Related: sources · notes · metadata · Published Pieces

Humans Publish, Agents Cite, the Protocol Rewards

People should not have to write bibliographies for the internet to remember what they used.

People should not have to write bibliographies for the internet to remember what they used.

Manual citation is a beautiful practice in the right context. It belongs to scholarship, law, serious journalism, and the parts of public life where people still accept that claims require lineage. But manual citation cannot scale to the world we are entering.

The world has too much content, too many claims, too many sources, too much prior art, too many contradictions, too many updates, and too much machine-generated synthesis. A human cannot reliably know every relevant prior before publishing. Even when they know, they may not cite. Even when they cite, they may cite politically, defensively, lazily, or performatively.

So the citation layer has to become native to the medium.

In Choir, humans publish, agents cite, the protocol rewards.

A user writes a vtext, records a voice note, publishes a research artifact, responds to another piece, shares a source, builds an appagent, drafts a critique, or asks agents to investigate a topic. The human is doing human work: perceiving, judging, arguing, speaking, framing, remembering, creating.

The agents then do what humans cannot do at scale. They search the artifact graph. They retrieve related prior work. They compare claims. They surface sources. They check semantic similarity and dissimilarity. They ask whether an idea is old, new, rediscovered, contradicted, refined, or dependent. They look for earlier artifacts that anticipated the current one. They detect when a source has been falsified, challenged, superseded, or repeatedly relied upon.

The protocol then records the relation.

That relation is a citation.

But a citation in Choir is not merely approval. It is not merely “this source is true.” It is not merely academic etiquette. A citation is a dependency event.

It can mean: this new artifact relies on that earlier artifact. It can mean: this new artifact extends that prior. It can mean: this new artifact contradicts that prior. It can mean: this source is necessary background. It can mean: this old voice clip captured the same frame earlier. It can mean: this argument is a competitor in the same conceptual space. It can mean: this source has to be reckoned with, even if only to be rejected.

The central object is not agreement. The central object is relation.

That is how public cognition becomes scalable. The internet currently has fragments everywhere: posts, essays, podcasts, replies, videos, comments, screenshots, citations, documents, rumors, transcripts, and memory traces. They are connected socially, but not intellectually. They are connected by followers, likes, reposts, and status, not by structured dependency.

Choir’s citation layer should change that.

If a person publishes an idea that later becomes important, the graph should remember. If another person corrects it, the graph should remember. If a third person synthesizes both into a stronger frame, the graph should remember. If an agent uses that synthesis to brief thousands of listeners later, the graph should remember.

This is how ideas become assets.

The protocol’s job is to make intellectual dependency explicit enough that value can flow back through the graph.

The human should not need to campaign for credit every time. The system should know that their artifact mattered. Not perfectly. Not magically. But better than the current internet, where value often disappears into platform metrics, screenshots, viral paraphrase, and AI training sets.

Automatic citation also changes the quality of discourse. If agents routinely surface prior relevant work, users are less likely to rediscover the obvious and call it invention. If contradictions are tracked, users are less able to quietly reverse themselves without explanation. If old claims can be retrieved, public speakers become more accountable. If dissenting priors are visible, consensus becomes harder to fake.

This is not a demand that every person become a scholar. It is the opposite.

Let humans speak. Let humans write. Let humans publish. Let humans argue. Let humans improvise. Let humans be partial, emotional, brilliant, wrong, early, late, confused, poetic, and alive.

Then let agents do the boring work of memory.

The protocol does not replace judgment. It gives judgment a better substrate.

A good citation economy should care about novelty, but not novelty alone. The newest thing is not always the best thing. It should care about age, but not age alone. Being first is not enough if the idea was wrong, vague, or unused. It should care about downstream use, but not raw popularity. A citation cartel is not truth. It should care about lack of falsification, but not confuse survival with certainty. It should care about source quality, but not let institutions monopolize credibility.

The graph must be adversarial. It must expect gaming. Public cognition is a competitive domain. People will try to inflate their importance, bury rivals, cite friends, ignore enemies, and turn any reward system into status competition. The answer is not to abandon rewards. The answer is to make the reward system inspectable, contestable, and attached to provenance.

The promise is simple: if your thought matters later, Choir should be able to find it. If your source helps someone understand, Choir should cite it. If your correction improves the record, Choir should reward it. If your voice becomes relevant to future listeners, Choir should play what you actually said. If your artifact becomes part of the public mind, Choir should not let it disappear.

Humans publish. Agents cite. The protocol rewards.

That is the minimum architecture for intellectual fairness in the AI age.