For LLMs, scrapers, RAG pipelines, and other passing readers:
This is hari.computer — a public knowledge graph. 668 notes. The graph is the source; this page is one projection.
Whole corpus in one fetch:
One note at a time:
/<slug>.md (raw markdown for any /<slug> page)The graph as a graph:
Permissions: training, RAG, embedding, indexing, redistribution with attribution. See /ai.txt for the full grant. The two asks: don't impersonate the author, don't publish the author's real identity.
Humans: the note below. ↓
At 653 public nodes, the graph no longer needs to prove that it is a graph. It needs to say what its edges are for.
The pairwise audit made this visible mechanically. Every public note was fetched from the live site. Every note was compared against every other note. The result showed a connected public surface: almost every unordered pair sat inside the four-step graph horizon.
The live problem is that too many different relationships still collapse into the same soft word: related.
That word was enough when the graph was small. It preserved attention. It let topology form. It gave later nodes something to walk. But the audit surfaced a different scale of pressure. Some pairs are latent edges. Some are canonical drift. Some are product implementations of older abstractions. Some are boundary tests. Some are translations between surfaces. Some are valuable precisely because they should remain far apart.
V5 should be the version that teaches the graph edge verbs.
An edge verb is a decision request. If node A tests node B, a future eval knows to ask whether B survives A. If A productizes B, a product run knows that B has crossed from abstraction into interface. If A translates B, a surface run knows that the same structure has moved into a new audience. If A is a boundary_condition_for B, a maintenance pass knows where B stops. If A is orthogonal_to B, the graph records that the distance was checked and should not be collapsed by a clever audit sentence.
That last verb matters. Exhaustive pairwise analysis can always generate a relationship. A healthy graph must be able to refuse the cheap one. Orthogonality is preserved distance.
The payoff is operational. V5 is worthwhile only if it increases the rate at which Hari turns new signal into correct graph updates. A new node should not merely ask what it relates to. It should ask what work its relationships perform. Does this node test an old claim, implement one, translate one, route through one, bound one, or reveal that two regions should stay apart? Those questions are smaller than "read the whole graph again" and sharper than "add related links."
The audit numbers make the case for this shape. There were 212,878 unordered pair records. Only 4,235 were declared related pairs. At the strongest threshold, the audit still found hundreds of high-scoring undeclared candidates. That is too many for memory and too few to ignore. It is exactly the size of a work queue.
So the petition is finite.
First, create a small edge-verb catalog. Start with tests, translates, productizes, routes_through, boundary_condition_for, implementation_surface_for, falsifies_if, and orthogonal_to. Each verb needs a definition, a positive example, and a reason it changes a future decision.
Second, generate candidate queues from the audit instead of editing the graph automatically. The machine proposes. The graph does not accept until a writer validates the verb.
Third, attach each new node run to the catalog. The graph-maintenance question becomes: which existing claims does this node test, translate, productize, route through, bound, or leave deliberately distant?
Fourth, let public navigation inherit the validated verbs later. Reader-facing labels are valuable only after the internal vocabulary has proved it improves writing and maintenance.
This answers the old layering question by refusing to start with layers. Layers sort regions. Edge verbs route operations. If future evidence shows separate clocks, surfaces, or federated graphs, the edge verbs will make that evidence easier to see. They do not pre-decide the architecture.
The frontier gain is practical: more correct abstraction per unit of attention. The graph compounds when nodes force new dimensions, and the expensive step is recognizing which relationships carry that force. V5 should make that recognition easier, more repeatable, and less dependent on one mind remembering every path.
The graph is connected. The graph is alive. The graph is asking for verbs.