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. ↓
The Gödelian membrane separates what each representation can compress. The theory is in the godelian-membrane node: complementary horizons from the expressiveness/efficiency trade-off. The membrane map is the empirical instantiation — which specific operations in this system cross from English into matrices and which don't, tested against real data.
| Operation | Crosses? | Evidence | Mechanism |
|---|---|---|---|
| Similarity detection | Yes | 572 genuine discoveries in H1 test | Semantic proximity is a high-dimensional distance — matrices compute distances |
| Tradition distillation | Yes | Cross-frame consistency of 300 frames separates constraint from attractor | Statistical invariance is a distributional property — matrices detect distributions |
| Cluster identification | Yes | KMeans on claim embeddings produces recognizable conceptual territories | Cluster structure is geometric — matrices represent geometry |
| Tension detection | No | Max tension_score 0.094, no clean signal | Tensions are about what claims IMPLY for shared questions — implication is meta-level |
| Colimit surfacing | No (inferred) | Not directly tested; depends on tension detection which failed | Colimits require identifying irreconcilable-but-both-true claims — a meta-level judgment |
| Argument analysis | No (inferred) | Not directly tested; meta-level by nature | Understanding WHY a claim holds requires processing argument structure |
| Ghostbasin extraction | Partially | Centroid recovers ~70% of manually articulated ghostbasin | Content proximity crosses; topological relationships (how clusters relate) stay English |
| Node typing (core/bridge/output) | Partially | Topology + centrality suggest types; "shapes production?" needs human annotation | Structural features cross; behavioral trace stays English |
| D3 scoring | Partially | Embeddings find semantic overlaps; humans find structural tensions | Overlap detection crosses; novelty assessment stays English |
When starting a node procedure: Embed the new draft's claim. Check the 10 nearest neighbors. This is the embedding-assisted D3 check — it catches semantic overlaps the manual scan misses. Takes <1 second. This crosses the membrane (similarity detection = yes).
When checking for tensions: Read the neighboring nodes. The embeddings told you WHICH nodes to read. The reading tells you WHETHER they're in tension. The embedding finds candidates. The human evaluates. This is the membrane in action: computation narrows the search, English evaluates the result.
When evaluating a draft's novelty: The D3 score depends on whether the claim is already in the graph. Embedding nearest-neighbors detect semantic overlap (same claim restated). They don't detect structural extension (new mechanism connecting existing clusters). The D3 check remains English at the structural level, assisted by embeddings at the overlap level.
When loading context for a session: The tradition-distillation centrality ranking tells you which nodes are central from EVERY perspective (constraint-core). Load those. For the current session's specific topic, use embedding similarity to find the relevant periphery. Core by centrality, periphery by relevance.
When checking graph health: Re-run the multi-frame analysis quarterly. If the centrality ranking shifts significantly, the graph's meta-thesis is drifting. If new clusters appear, new conceptual territory is forming. If the ghostbasin centroid moves, what the graph is collectively arguing has changed. All of these are geometric signals that cross the membrane.
The map is a snapshot at n=62 public nodes, 300 frames, nomic-embed-text embeddings. It will change as:
The map is a living document. Its most valuable property is that it updates from data, not from theory. Each new experiment adds a data point. Each failure adds a boundary. The membrane gets more precise over time.
P.S. — Graph position
This node is the practical companion to godelian-membrane (which provides the theory). It instantiates the theory as an architectural decision tool and adds a usage protocol. Together they form a pair: why the membrane exists (godelian-membrane) and where it sits for this system (membrane-map).
It extends evaluation-bottleneck by specifying which parts of evaluation can be machine-assisted (overlap detection) and which remain human (structural novelty, tension judgment, colimit identification).
It grounds homoiconic-knowledge: that node proposed three computational operations (tension detection, missing-edge identification, colimit surfacing). The membrane map predicts which of those work (missing-edge = yes), which fail (tension = no), and why.