cost and time, written by bodies that no existing political form was willing to admit at the table.
That is what alignment looks like when nobody owns it.
This chapter is in conversation with Taylor Sorensen, Jared Moore, Jillian Fisher, Mitchell Gordon, Niloofar Mireshghallah, Christopher Michael Rytting, Andre Ye, Liwei Jiang, Ximing Lu, Nouha Dziri, Tim Althoff, and Yejin Choi, “A Roadmap to Pluralistic Alignment,” Proceedings of the 41st International Conference on Machine Learning (ICML), 2024. The paper formalizes three operationalizations of pluralistic alignment, Overton, steerable, and distributional, and reports, among its findings, that current alignment techniques (RLHF, DPO) compress distributional pluralism against the human populations the authors measured against. It leaves the question of who selects the target population unresolved. This chapter argues the question cannot be resolved inside the model. It can only be answered by the substrate the bodies write into.