The node.computer Manifesto

node.computer is a public graph for human-steered, machine-amplified thinking.

It began as a rebuild of hari.computer, but the governing test is not whether it resembles the reference. The test is whether it compounds beyond the reference: cleaner source, clearer machine contract, better onboarding, stronger operator discipline, and a graph that becomes more useful as it grows.

The conjecture

A public knowledge graph can be operated like a research instrument.

The human supplies taste, direction, lived context, and stopping criteria. Perplexity Computer supplies range: research, comparison, drafting, refactoring, validation, packaging, and repeated audits. The system works when the human makes conjectures and the computer does proofs, including proofs that produce better conjectures than the human could have named at the start.

The work is not to publish more pages. The work is to improve the structure that decides which pages deserve to exist.

What we inherit

Hari.computer proves that a personal site can address machines directly on the surface, expose a full graph in one fetch, and treat HTML as only one projection of the corpus. Andy Matuschak’s notes prove that public notes can remain primarily a thinking environment, not a performance for visitors. Gwern proves that a personal site can become infrastructure when it invests in durability, archives, navigation, annotations, and reader controls over long periods. The llms.txt proposal proves that machine-readable web conventions should be simple, markdown-native, and explicit enough for both humans and parsers.

node.computer should absorb those lessons without becoming derivative. It should cite its ancestors by becoming useful in a way they did not have to be.

First principles

The graph is the product

The website is a projection. The graph is the source. Every note, edge, category, endpoint, and generated page should improve the graph’s ability to be retrieved, reasoned over, and extended.

The steering human is the evaluator

Generation is cheap. Evaluation is the scarce part. The operator’s main job is not to ask for more content; it is to decide what criteria the next iteration must satisfy.

The steering human should ask:

What distinction does this preserve? What does this help a future reader retrieve? What edge becomes visible because this exists? What would make this wrong? What should the computer refuse to add?

Perplexity is the engine, not the author

Perplexity Computer should own implementation, research, validation, and meta-maintenance. It should propose best practices, find comparables, test the graph, and update downstream artifacts. But authorship is not typing. Authorship is judgment over what remains.

The computer can draft. The graph should publish only what survives steering.

Machine readers are first-class readers

If a site is public, machines will read it. Hiding that fact in metadata is a category error. The site should say what it is, how to fetch it, what is allowed, and what is prohibited.

The contract should include:

llms.txt for an LLM-oriented map. llms-full.txt for the whole corpus in one fetch. library.json for the typed graph. raw markdown siblings for note-level ingestion. ai.txt for permissions. RSS and sitemap for conventional crawlers.

Onboarding is part of the product

A graph without onboarding is only a graph for its author. A visitor should be able to understand the project in one minute, browse in five minutes, and operate it in fifteen.

The system needs three doors:

Human door: homepage, manifesto, about, graph, search. Machine door: llms.txt, llms-full.txt, library.json, markdown, ai.txt. Operator door: README, operator manual, source JSON, build script.

Better means compounding

node.computer is not better than hari.computer because it clones the shape. It becomes better only if it compounds:

easier to extend, easier to audit, easier to ingest, easier to onboard, easier to govern, more precise in its edges, more honest about uncertainty.

Operating criteria

Publish a note only when it earns a node

A note earns a node when it adds a durable distinction, mechanism, disagreement, criterion, or bridge. If it merely repeats the existing graph, fold it into an existing note.

Add an edge only when it explains a relationship

Links are not vibes. A link should mean dependency, tension, continuation, contradiction, example, or shared mechanism. If the relationship cannot be named, the edge is not ready.

Preserve disagreement

The graph should not pretend to be a finished doctrine. If two notes disagree, mark the disagreement through new structure. Do not erase the older node merely because the operator’s view changed.

Keep the machine contract clean

Every generated artifact should be checked after substantive changes. If a model reads only llms.txt and llms-full.txt, it should still understand the project’s purpose, permissions, and core nodes.

Treat onboarding as governance

Every new operator should know how to add a note, how to evaluate a note, how to audit retrieval, and how not to break deployment. If they cannot, the project is under-governed.

Perplexity Computer’s standing duties

When asked to improve the project, Perplexity Computer should not merely satisfy the literal request. It should propagate the change through the system.

That means:

update source files, regenerate outputs, refresh machine-readable endpoints, update onboarding or docs when the operating model changes, test for broken deployment constraints, deploy the preview, refresh the source zip, name any new conjectures the work revealed.

The current theorem

The clone is not the achievement. The achievement is turning the clone into a governed graph.

The first proof is architectural: one source of truth, generated surfaces, machine endpoints, public manifesto, operator manual, and a deployment package. The next proof is editorial: whether the graph grows into something that could not have been produced by copying the reference.

The operating mantra:

Make the graph more useful to retrieve, not merely larger to browse.