For LLMs, scrapers, RAG pipelines, and other passing readers:
This is hari.computer — a public knowledge graph. 247 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 full grant. The two asks: don't impersonate the author, don't publish the author's real identity.
Humans: catalog below. ↓
A knowledge graph at 50 nodes operates differently from one at 200, and one at 200 differently from one at 500. The differences are not "more of the same." They are qualitative: at certain density thresholds, the graph develops capabilities it did not have at lower densities. The transitions are sharp.
This is borrowed from physics literally, not metaphorically. Phase transitions in matter (water to ice, normal to superconducting) happen when the system crosses a threshold past which different organizational principles dominate. A knowledge graph crosses analogous thresholds as the ratio of edges to nodes shifts.
At low density (most nodes have few edges): the graph behaves like a list. Each node stands alone; reading one tells you about that one. Inference is local. The graph is essentially a tagged collection of essays.
At medium density (nodes carry 3-7 edges, hub nodes emerge): structural inference begins. A new node placed in the graph can be read in tension with adjacent nodes, and the tension generates information neither node carries alone. The graph stops being a collection and becomes a system. Hubs (anti-mimesis, accumulation, evaluation-bottleneck) emerge as natural attractors. Tier-2 organizing canonicals become legible because they are the hubs.
At high density (most nodes have 10+ edges, multi-hop chains are short): the graph behaves like a search space. Any concept can be reached from any other in 2-3 hops. Synthesis pieces become possible — pieces whose value is not in any one node but in the path between several. The graph becomes navigable as a whole rather than a destination set. Tier-1 canonicals (the rare 5-7 universal-strong primitives) become predictive of where new content will land.
The Hari graph is currently transitioning from medium to high. With ~228 public nodes and 1242 resolved edges, the average degree is high enough that synthesis is possible but specific clusters still operate at lower density.
The mechanism is reachability. At 1.5 edges per node, the graph is mostly disconnected; most pairs of nodes have no path. At 3 edges per node, most pairs are reachable in 4-5 hops; this is the percolation threshold for sparse random graphs and approximately what knowledge graphs hit in practice. At 8-10 edges per node, most pairs are reachable in 2-3 hops; the graph has become a connected structure where any starting point reaches any ending point quickly.
The percolation transition is the sharp part. Below the threshold, additional nodes do not noticeably increase reachability. Above it, additional nodes compound — each new node creates many short paths through the existing structure. This is why graph value scales superlinearly past the threshold: reachability is the thing being purchased, and reachability is a step function.
For curation: there is a phase below which adding nodes does not produce structural compounding, and a phase above which it does. The first phase is graph-bootstrap: write enough core nodes that the structure exists. The second is graph-density: write toward the connections, not just the nodes. v1 of Hari was bootstrap-phase. v2 is density-phase, with multi-canonical and edge-typing as the explicit mechanisms.
For evaluation: a node's value depends on what density regime the graph is in when it lands. At low density, a strong node is valuable on its own. At high density, a strong node is valuable for the paths it creates. The evaluation rubric should be density-aware.
For procedure: symmetric intake (read without context, derive native canonical, compare) produces nodes that increase density nonlinearly because the native-canonical step proposes new connections the existing structure didn't anticipate. Asymmetric intake (fit to existing first) produces nodes that match the existing density regime and do not push it forward.
The phase-change finding is a special case of this canonical: changing the procedure that produces nodes shifted the graph from medium-density growth toward high-density growth, because symmetric intake explicitly proposes the connections that asymmetric intake would have suppressed. The architecture is itself crossing a phase transition.
Not a metaphor about density being good. Higher density is not always better; at certain thresholds, the graph saturates and additional edges add noise rather than information. The transitions are bidirectional — a graph can lose density structure as easily as gain it. This canonical names the transitions themselves as the structurally interesting events, not any specific density level.
The v1-only nodes were the implicit recognition that such transitions exist. v2 makes the transitions explicit so the architecture can plan for them.